JP7495398B2 - Machine learning health analysis using mobile devices - Google Patents

Description

CROSS-REFERENCE TO RELATED APPLICATIONS This application claims priority to U.S. Application No. 16/580,574, filed September 24, 2019, which is a continuation of U.S. Application No. 16/153,403, filed October 5, 2018, the entire contents of which are incorporated herein by reference.

Indicators of an individual's physiological health ("health indicators"), such as, but not limited to, heart rate, heart rate variability, blood pressure, and ECG (electrocardiogram), to name a few, may be measured or calculated at any discrete time from the data collected to measure the health indicator. Often, the value of a health indicator at a particular time, or changes over time, provide information about an individual's health status. For example, a low or high heart rate or blood pressure, or an ECG clearly indicating myocardial ischemia, may indicate the need for immediate intervention. However, a reading, a series of readings, or changes in readings over time of these indicators may provide information that is not recognized as requiring attention by a user, or even a medical professional.

For example, arrhythmias may occur continuously or intermittently. Continuous arrhythmias may be most clearly diagnosed from an individual's electrocardiogram. Since continuous arrhythmias are always present, ECG analysis may be applied at any time to diagnose the arrhythmia. ECG may also be used to diagnose intermittent arrhythmias. However, intermittent arrhythmias present the challenge of applying diagnostic techniques when an individual is experiencing an arrhythmia, since they may be asymptomatic and/or are continuous by definition. Thus, the actual diagnosis of continuous arrhythmias is notoriously difficult. This particular difficulty is compounded by asymptomatic arrhythmias, which account for nearly 40% of arrhythmias in the United States. Boriani G. and Pettorelli D., Atrial Fibrillation Burden and Atrial Fibrillation type: Clinical Significance and Impact on the Risk of Stroke and Decision Making for Long-term Anticoagulation, Vascul Pharmacol., 83:26-35 (August 2016), 26 pages.

Sensors and mobile electronic technologies exist that allow frequent or continuous monitoring and recording of health indicators. However, the capabilities of these sensor platforms often exceed the ability of traditional medicine to interpret the data they generate. The physiological significance of health indicator parameters such as heart rate is often only clearly defined in specific medical situations, for example, heart rate is typically evaluated as a single scalar value out of context from other data/information that may affect the health indicator. A resting heart rate in the range of 60-100 beats per minute (BPM) may be considered normal. A user may generally manually measure their resting heart rate once or twice a day.

A mobile sensor platform (e.g., a mobile blood pressure cuff, a mobile heart rate monitor, or a mobile ECG device) may be able to continuously monitor a health indicator (e.g., heart rate) while simultaneously acquiring other data about the user such as, but not limited to, activity level, body position, and environmental parameters such as temperature, pressure, location, and may be able to generate measurements, for example, every 1 or 5 seconds. Over the course of a 24-hour period, this may result in thousands of independent measurements of a health indicator. Relatively little data or medical consensus exists on what a "normal" series of thousands of measurements looks like, as opposed to measurements once or twice a day.

The devices currently used to continuously measure a user/patient's health metrics range from bulky, invasive, and inconvenient to simple wearable or handheld mobile devices. Currently, these devices do not provide the ability to effectively utilize the data to continuously monitor a person's health status. It is left to the user or a medical professional to evaluate the health metrics in light of other factors that may affect these health metrics to determine the user's health status.

U.S. Provisional Application No. 61/872,555 U.S. Patent No. 9,420,956 U.S. Patent No. 9,572,499 U.S. Patent No. 9,351,654 U.S. Patent No. 9,247,911 U.S. Patent No. 9,254,095 U.S. Patent No. 8,509,882 US Patent Application Publication No. 2015/0018660 US Patent Application Publication No. 2015/0297134 US Patent Application Publication No. 2015/0320328

Boriani G. and Pettorelli D., Atrial Fibrillation Burden and Atrial Fibrillation Type: Clinical Significance and Impact on the Risk of Stroke and Decision Making for Long-term Anticoagulation, Vascul Pharmacol., 83:26-35 (August 2016), 26 p.

Certain features described herein are set forth with particularity in the appended claims. A better understanding of the features and advantages of the disclosed embodiments can be obtained by reference to the following detailed description that sets forth illustrative embodiments in which the principles described herein are utilized, and the accompanying drawings, in which:

FIG. 1 illustrates a convolutional neural network that may be used in accordance with some embodiments described herein. FIG. 1 illustrates a convolutional neural network that may be used in accordance with some embodiments described herein. FIG. 1 illustrates a recurrent neural network that may be used in accordance with some embodiments described herein. FIG. 1 illustrates a recurrent neural network that may be used in accordance with some embodiments described herein. FIG. 1 illustrates an alternative recurrent neural network that may be used in accordance with some embodiments described herein. FIG. 1 shows hypothetical data plots to demonstrate the application of some embodiments described herein. FIG. 1 shows hypothetical data plots to demonstrate the application of some embodiments described herein. FIG. 1 shows hypothetical data plots to demonstrate the application of some embodiments described herein. FIG. 1 illustrates an alternative recurrent neural network according to some embodiments described herein, as well as hypothetical plots used to explain some of these embodiments. FIG. 1 illustrates an alternative recurrent neural network according to some embodiments described herein, as well as hypothetical plots used to explain some of these embodiments. FIG. 1 illustrates an alternative recurrent neural network according to some embodiments described herein, as well as hypothetical plots used to explain some of these embodiments. FIG. 1 illustrates an alternative recurrent neural network according to some embodiments described herein, as well as hypothetical plots used to explain some of these embodiments. FIG. 1 illustrates an alternative recurrent neural network according to some embodiments described herein, as well as hypothetical plots used to explain some of these embodiments. FIG. 2 illustrates a deployed recurrent neural network according to some embodiments described herein. FIG. 1 illustrates a system and device according to some embodiments described herein. FIG. 1 illustrates a system and device according to some embodiments described herein. FIG. 1 illustrates a method according to some embodiments described herein. FIG. 1 illustrates a method according to some embodiments described herein and a hypothetical plot of heart rate versus time to demonstrate one or more embodiments. FIG. 1 illustrates a method according to some embodiments described herein and a hypothetical plot of heart rate versus time to demonstrate one or more embodiments. FIG. 1 illustrates a method according to some embodiments described herein. FIG. 1 shows hypothetical data plots to demonstrate the application of some embodiments described herein. FIG. 1 illustrates a system and device according to some embodiments described herein.

The large volume of data, the complexity of the interconnections between health indicators and other factors, and limited clinical guidance may limit the effectiveness of any monitoring system that attempts to detect anomalies in continuous and/or ambulatory sensor data through specific rules based on traditional medical practice. The embodiments described herein include devices, systems, methods, and platforms that can utilize predictive machine learning models to detect anomalies in an unsupervised manner from time sequences of health indicator data, either alone or in combination with other factor data (as defined herein).

Atrial fibrillation (AF or AFib) is present in 1-2% of the general population, and the presence of AF increases the risk of morbidity and adverse outcomes such as stroke and heart failure. Boriani G. and Pettorelli D., Atrial Fibrillation Burden and Atrial Fibrillation type: Clinical Significance and Impact on the Risk of Stroke and Decision Making for Long-term Anticoagulation, Vascul Pharmacol., 83:26-35 (August 2016), 26 p. AFib in many, estimated to be as much as 40% of AF patients, may be asymptomatic, and these asymptomatic patients have a similar risk profile for stroke and heart failure as symptomatic patients. Id. However, symptomatic patients can take aggressive measures, such as taking anticoagulants or other medications, to reduce the risk of negative outcomes. The use of implantable electrical devices (CIEDs) can detect asymptomatic AF (so-called silent AF or SAF) and the duration a patient is in AF. Id. From this information, the amount of time these patients spend in AF, i.e., AF burden, can be determined. Id. AF burdens greater than 5-6 minutes, and especially greater than 1 hour, are associated with significantly elevated risk of stroke and other negative health outcomes. Id. Thus, the ability to measure AF burden in asymptomatic patients can lead to earlier interventional treatments to reduce the risk of negative health outcomes associated with AF. Id. Detection of SAF is difficult and typically requires some form of continuous monitoring. Currently, continuous monitoring of AF requires bulky, sometimes invasive, and expensive devices, and such monitoring requires advanced medical professional oversight and review.

Many devices continuously acquire data to provide measurements or calculations of health indicator data, for example, but not limited to, FitBit®, Apple Watch®, Polar®, smartphones, tablets, etc., fall under the category of wearable and/or mobile devices. Other devices include permanent or semi-permanent devices (e.g., Holter) on or in the user/patient, and other devices may include larger devices in hospitals that may be mobile by being on a cart. However, little is done with this measured data other than periodically monitoring it on a display or setting simple data thresholds. Even by trained medical professionals, observations of the data may often appear normal, with one major exception being in cases with easily identifiable acute symptoms. It is very difficult, and practically impossible, for medical professionals to continuously observe health indicators to observe abnormalities and/or trends in the data that may indicate something more serious.

As used herein, a platform includes one or more customized software applications (or "applications") configured to interact with each other locally or over a distributed network, including the cloud and the Internet. The applications of the platform described herein are configured to collect and analyze user data and may include one or more software models. In some embodiments of the platform, the platform includes one or more hardware components (e.g., one or more sensing devices, processing devices, or microprocessors). In some embodiments, the platform is configured to operate with one or more devices and/or one or more systems. That is, the devices described herein are configured in some embodiments to execute the platform's applications using an embedded processor, and in some embodiments, the platform is utilized by a system that includes one or more computing devices that interact with or execute one or more applications of the platform.

The present disclosure describes systems, methods, devices, software, and platforms for continuously monitoring user data related to one or more health indicators (e.g., but not limited to, PPG signal, heart rate, or blood pressure) from a user device in combination with corresponding (e.g., in time) data related to factors (herein referred to as "other factors") that may affect the health indicators to determine whether the user has a normal health status as determined by or compared to, for example, but not limited to, either (i) a group of individuals affected by similar other factors, or (ii) the user himself/herself affected by similar other factors. In some embodiments, the measured health indicator data, alone or in combination with the other factor data, is input into a trained machine learning model that determines the probability that the user's measured health indicator is considered to be within a healthy range and notifies the user accordingly if not. A user not within a healthy range may increase the likelihood that the user may be experiencing a health event that warrants high fidelity information for confirming a diagnosis, such as an arrhythmia that may be symptomatic or asymptomatic. The notification may take the form of, for example, requesting the user to obtain an ECG. Other high fidelity measurements may be required such as blood pressure, pulse oximetry, to name two, and ECG is just one example. The high fidelity measurement, in this embodiment the ECG, may be evaluated by an algorithm and/or a medical professional to provide an indication or diagnosis (collectively referred to herein as "diagnosis", recognizing that only a physician may provide a diagnosis). In the ECG example, the diagnosis may be AFib, or any number of other well-known conditions diagnosed utilizing ECG.

In further embodiments, the diagnosis is used to label a low-fidelity data sequence (e.g., heart rate or PPG), which may include data sequences of other factors. This low-fidelity data sequence with the high-fidelity diagnosis label is used to train a high-fidelity machine learning model. In these further embodiments, the training of the high-fidelity machine learning model may be trained by unsupervised learning or may be updated from time to time with new training examples. In some embodiments, the user's measured low-fidelity health indicator data sequence, and optionally the corresponding (in time) data sequences of other factors, are input into a trained high-fidelity machine learning model to determine a probability and/or prediction that the user is experiencing or has experienced the diagnosed condition for which the high-fidelity machine learning model was trained. This probability may include a probability of when an event will start and when it will end. Some embodiments may, for example, calculate the user's atrial fibrillation (AF) burden, or the amount of time the user will experience AF over time. Previously, AF burden could only be determined using cumbersome and expensive Holter, or implantable continuous ECG monitoring devices. Thus, some embodiments described herein may continuously monitor the health status of a user and notify the user of changes in health status by continuously monitoring health indicator data (e.g., but not limited to, PPG data, blood pressure data, and heart rate data) obtained solely from a device worn by the user, or in combination with corresponding data regarding other factors. As used herein, "other factors" include those that may affect the health indicators and/or may affect the data representing the health indicators (e.g., PPG data). These other factors may include various factors such as, by way of example and not by way of limitation, temperature, altitude, activity level, weight, sex, diet, standing position, sitting position, falling, lying position, weather, and BMI, to name a few. In some embodiments, a mathematical or empirical model, rather than a machine learning model, may be used to determine when to notify the user to obtain high-fidelity measurements, which may then be analyzed and used to train the high-fidelity machine learning models described herein.

Some embodiments described herein can detect user anomalies in an unsupervised manner by receiving a primary time sequence of health indicator data, optionally receiving one or more secondary time sequences of other factor data that correspond in time to the primary time sequence of health indicator data, where the secondary time sequences may come from a sensor or from an external data source (e.g., via a network connection, computer API, etc.), providing the primary and secondary time sequences to a preprocessor that may perform operations on the data such as filtering, caching, averaging, time alignment, buffering, upsampling, and downsampling, providing the time sequence of data to a machine learning model trained and/or configured to utilize values of the primary and secondary time sequences to predict the next value of the primary sequence at a future time, comparing the predicted primary time sequence value generated by the machine learning model at a particular time t with the measured value of the primary time sequence at time t, and alerting or prompting the user to take action if the difference between the predicted future time sequence and the measured time sequence exceeds a threshold or criterion.

Thus, some embodiments described herein detect when the observed behavior of a primary sequence of physiological data over time and/or in response to a secondary sequence of observed data differs from that predicted given the training examples used to train the model. If the training examples were collected from normal individuals or from data previously classified as normal for a particular user, the system can function as an anomaly detector. If the data was simply collected from a particular user without any other classification, the system can function as a change detector that detects changes in the health indicator data against which the primary sequence is being measured compared to the time the training data was captured.

Described herein are software platforms, systems, devices, and methods for generating and using trained machine learning models to predict or determine the probability that a user's measured health indicator data (primary sequence) under the influence of other factors (secondary sequence) is outside the range of normal for a healthy population under the influence of similar other factors (i.e., global model), or outside the range of normal for that particular user under the influence of similar other factors (i.e., personalized model). In some embodiments, the user may be prompted to obtain additional measured high-fidelity data that may be used to label previously obtained low-fidelity user health indicator data to generate a different trained high-fidelity machine learning model that has the ability to predict or diagnose anomalies or events using only the low-fidelity health indicator data, such anomalies typically being identified or diagnosed only using the high-fidelity data.

Some embodiments described herein may include inputting a user's health indicator data and, optionally, inputting corresponding (in time) data of other factors into a trained machine learning model, which predicts the user's health indicator data or a probability distribution of health indicator data at future time steps. The prediction in some embodiments is compared to the user's measured health indicator data at the time step of the prediction, and if the absolute value of the difference exceeds a threshold, the user is notified that the user's health indicator data is outside of the normal range. This notification, in some embodiments, may include a diagnosis or instructions to do something, such as, but not limited to, to obtain additional measurements or to contact a medical professional. In some embodiments, the health indicator data and corresponding (in time) data of the other factors from a population of healthy people are used to train the machine learning model. It will be understood that the other factors in the training example used to train the machine learning model may not be population averages, and the data for each of the other factors corresponds in time to the collection of health indicator data for the individuals in the training example.

Some embodiments are described as receiving discrete data points in time, predicting discrete data points at future times from the input, and then determining whether the loss between the discrete measurement input at the future time and the predicted value at the future time exceeds a threshold. Those skilled in the art will readily appreciate that the input data and output predictions may take forms other than discrete data points or scalars. For example, and without limitation, health indicator data sequences (also referred to herein as primary sequences) and other data sequences (also referred to herein as secondary sequences) may be divided into segments in time. Those skilled in the art will recognize that the manner in which the data is segmented is a matter of design choice and may take many different forms.

Some embodiments divide the health indicator data sequence (also referred to herein as the primary sequence) and other data sequences (also referred to herein as the secondary sequence) into two segments, the past, which represents all data before a particular time t, and the future, which represents all data from time t onwards. These embodiments input the health indicator data sequence for the past time segments and all other data sequences for the past time segments into a machine learning model configured to predict the most likely future segment (or distribution of possible future segments) of health indicator data. Alternatively, these embodiments input the health indicator data sequence for the past time segments, all other data sequences for the past time segments, and other data sequences from the future segments into a machine learning model configured to predict the most likely future segment (or distribution of possible future segments) of health indicator data. The predicted future segment of health indicator data is compared to the user's measured health indicator in the future segment to determine a loss and whether the loss exceeds a threshold, and if the loss exceeds a threshold, some action is taken. Actions may include, for example, without limitation, an action of notifying the user to obtain additional data (e.g., ECG or blood pressure), an action of notifying the user to contact a medical professional, or an action of automatically triggering the acquisition of additional data. Automatic acquisition of additional data may include, for example, without limitation, an ECG acquisition via a sensor operably coupled (wired or wirelessly) to a computing device worn by the user, or a blood pressure via a mobile cuff around the user's wrist or other suitable body part and coupled to a computing device worn by the user. A segment of data may include a single data point, many data points over a period of time, an average value of these data points over a period of time, where the average value may include a true mean, a median, or a mode. In some embodiments, segments may overlap in time.

These embodiments detect when the observed behavior or measurement of a sequence of data of a health indicator over time influenced by a sequence of data of a (temporally) corresponding other factor differs from that expected from training examples, where the training examples are collected under similar other factors. If the training examples are collected from healthy individuals under similar other factors, or from data previously classified as healthy for a particular user under similar other factors, these embodiments function as anomaly detectors from a healthy population or from a particular user, respectively. If the training examples are simply collected from a particular user without any other classification, these embodiments function as change detectors that detect a transformation in the health indicator at the time of measurement compared to the time the training examples were collected for the particular user.

Some embodiments described herein utilize machine learning to continuously monitor a person's health indicators under the influence of one or more other factors and assess whether a person is healthy, taking into account a population classified as healthy under the influence of similar other factors. As one of ordinary skill in the art would readily appreciate, a number of different machine learning algorithms or models (including, but not limited to, Bayesian, Markovian, Gaussian processes, clustering algorithms, generative models, kernels, and neural network algorithms) may be used without going beyond the scope described herein. As will be appreciated by one of ordinary skill in the art, a typical neural network may use, by way of example and not limitation, one or more layers of nonlinear activation functions to predict an output for a received input, and may include one or more hidden layers in addition to an input layer and an output layer. The output of each hidden layer in some of these networks is used as an input to the next layer in the network. Examples of neural networks include, by way of example and not limitation, generative neural networks, convolutional neural networks, and recurrent neural networks.

Some embodiments of the health monitoring system monitor an individual's heart rate and activity data as low fidelity data (e.g., heart rate or PPG data) and detect conditions (e.g., AFib) that are typically detected using high fidelity data (e.g., ECG data). For example, the individual's heart rate may be provided by a sensor continuously or at discrete intervals (e.g., every 5 seconds). The heart rate may be determined based on PPG, pulse oximetry, or other sensors. In some embodiments, activity data may be generated as a number of steps, an amount of sensed movement, or other data points indicative of activity level. The low fidelity (e.g., heart rate) data and the activity data may then be input into a machine learning system to determine a high fidelity outcome prediction. For example, the machine learning system may use the low fidelity data to predict arrhythmia or other signs of the user's cardiac health. In some embodiments, the machine learning system may use input of a segment of data input to determine a prediction. For example, an hour of activity level data and heart rate data may be input into the machine learning system. The system may then use the data to generate a prediction of a condition such as atrial fibrillation. Various embodiments of the invention are discussed more fully below.

Referring to FIG. 1A, a trained convolutional neural network (CNN) 100 (an example of a feedforward network) puts input data 102 (e.g., a photo of a boat) into a convolutional layer (aka hidden layer) 103 and applies a series of trained weights or filters 104 to the input data 103 in each of the convolutional layers 103. The output of the first convolutional layer is an activation map (not shown) that is the input to the second convolutional layer, to which trained weights or filters (not shown) are applied, and the output of the subsequent convolutional layers results in activation maps that represent increasingly more complex features of the input data to the first layer. After each convolutional layer, a nonlinear layer (not shown) is applied to introduce nonlinearity to the problem, which may include tanh, sigmoid, or ReLU. In some cases, a pooling layer (not shown) is applied after the nonlinear layer, which is also called a downsampling layer, and basically applies the same length filter and stride to the input and outputs the maximum number in all sub-regions that the filter convolves. Other options for pooling are average pooling and L2-norm pooling. The pooling layer reduces the spatial dimensions of the input volume, reducing computational costs and controlling overfitting. The final layer of the network is a fully connected layer, which receives the output of the last convolutional layer and outputs an n-dimensional output vector representing the quantity to be predicted, e.g., the probability of the image classification, 20% car, 75% boat, 5% bus, and 0% bicycle, i.e., resulting in a predicted output 106 (O*), e.g., this is likely to be a photo of a boat. The output can be a scalar-valued data point predicted by the network, e.g., a stock price. The trained weights 104 can be different for each of the convolutional layers 103, as described more fully below. To achieve this real-world prediction/detection (e.g., it's a boat), the neural network needs to be trained on known data inputs or training examples, resulting in a trained CNN 100. To train the CNN 100, many different training examples (e.g., many pictures of boats) are input to the model. Those skilled in the art of neural networks will understand that
It will be appreciated that the above description provides a somewhat simplistic view of CNNs to provide some context to the current discussion, and that the application of any CNN alone or in combination with other neural networks is equally applicable and within the scope of some of the embodiments described herein.

FIG. 1B illustrates a training CNN 108. In FIG. 1B, the convolutional layers 103 are shown as individual hidden convolutional layers 105, 105' up to convolutional layer 105 ^n-1 , with the last nth layer being a fully connected layer. It will be understood that the last layer may be two or more fully connected layers. Training examples 111 are input to the convolutional layers 103, and nonlinear activation functions (not shown) and weights 110, 110' through 110 ⁿ are applied serially to the training examples 111, with the output of any hidden layer being input to the next layer, until the last nth fully connected layer 105 ⁿ produces an output 114. The output or prediction 114 is compared to the training examples 111 (e.g., photos of boats), resulting in a difference 116 between the output or prediction 114 and the training examples 111. If the difference or loss 116 is less than some pre-set loss (e.g., the output or prediction 114 predicted that the object was a boat), the CNN is considered to have converged and trained. If the CNN has not converged, then using the technique of backpropagation, the weights 110 and 110′ through 110 ⁿ are updated depending on how close the prediction is to the known input. Those skilled in the art will appreciate that methods other than backpropagation can be used to adjust the weights. A second training example (e.g., a different photo of a boat) is input, and the process is repeated again with the updated weights, which are then updated again until the nth training example (e.g., the nth photo of the nth boat) is input. This is repeated over and over with the same n training examples until the convolutional neural network (CNN) is trained or converges to the correct output for the known input. Once the CNN 108 is trained, the weights 110, 110′ through 110 ⁿ are fixed and used in the trained CNN 100, and these weights are the weights 104 as shown in FIG. 1A. As explained, there are different weights for each convolutional layer 103 and each of the fully connected layers. The trained CNN 100 or model is then fed image data to determine or predict what it was trained to predict/identify (e.g., boats) as explained above. Any trained model, CNN, RNN, etc., may be further trained with additional training examples or with prediction data output by the model then used as training examples, i.e., weights may be allowed to be modified. Machine learning models may be trained "offline", e.g., trained on a computing platform separate from the platform that uses/executes the trained model and then transferred to that platform. Alternatively, the embodiments described herein may periodically or continuously update the machine learning model based on newly acquired training data. This updated training may occur on a separate computing platform that delivers the updated training model to the platform that uses/executes the retrained model via a network connection, or the training/retraining/updating process may occur on the platform itself as new data is acquired. Those skilled in the art will appreciate that CNNs are applicable to data in a fixed sequence (e.g., photos, characters, words, etc.), or time sequences of data. For example, the ordered health indicator data and other factor data can be modeled using CNN. Some embodiments utilize feedforward, CNN with skip connections, and Gaussian mixture model output to determine a probability distribution for a predicted health indicator, such as heart rate, PPG, or arrhythmia.

Some embodiments may utilize other types and configurations of neural networks. The number of convolutional layers may be increased or decreased, as may the number of fully connected layers. In general, the optimal number and ratio of convolutional and fully connected layers may be set empirically by determining which configuration gives the best performance on a given dataset. The number of convolutional layers may be reduced to 0, leaving a fully connected network. The number of convolutional layers and the width of each filter may also be increased or decreased.

The output of the neural network may be a single scalar value corresponding to an accurate prediction for the primary time sequence. Alternatively, the output of the neural network may be a logistic regression where each category corresponds to a particular range or class of primary time sequence values, with any number of alternative outputs readily understood by one of ordinary skill in the art.

The use of Gaussian mixture model outputs in some embodiments is intended to constrain the network to learn a shaped probability distribution, improving generalization on limited training data. The use of multiple elements in some embodiments of a Gaussian mixture model is intended to allow the model to learn multi-modal probability distributions. Machine learning models that combine or aggregate the results of different neural networks may also be used, and the results may also be combined.

Machine learning models with updatable memory or states from previous predictions to apply to subsequent predictions are another approach to model sequence data. In particular, some embodiments described herein utilize recurring neural networks. Referring to the example of FIG. 2A, a diagram of a trained recurrent neural network (RNN) 200 is shown. The trained RNN 200 has updatable states (S) 202 and trained weights (W) 204. Input data 206 is input to the state 202, where the weights (W) are applied and a prediction 206 (P*) is output. In contrast to linear neural networks (e.g., CNN 100), the state 202 is updated based on the input data, thereby serving as a memory from the previous state for the next prediction using the next data in the sequence. Updating the states gives the RNN a cyclic or looping mechanism. To better demonstrate, FIG. 2B shows the trained RNN 200 deployed and its applicability to sequence data. When unrolled, the RNN looks similar to a CNN, but in an unrolled RNN, each of the seemingly similar layers appears as a single layer with updated states, with the same weights applied at each iteration of the loop. While a single layer is depicted here for clarity of explanation, one skilled in the art will recognize that the single layer may itself have sub-layers. Input data (I _t ) 208 at time t is input to state (S _t ) 210 at time t, and the trained weights 204 are applied within a cell (C _t ) 212 at time t. The output of C _t 212 represents the prediction at time step t+1.

214, and the updated state S _t+1 216. Similarly, in C _t+1 220, I _t+1 218 is input to S _t+1 216 and the same trained weights 204 are applied, and the output of C _t+1 220 is

222. As above, S _t+1 is updated from S _t , so S _t+1 has memory from S _t from the previous time step. For example, and without limitation, this memory may include previous health indicator data or previous other factor data from one or more previous time steps. This process continues for n steps, with I t _+n 224 being input to S _t+n 226 and the same weights 204 being applied. The output of cell C _t+n is the prediction

In particular, the state is updated from the previous time step, giving the RNN the advantage of memory from previous states. This property makes the RNN an alternative choice for making predictions on sequence data for some embodiments. However, as explained above, there are other suitable machine learning techniques for performing such predictions on sequence data, including CNNs.

An RNN, like a CNN, can process a data string as input and output a predicted data string. A simple way to illustrate this aspect of using an RNN is to use the example of natural language prediction. Use the phrase The sky is blue. The word (i.e., data) string has a context. So as the state is updated, the data string is updated every iteration to provide the context to predict blue. As just described, an RNN has a memory component to aid in making predictions on sequence data. However, the memory in the updated state of the RNN may be limited in the extent to which it can look back, similar to short-term memory. Similar to long-term memory, when predicting sequence data where a longer look back is desired, the fine-tuning of the RNN just described may be used to achieve this. The sentence Mary speaks fluent French, where the word to be predicted is unclear from the immediately preceding or surrounding words, is again a simple example to illustrate. It is unclear from the immediately preceding word that French is the correct prediction, i.e., it is unclear which language, only that one language is the correct prediction. The correct prediction may lie in the context of words separated by gaps larger than a single word string. Long Short Term Memory (LSTM) networks are a special kind of RNN that can learn these long-term(er) dependencies.

As explained above, RNNs have a relatively simple repetitive structure, for example, they may have a single layer with a nonlinear activation function (e.g., tanh or sigmoid). LSTMs similarly have a chain-like structure, but with (for example) four neural network layers instead of one. These additional neural network layers give LSTMs the ability to remove or add information to a state by using a structure called a cell gate. Id. FIG. 3 shows a cell 300 for an LSTM RNN. Line 302 represents the state (S) of the cell and can be thought of as an information highway, where it is relatively easy for information to flow along the state of the cell that does not change. Id. Cell gates 304, 306, and 308 determine how much information is allowed through the state, or along the information highway. Cell gate 304 initially determines how much information is removed from the cell state S _t , the so-called forget gate layer. Id. Cell gates 306 and 306' then determine what information is added to the cell state, and cell gates 308 and 308' determine the predicted

The information highway or cell state is now updated to cell state S _t+1 for use in the next cell. LSTMs allow RNNs to have more persistent or long-term memory. LSTMs provide an additional advantage to RNN-based machine learning models in that the output prediction takes into account context that is separated from the input data by a longer amount of space or time than simpler RNN structures, depending on how the data is ordered.

In some embodiments utilizing RNNs, the primary and secondary time sequences may not be provided to the RNN as vectors at each time step. Instead, the RNN may be provided only with the current values of the primary and secondary time sequences along with future values or aggregation functions of the secondary time sequence within the prediction horizon. In this way, the RNN uses a persistent state vector to hold information about previous values for use in making predictions.

Machine learning is well suited for continuous monitoring of one or more criteria to identify minor or major anomalies or trends in input data compared to training examples used to train the model. Thus, some embodiments described herein input a user's health indicator data and, optionally, other factor data into a trained machine learning model, which predicts what a healthy person's health indicator data will look like in the next time step and compares the prediction to the user's measured health indicator data in future time steps. If the absolute value of the difference (e.g., loss, described below) exceeds a threshold, the user is notified that the user's health indicator data is not within a normal or healthy range. The threshold is a number set by the designer, and in some embodiments, can be changed by the user to allow the user to adjust the notification sensitivity. The machine learning models of these embodiments can be trained on the health indicator data alone, or in combination with corresponding (in time) other factor data from a population of healthy people, or on other training examples to fit the design needs for the model.

Data from health indicators such as heart rate data is sequence data, and more specifically, time sequence data. Heart rate can be measured in a number of different ways, for example, without limitation, by measuring an electrical signal from a chest strap, or derived from a PPG signal. Some embodiments obtain a derived heart rate from a device, where each data point (e.g., heart rate) is generated at approximately equal intervals (e.g., 5 seconds). However, in some cases, in other embodiments, the derived heart rate is not provided at approximately equal time steps, for example because the data required for the derivation is unreliable (e.g., the device has moved, or the PPG signal is unreliable due to light pollution). The same is true for obtaining a secondary sequence of data from a motion sensor or other sensor used to collect other factor data.

The raw signal/data (electrical signal from EEC, chest strap, or PPG signal) itself is a time sequence of data that may be used in accordance with some embodiments. For purposes of clarity, and not limitation, this description uses PPG to refer to data representing health indicators. Those skilled in the art will readily recognize that any form of data relating to health indicators, raw data, waveforms, or numbers derived from raw data or waveforms, may be used in accordance with some embodiments described herein.

Machine learning models that may be used in the embodiments described herein include, by way of example and without limitation, Bayesian, Markovian, Gaussian processes, clustering algorithms, generative models, kernels, and neural network algorithms. Some embodiments utilize machine learning models based on trained neural networks, other embodiments utilize recurrent neural networks, and additional embodiments use LTSM RNNs. For purposes of clarity, and without limitation, recurrent neural networks are used to describe some embodiments herein.

4A-4C show hypothetical plots of PPG versus time (FIG. 4A), steps versus time (FIG. 4B), and temperature versus time (FIG. 4C). PPG is an example of health indicator data, and steps, activity level, and temperature are examples of multifactor data related to other factors that may affect the health indicator data. As will be recognized by those skilled in the art, the other data may be obtained from any of many known sources, including, but not limited to, GPS data, weight scales, user input, and the like, and may include, but is not limited to, temperature, activity (running, walking, sitting, cycling, falling, climbing stairs, walking, etc.), BMI, weight, height, age, and the like. The first dotted line running vertically through all three plots represents the time t at which user data was obtained for input into a trained machine learning model (discussed below). The hashed plot line in FIG. 4A represents predicted or likely output data 402, and the solid line 404 in FIG. 4A represents measured data. Figure 4B is a hypothetical plot of a user's step count at various times, and Figure 4C is a hypothetical plot of temperature at various times.

5A-5B show schematic diagrams of a trained recurrent neural network 500 for receiving the input data shown in FIGS. 4A-4C, i.e., PPG (P), step count (R), and temperature (T). It is emphasized again that these input data (P, R, and T) are merely examples of health indicator data and other factor data. It will also be understood that data regarding more than one health indicator may be input and predicted, and more or less other factor data than two may be used, with the selection depending on what the model is designed for. It will be further understood by those skilled in the art that the other factor data is collected to correspond in time with the collection or measurement of the health indicator data. In some cases, the other factor data, e.g., weight, remains relatively constant over a certain period of time.

FIG. 5A shows the trained neural network 500 as a loop. P, T, and R are input to state 502 of the RNN 500, where weight W is applied, and the RNN 500 outputs a predicted PPG 504 (P ^* ). In step 506, the difference PP ^* (ΔP ^* ) is calculated, and in step 508, it is determined whether |ΔP ^* | is greater than a threshold. If so, step 510 notifies/warns the user that the user's health indicators are outside of the boundaries/thresholds predicted as normal or predicted for a healthy person. The warning/notification/detection can be, for example, but not limited to, a suggestion to see/consult a doctor, a simple notification like haptic feedback, a request to take additional measurements like an ECG, or a simple note without a recommendation, or any combination thereof. If |ΔP ^* | is less than or equal to the threshold, step 512 does nothing. In both steps 510 and 512, the process is repeated with new user data in the next time step. In this embodiment, the state may be updated following output of the predicted data, and the predicted data may be used in updating the state.

In another embodiment not shown, the primary sequence of heart rate data (e.g., derived from a PPG signal) and the secondary sequence of other factor data are provided to a trained machine learning model, which may be an RNN, a CNN, another machine learning model, or a combination of models. In this embodiment, the machine learning model is
A. A vector ( _VH ) of length 300 of the last 300 health indicator samples (e.g., heart rate per minute) up to and including any health indicator data at time t;
At least one vector (V _O ) of length 300 containing the most recent other factor data, e.g., step counts, at the approximate time of each sample in BV _H ;
C. A vector (V _TD ) of length 300 whose entry V _DT (i) at index i contains the time difference between the timestamps of health index samples V _H (i) and V _H (i-1);
The method is configured to receive as input at a reference time t a scalar prediction interval other-factor rate O _rate (e.g., but not limited to, step rate) that represents an average other-factor rate (e.g., step _rate ) measured over a time period from Dt to t+τ, where τ is, for example, but not limited to, 2.5 minutes, a future prediction interval.

The output of this embodiment can be, for example, a probability distribution characterizing the predicted heart rate measured over the time period from t to t+τ. In some embodiments, the machine learning model is trained using training examples that include a continuous time sequence of health indicator data and a multifactor data sequence. In one alternative embodiment, the notification system assigns a timestamp to each predicted health indicator (e.g., heart rate) distribution for t+τ/2, thus centering the prediction distribution within the prediction interval (τ). The notification logic, in this embodiment, then considers all samples within a sliding window (W) of length W _L =2*(τ), or 5 minutes in this example, and calculates the prediction interval (τ) based on three parameters:
1. Average value of all health indicator sequence data in the time window

and,
2. The average of all model predictions for the health indicators whose prediction timestamp falls within the time window

and,
3. The root mean square median of each predicted health indicator distribution within the time window

Calculate where:
4. In one embodiment,

or

A notification is generated if , where Ψ is the threshold.

In this embodiment, an alert is generated if the measured health index is more than a certain multiple of a standard deviation from the mean of the predicted health index values within a certain window W. The window W may be applied in a sliding fashion over the sequence of measured and predicted health index values, with each window overlapping the previous window by a designer-specified portion, e.g., 0.5 minutes.

The notification may take any number of different forms. For example, without limitation, the notification may notify the user to obtain an ECG and/or blood pressure, may instruct a computing system (e.g., a wearable, etc.) to automatically obtain an ECG and/or blood pressure (for example), may notify the user to seek medical attention, or may simply notify the user that a health indicator is not normal.

The selection of _VDT as an input to the model in this embodiment is intended to allow the model to utilize information contained in the variable intervals between health indicator data in _VH , which may result from an algorithm that derives health indicator data from inconsistent raw data. For example, heart rate samples are generated by the Apple Watch algorithm only when it has sufficiently reliable raw PPG data to output a reliable heart rate value, which results in irregular time gaps between heart rate samples. Similarly, this embodiment utilizes a vector (VO) for other factor data that has the same length as the other vectors to handle the different irregular sample rates between the primary sequence (health indicator) and the _secondary sequence (other factors). The secondary sequence is remapped or interpolated to the same time point as the primary time sequence in this embodiment.

Furthermore, in some embodiments, the composition of data from the secondary time sequences presented as input to the machine learning model from future prediction time intervals (e.g., after t) may be modified. In some embodiments, a single scalar value containing the average multifactor data rate over the prediction horizon may be modified using multiple scalar values, e.g., one for each secondary time sequence. Or, a vector of values may be used over the prediction horizon. In addition, the prediction horizon itself may be adjusted. For example, a shorter prediction horizon may provide faster response to changes and improved detection of events with short(er) underlying timescales, but may also be more sensitive to interference from noise sources such as motion artifacts.

Similarly, the output predictions of the machine learning model itself need not be scalar. For example, some embodiments may generate a time series of predictions for multiple times t in the time interval between t and t+τ, and the alerting logic may compare each of these predictions to measurements in the same time interval.

In this prior embodiment, the machine learning model itself may comprise, for example, a seven-layer feedforward neural network. The first three layers may be convolutional layers containing 32 kernels each with a kernel width of 24 and a stride of 2. The first layer may have arrays VH, _VO , and _VTD in three channels as inputs. The last four layers may be fully connected _layers , all utilizing a hyperbolic tangent activation function except for the last layer. The output of the third layer may be flattened into one array for input to the first fully connected layer. The last layer outputs 30 values (mean, variance, and weight for each mixture) that parameterize a Gaussian mixture model with 10 mixtures. The network uses skip connections between the first and third fully connected layers such that the output of layer 6 is summed with the output of layer 4 to generate the input to layer 7. Standard batch normalization may be used for all layers except the last layer, with a decay of 0.97. The use of skip connections and batch normalization may improve the ability to propagate gradients through the network.

The choice of machine learning model can affect the performance of a system. Machine learning model configuration can be divided into two types of considerations. The first is the internal architecture of the model, meaning the choice of the type of model (convolutional neural network, recurrent neural network, generalized nonlinear regression, such as random forests) as well as the parameters that characterize the model's implementation (generally the number of parameters and/or the number of layers, number of decision trees, etc.). The second is the external architecture of the model, the arrangement of the data fed to the model and the specific parameters of the problem the model is being asked to solve. The external architecture is characterized in part by the dimensionality and type of data provided as input to the model, the time range that that data spans, and any pre- or post-processing performed on the data.

Generally speaking, the choice of external architecture is a balance between increasing the number of parameters and amount of information provided as input, which can increase the predictive power of the machine learning model, the available storage and computational power to train and evaluate larger models, and the availability of a sufficient amount of data to prevent overfitting.

Many variations of the external architecture of the model discussed in some embodiments are possible. The number of input vectors can be varied, as well as the absolute length (number of elements) and time covered. Each input vector need not be the same length or cover the same period of time. The data need not be evenly sampled in time, for example, without limitation, one could provide a 6 hour history of heart rate data, where data less than 1 hour before t is sampled at a rate of 1 Hz, data more than 1 hour before t and less than 2 hours before t is sampled at a rate of 0.5 Hz, and data older than 2 hours is sampled at a rate of 0.1 Hz, where t is a reference time.

5B shows the trained RNN 500 deployed. Input data 513 ( _Pt , _Rt , and _Tt ) are input to state ( _St ) 514 at time t, and trained weights 516 are applied. The output of cell ( _Ct ) 518 is the prediction at time t+1.

520, and the updated state S _t+1 522. Similarly, in C _t+1 524, the input data (P _t+1 , R _t+1 , and T _t+1 ) 513′ is input to S _t+1 522, the trained weights 516 are applied, and the output of C _t+1 524 is

523. As above, S _t+1 is the result of updating S _t , so S _t+1 has memory from S _t from the operation in cell (C _t ) 518 in the previous time step. This process continues for n steps, with input data (P _n , R _n , and T _n ) 513″ being input to S _n 530 and trained weights 516 being applied. The output of cell C _t is the prediction 532

In particular, the trained RNN applies the same weights throughout, but importantly, the state is updated from the previous time step, providing the RNN with the benefit of memory from the previous time step. Those skilled in the art will appreciate that the time order of inputting dependent health index data can vary and still produce the desired results. For example, measured health index data from the previous time step (e.g., P _t-1 ) and other factor data from the current time step (e.g., R _t and T _t ) can be input into the state (S _t ) at the current time step, where the model determines the health index at the current time step.

and the health indicators are compared to the measured health indicator data at the current time step to determine whether the user's health indicators are normal or within a healthy range, as described above.

Figure 5C shows an alternative embodiment of a trained RNN to determine whether a user's health indicator sequence data, in this example PPG, is within a healthy person band or threshold. The input data in this embodiment is a linear combination of

where,

is the predicted health index value at time t, and _Pt is the measured health index at time t. In this embodiment, α ranges from 0 to 1 nonlinearly as a function of loss (L), where loss and α are discussed in more detail below. It is worth noting here that when α is close to zero, the measured data _Pt is input to the network, and when α is close to 1, the predicted data

is input to the network to make a prediction at the next time step. Other factor data at time t (O _t ) can also be optionally input.

I _t and O _t are input to state S _t , _which in some embodiments represents predicted health index data at time step t+1.

Probability distribution of (β),

where β _(p*) is the probability distribution function of the predicted health index (P ^* ). In some embodiments, the probability distribution function is the predicted health index value at t+1

As will be appreciated by those skilled in the art, β _(p*) may be sampled using different methods depending on the network designer's goals, which may include taking the mean, maximum, or random sampling of a probability distribution. Evaluating β ^t+1 using data measured at time t+1 provides the probability that state S _t+1 would have predicted given the measured data.

To illustrate this concept, Figure 5D shows a hypothetical probability distribution for the range of hypothetical health indicator data at time t+1. This function is the predicted health indicator at time t+1.

To determine β t+1 , the probability distribution (β ^t+1 ) is also sampled with a maximum probability of 0.95.

to determine the probability that the model would have predicted if real data had been fed into the model. In this example,

is 0.85.

A loss may be defined to help determine whether to notify the user that their health condition is not within the normal range as predicted by the trained machine learning model. The loss is selected to model how close the predicted data is to the actual or measured data. Those skilled in the art will recognize many ways to define the loss. In other embodiments described herein, for example, the absolute value of the difference between the predicted data and the actual data (|ΔP ^* |) is the loss. In some embodiments, the loss (L) may be L=-ln[β _(p) ], where:

where L is a measure of how close the predicted data is to the measured or actual data. β _(p) ranges from 0 to 1, where 1 means the predicted and measured values are the same. Thus, a low loss indicates that the predicted value is likely to be the same or close to the measured value, which in this context means that the measured data appears to be from a healthy/normal person. In some embodiments, a threshold on L is set, e.g., L>5, where the user is notified that the health index data is outside of a range considered healthy. Other embodiments may take an average of the loss over a period of time and compare the average to a threshold. In some embodiments, the threshold itself may be a function of a statistical calculation of the predicted values or the average of the predicted values. In some embodiments, the following formula may be used to notify the user that the health index is not within a healthy range:

(P _range ) is determined by a method for averaging the measured health index data over a time range,

is determined by averaging the predicted health indicator data over the same time range,

is the median of the sequence of standard deviations derived from the network over the same time range,

teeth,

is a function of the standard deviation evaluated at and can act as a threshold.

Averaging methods that may be used include, by way of non-limiting example, the mean, the arithmetic mean, the median, and the mode. In some embodiments, outliers are removed so as not to skew the calculated numerical values.

Input data for the embodiment shown in Figure 5C

Referring to FIG _{. 5E} , α(L) is defined as a function of L and ranges from 0 to 1. For example, α(L) may be a linear or nonlinear function, or may be linear over one range of L and nonlinear over another range of L. In one embodiment, as shown in FIG. 5E, the function α(L) is linear for L between 0 and 3, quadratic for L between 3 and 13, and 1 for L greater than 13. For this embodiment, when L is between 0 and 3 (i.e., when the predicted health index data and the measured health index data are nearly identical), α-1 is close to zero, so that the input data I _t+1 is nearly the measured data _Pt+1 . When L is large, e.g., greater than 13, α(L) is 1, which means that the input data

Let α(L) be the predicted health index at time t+1. When L is between 1 and 13, α(L) varies quadratically, and the relative contribution of the predicted health index data and the measured health index data to the input data also varies. The linear combination of the predicted health index data and the measured health index data weighted by α(L) allows the input data to be weighted between the predicted data and the measured data at any particular time step in this embodiment. In all these examples, the input data may also include other factor data (O _t ). This is just one example of self-sampling, and some combination of the predicted data and the measured data is used as input to the trained network. Those skilled in the art will understand that many others may be used.

The machine learning model in an embodiment uses a trained machine learning model. In some embodiments, the machine learning model uses a recurrent neural network, which requires a trained RNN. By way of example, and not by way of limitation, FIG. 6 shows an unfolded RNN to demonstrate training of the RNN in accordance with some embodiments. Cell 602 has an initial state _S0 604 and a weight matrix W606. Step rate data _R0 , temperature data _T0 , and initial PPG data _P0 at time step zero are input into state _S0 , and weights W are applied to obtain the predicted PPG at the first time step.

is output from cell 602,

is calculated using the PPG (P ₁ ) obtained at time step 1. Cell 602 also outputs an updated state 608 (S ₁ ) at time step 1, which enters cell 610. The step rate data R ₁ , temperature data T ₁ , and PPG data P ₁ at time step 1 are input to state S ₁ and weight 606W is applied to obtain the predicted PPG at time step 2.

is output from cell 610,

is calculated using the PPG ( _P2 ) obtained at time step 2. Cell 610 also outputs an updated state 612 ( _S2 ) at time step 2, which goes into cell 614. The step rate data _R3 , temperature data _T3 , and PPG data _P3 at time step 3 are input into _S2 and weight 606W is applied to obtain a predicted PPG at time step 3.

is output from cell 614,

is calculated using the PPG (P ₃ ) obtained at time step 3. This is because the state 616 at time step n is output,

ΔP ^* ' is used in backpropagation to adjust the weight matrix, similar to training a convolutional neural network. However, unlike a convolutional network, the same weight matrix in a recurrent neural network is applied at each iteration and is only changed in backpropagation during training. Many training examples with health indicator data and corresponding other factor data are input to the RNN 600 many times until the RNN 600 converges. As previously discussed, LTSM RNNs may be used in some embodiments where the state of such a network provides a longer-term contextual analysis of the input data, which may provide better predictions when the network learns (more) long-term correlations. As also mentioned and as those skilled in the art will readily understand, other machine learning models fall within the scope of the embodiments described herein and may include, by way of example and without limitation, CNNs or other feed-forward networks.

7A shows a system 700 that predicts whether a user's measured health indicator is within or outside of a normal threshold range for a health indicator of a healthy person under similar other factors. The system 700 has a machine learning model 702 and a health detector 704. Embodiments relating to the machine learning model 702 include, for example (but not limited to), a trained machine learning model, a trained RNN, a CNN, or other feed-forward network. The trained RNN, other network, or combination of networks may be trained on training examples from a population of healthy people from whom the health indicator data and corresponding other factor data (in time) were collected. Alternatively, the trained RNN, other network, or combination of networks may be trained on training examples from a particular user, resulting in a personalized trained machine learning model. Those skilled in the art will appreciate that training examples from different populations may be selected depending on the application or design for the trained network and system in general. Those skilled in the art will also readily appreciate that the health indicator data in this and other embodiments may be one or more health indicators. For example, but not limited to, one or more of PPG data, heart rate data, blood pressure data, temperature data, blood oxygen level data, etc., may be used to train a model and predict the health of the user. The health detector 704 uses the predictions 708 from the machine learning model 702 and the input data 710 to determine whether the loss or other metric determined by analyzing the prediction output with the measurement data exceeds a threshold considered normal and therefore unhealthy. The system 700 outputs a notification, or a status of the health of the user. This notification may take many forms as discussed herein. The input generator 706 continuously acquires sensory data from a user wearing or in contact with a sensor (not shown), the data being representative of one or more health indicators of the user. Corresponding (in time) other factor data may be collected by another sensor or acquired via other means, as described herein or as would be readily apparent to one of ordinary skill in the art.

The input generator 706 may also collect data to determine/calculate other factor data. The input generator may include, for example, but not limited to, a smartwatch, a wearable or mobile device (e.g., an Apple Watch® or FitBit® smartphone, a tablet, or a laptop computer), a combination of a smartwatch and a mobile device, a surgically implanted device capable of transmitting data to a mobile device or other portable computing device, or a device on a cart in a medical facility. Preferably, the user input generator 706 has a sensor (e.g., a PPG sensor, an electrode sensor) for measuring data related to one or more health indicators. The smartwatch, tablet, mobile phone, or laptop computer of some embodiments may be equipped with a sensor or the sensor may be remotely located (may be surgically implanted, may be in contact with the body away from the mobile device, or may be some separate device), in all these cases the mobile device communicates with the sensor to collect the health indicator data. In some embodiments, the system 700 may be provided on a mobile device alone, in combination with other mobile devices, or in combination with other computing systems via communication over a network through which these devices may communicate. For example, and without limitation, the system 700 may be a smart watch or wearable in which the machine learning model 702 and the health detector 704 are located on the device, e.g., in the memory of the watch or in firmware on the watch. The watch may have a user input generator 706 and communicate with other computing devices (e.g., mobile phones, tablets, laptop computers, or desktop computers) via direct communication, wireless communication (e.g., WiFi, sound, Bluetooth, etc.), or a network (e.g., the Internet, an intranet, an extranet, etc.), or combinations thereof, in which the trained machine learning model 702 and the health detector 704 may be located on the other computing devices. Those skilled in the art will appreciate that any number of configurations of the system 700 may be utilized without departing from the scope of the embodiments described herein.

7B, a smartwatch 712 according to one embodiment is shown. The smartwatch 712 includes a wristwatch 714 that includes all circuitry and microprocessors, as well as processing devices (not shown) known to those skilled in the art. The wristwatch 714 also includes a display 716 on which the user's health indicator data 718, in this example, heart rate data, may be displayed. Also displayed on the display 716 may be a predicted health indicator band 720 for a normal or healthy population. In FIG. 7B, the user's measured heart rate data does not exceed the predicted health band, so in this particular example, no notification would be made. The wristwatch 714 may also include a watchband 722 and a high fidelity sensor 724, for example, an ECG sensor. Alternatively, the watchband 722 may be an expandable cuff for measuring blood pressure. For example, a low fidelity sensor 726 (shown in shading) is provided on the back of the wristwatch 714 to collect the user's health indicator data, such as PPG data, which may be used to derive heart rate data, or other data, such as blood pressure. Alternatively, as will be appreciated by those skilled in the art, in some embodiments, fitness bands such as FitBit or Polar may be used, which have similar processing capabilities and other factor measuring devices (e.g., ppg and accelerometers).

FIG. 8 illustrates one embodiment of a method 800 for continuously monitoring a user's health status. Step 802 receives user input data, which may include data on one or more health indicators (aka a primary sequence of data) and data (corresponding in time) on other factors (aka secondary sequences of data). Step 804 inputs the user data into a trained machine learning model, which may include a trained RNN, CNN, other feed-forward network as described herein, or other neural network known to those skilled in the art. In some embodiments, the health indicator input data may be one or a combination, e.g., a linear combination, of predicted and measured health indicator data, as described in some embodiments herein. Step 806 outputs data on one or more predicted health indicators at a time step, which may include, by way of example and not limitation, a single predicted value, a probability distribution as a function of the predicted value. Step 808 determines a loss based on the predicted health index, where for example, but not by way of limitation, the loss can be a simple difference between the predicted health index and the measured health index, or some other appropriately selected loss function (e.g., the negative logarithm of a probability distribution evaluated at values related to the measured health index). Step 810 determines whether the loss exceeds a threshold considered normal or unhealthy, where the threshold can be, for example, but not by way of limitation, a simple numerical value selected by a designer, or a more complex function of some parameter related to the prediction. If greater than the threshold, step 812 notifies the user that the user's health index exceeds a threshold considered normal or healthy. As described herein, the notification can take many forms. In some embodiments, this information can be visualized to the user. For example, but not by way of limitation, the information can be displayed on a user interface, such as a graph showing (i) the distribution of measured health index data (e.g., heart rate) and other factor data (e.g., step count) as a function of time, and (ii) the predicted health index data (e.g., predicted heart rate values) generated by the machine learning model. In this manner, a user can visually compare the measured data points to the predicted data points and determine by visual inspection whether the user's heart rate falls within the range predicted by, for example, a machine learning model.

Some embodiments described herein refer to using a threshold to determine whether to notify the user. In one or more of these embodiments, the user may change the threshold to adjust or tune the system or method to more closely match the user's personal health knowledge. For example, if the physiological indicator used is blood pressure and the user has a higher blood pressure, an embodiment may frequently alert/notify the user that the user's indicator is outside of a normal or healthy range from a model trained on a healthy population. Thus, certain embodiments allow the user to raise the threshold so the user is not notified less frequently that the user's health indicator exceeds what is considered normal or healthy.

Some embodiments preferably use raw data for health indicators. If the raw data is processed to derive a particular measurement, e.g., heart rate, this derived data may be used according to the embodiments. In some situations, the provider of the health monitoring device does not have control over the raw data, and what is received is processed data in the form of a calculated health indicator, e.g., heart rate or blood pressure. As will be appreciated by those skilled in the art, the form of data used to train the machine learning model should match the format of the data collected from the user and input to the trained model, otherwise predictions may prove to be incorrect. For example, the Apple Watch provides heart rate measurement data at unequal time steps and does not provide raw PPG data. In this example, the user wears an Apple Watch that outputs heart rate data according to Apple's PPG processing algorithm with heart rate data at unequal time steps. The model is trained on this data. Apple deciding to change its algorithm for providing heart rate data may render models trained on data from the previous algorithm unavailable for use on data input from the new algorithm. To account for this potential problem, some embodiments resample irregularly spaced data (such as heart rate, blood pressure data, or ECG data) onto a regularly spaced grid, and sample from the regularly spaced grid when collecting data to train the model. If Apple, or other providers of data, change their algorithms, the model only needs to be retrained on newly collected training examples; the model does not need to be rebuilt to account for the algorithm changes.

In further embodiments, the trained machine learning model may be trained on the user's data, resulting in a personalized trained machine learning model. This trained personalized machine learning model may be used in place of or in combination with the machine learning model trained on a population of healthy people described herein. When used alone, the user's data is input into the personalized trained machine learning model, which outputs a prediction of that individual's health indicators that are normal for that user in the next time step, which prediction is then compared to actual/measured data from the next time step in a manner consistent with the embodiments described herein to determine whether the user's health indicators differed by some threshold from what was predicted to be normal for that user. In addition, this personalized machine learning model may be used in combination with the machine learning model trained on training examples from a population of healthy people to generate predictions and associated notifications related to both what was predicted to be normal for that individual user and what was predicted to be normal for a population of healthy people.

Figure 9A shows a method 900 according to another embodiment, and Figure 9B shows, for illustrative purposes, a hypothetical plot 902 of heart rate (by way of example, not limitation) as a function of time. Step 904 (Figure 9A) receives a user's heart rate data (or other health indicator data) and, optionally, corresponding (in time) other factor data, and inputs this data into a personalized trained machine learning model. In some embodiments, the personalized trained model is trained on the user's individual health indicator data and, optionally, corresponding (in time) other data as described herein. Thus, in step 906, the personalized trained machine learning model predicts normal heart rate data for that individual user under the conditions of the other factors, and step 908 identifies deviations or anomalies in the user's health indicator data compared to what was predicted as normal for that particular user. Some embodiments receive user health indicator data from a wearable device on the user (e.g., Apple Watch, smartwatch, FitBit, etc.) or from another mobile device (e.g., tablet, computer, etc.) that communicates with sensors on the user, as discussed throughout this specification.

A loss may be defined to help determine whether to notify the user in step 908 that the user's measurement data is abnormal relative to what is predicted to be normal for that particular user. The loss is selected to model how close the prediction is to the actual or measured data. Those skilled in the art will recognize many ways to define the loss. In other embodiments described herein and equally applicable here, for example, the absolute value of the difference between the predicted value and the absolute value, |ΔP ^* |, is a form of loss. In some embodiments, the loss (L) may be L=ln[β _(p) ], where:

L is generally defined as
_{β(p) is a measure of how close the predicted data is to the measured data. β(p)} , in this example a probability distribution, ranges from 0 to 1, where 1 means the predicted and measured data are the same. Thus, a low loss, in some embodiments, indicates that the predicted data is more likely to be the same or close to the measured data. In some embodiments, a threshold on L is set, e.g., L>5, where the user is notified that there is an anomalous condition from the predicted condition for that particular user. This notification can take many forms, as described elsewhere herein. As also described elsewhere herein, other embodiments may take an average of the loss over a time period and compare that average to a threshold. In some embodiments, the threshold itself may be a function of a statistical calculation of the predicted data or an average of the predicted data, as described in more detail elsewhere herein. Loss is described in more detail elsewhere herein and will not be discussed further here for brevity. Those skilled in the art will also appreciate that the input and predicted data may be scalar values, or segments of data over a time period. For example, and not by way of limitation, a system designer may be interested in a 5 minute segment of data, inputting all data prior to time t and all other data for t+5 minutes, predicting the health index data for t+5 minutes, and determining the loss between the measured health index data for the t+5 minute segment and the predicted health index data for the t+5 minute segment.

Step 908 determines whether an anomaly exists. As discussed, this may be determined if the loss exceeds a threshold. As previously mentioned, the threshold is set based on the objectives of the system being designed, by the designer's choice. In some embodiments, the threshold may be changed by the user, but in this embodiment, preferably not. If an anomaly does not exist, the process is repeated at step 904. If an anomaly exists, step 910 notifies or alerts the user to obtain a high fidelity measurement, such as, but not limited to, an ECG or blood pressure measurement. At step 912, the high fidelity data is analyzed by an algorithm, a medical professional, or both, and described as normal or not normal, and if not normal, some diagnosis may be assigned, such as AFib, tachycardia, bradycardia, atrial flutter, or hyper/hypotension, depending on the high fidelity measurement obtained. For clarity, it is noted that the notification to record high fidelity data is equally applicable and possible in other embodiments, and in the specific embodiment using the general model described above. The high fidelity measurements may in some embodiments be obtained directly by the user using a mobile monitoring system such as an ECG or blood pressure system, which in some embodiments may be associated with the wearable device. Alternatively, the notification step 910 triggers automatic acquisition of the high fidelity measurements. For example, the wearable device may communicate with a sensor (via a wired or wireless connection) to obtain ECG data, or the wearable device may communicate with a blood pressure cuff system (e.g., a wearable wristband or armband cuff) to automatically obtain blood pressure measurements, or the wearable device may communicate with an implanted device such as a pacemaker or ECG electrodes. Systems for remotely obtaining ECGs are provided, for example, by AliveCor, Inc., such systems include (without limitation) one or more sensors that contact the user at two or more locations, where the sensors collect electrical cardiac data that is transmitted, wired or wirelessly, to the mobile computing device, where an app generates an ECG strip from the data, and the ECG strip may be analyzed by an algorithm, a medical professional, or both. Alternatively, the sensor may be a blood pressure monitor, with blood pressure data being transmitted to the mobile computing device, either wired or wirelessly. The wearable itself may be a blood pressure system with a cuff capable of measuring health indicator data, and optionally with an ECG sensor similar to that described above. The ECG sensor may also include an ECG sensor, such as that described in commonly owned U.S. Provisional Application No. 61/872,555, the contents of which are incorporated herein by reference. The mobile computing device may be, for example, but not limited to, a computer tablet (e.g., iPad®), a smartphone (e.g., iPhone®), a wearable (e.g., Apple Watch), or a device in a medical facility (perhaps mounted on a cart). The mobile computing device may be, in some embodiments, a laptop computer or a computer that communicates with some other mobile device. Those skilled in the art will understand that a wearable or a smart watch would also be considered a mobile computing device in terms of the functionality provided in the context of the embodiments described herein. In the case of a wearable, the sensor may be located on a band of the wearable where the sensor may transmit data wirelessly or wired to a computing device/wearable, or the band may also be a blood pressure monitoring cuff, or both as previously described. In the case of a mobile phone, the sensor may be a pad attached to the phone or separate from the phone that senses electrical cardiac signals and communicates that data wirelessly or wired to a wearable or other mobile computing device. More detailed descriptions of some of these systems are provided in one or more of U.S. Patent Nos. 9,420,956, 9,572,499, 9,351,654, 9,247,911, 9,254,095, and 8,509,882, and one or more of U.S. Patent Application Publication Nos. 2015/0018660, 2015/0297134, and 2015/0320328, all of which are incorporated herein in their entirety for all purposes. Step 912, as previously described, analyzes the high fidelity data and provides an explanation or diagnosis.

In step 914, the diagnosis or classification of the high fidelity measurements is received by a computing system, which in some embodiments may be a mobile or wearable computing system used to collect the user's heart rate data (or other health indicator data), and in step 916, the low fidelity health indicator data sequence (in this example, the heart rate data) is labeled with the diagnosis. In step 918, the labeled user's low fidelity data sequence is used to train a high fidelity machine learning model, and optionally, other factor data sequences are also provided to train the model. The trained high fidelity machine learning model, in some embodiments, receives the measured low fidelity health indicator data sequence (e.g., heart rate data or PPG data) and optionally other factor data, and has the ability to provide a probability or prediction, or to diagnose or detect when the user is experiencing an event that is typically diagnosed or detected using high fidelity data. The trained high fidelity machine learning model is able to do this because it has been trained on the user's health indicator data (and optionally other factor data) labeled with the diagnosis of the high fidelity data. Thus, the trained model has the ability to predict when a user has an event associated with one or more of the labels (e.g., Afib, high blood pressure, etc.) based solely on the measured low fidelity health indicator input data sequence, e.g., heart rate or ppg data (and optionally other factor data). Those skilled in the art will appreciate that the training of the high fidelity model may be performed on the user's mobile device, remotely from the user's mobile device, a combination of the two, or in a distributed network. For example, and without limitation, the user's health indicator data may be stored in a cloud system and this data may be labeled in the cloud using the diagnosis from step 914. Those skilled in the art will readily appreciate any number of ways and manners to store, label, and access this information. Alternatively, a globally trained high fidelity model may be used, which is trained on labeled training examples from a population of people experiencing these conditions that are typically diagnosed or detected using high fidelity measurements. These global training examples provide low-fidelity data sequences (e.g., heart rate) labeled with conditions that are to be diagnosed using high-fidelity measurements (e.g., Afib called from an ECG by a medical professional or an algorithm).

9B, plot 902 shows a schematic diagram of heart rate plotted as a function of time. Anomalies 920 from the user's normal heart rate data occurred at times _t1 , _t2 , _t3 , _t4 , _t5 , _t6 , _t7 , _t8 . Normal means that the expected data for this particular user was within the threshold of measured data, as explained above, and anomalies are outside the threshold. In anomalies from normal, some embodiments prompt the user to obtain a clearer or higher fidelity reading, an ECG reading identified as, for example and without limitation, _ECG1 , _ECG2 , _ECG3 , _ECG4 , _ECG5 , _ECG6 , _ECG7 , _ECG8 . As explained above, the higher fidelity reading may be obtained automatically or the user may obtain it, which may be something other than an ECG, such as blood pressure. The high fidelity readings are analyzed by an algorithm, a medical professional, or both to identify the high fidelity data as normal/abnormal and further to identify/diagnose the abnormality, such as, but not limited to, AFib. This information is used to label health indicator data (e.g., heart rate or PPG data) at the location of the abnormality 920 in the user's sequence data.

The distinction between high-fidelity and low-fidelity data is that high-fidelity data or measurements are typically used to make decisions, detections, or diagnoses, whereas low-fidelity data cannot be readily used for such. For example, an ECG scan may be used to identify, detect, or diagnose arrhythmias, whereas heart rate or PPG data typically do not provide this function. Those skilled in the art will understand that the discussion herein relating to machine learning algorithms (e.g., Bayesian, Markov, Gaussian processes, clustering algorithms, generative models, kernels, and neural network algorithms) applies equally to all embodiments described herein.

In some situations, a user may remain asymptomatic despite the possibility that a problem exists, and even if symptoms are present, it may be impractical to obtain the high fidelity measurements necessary to make a diagnosis or detection. For example, but not by way of limitation, arrhythmias, particularly AF, may not manifest, and even if symptoms do manifest, recording an ECG at that moment is notoriously difficult, and continuously monitoring the user without expensive, bulky, and sometimes invasive monitoring devices is very difficult. As discussed elsewhere herein, AF may at least be a contributing factor to stroke, among other serious conditions, so it is important for a user to understand when they are experiencing AF. Similarly, as discussed elsewhere, AF burden may be of similar importance. Some embodiments allow for continuous monitoring of arrhythmias (e.g., AF) or other serious conditions using only low fidelity health indicator data such as heart rate or ppg, and continuous monitoring of optional other factor data.

FIG. 10 illustrates a method 1000 according to some embodiments of health monitoring systems and methods. Step 1002 receives measured or actual low-fidelity health indicator data of a user (e.g., heart rate or PPG from a sensor on a wearable) and, optionally, corresponding (in time) other factor data that may affect the health indicator data as described herein. As discussed elsewhere herein, the low-fidelity health indicator data may be measured by a mobile computing device, such as a smart watch, other wearable, or computer tablet. In step 1004, the user's low-fidelity health indicator data (and optionally other factor data) are input to a trained high-fidelity machine learning model, which in step 1006 outputs a predicted identification or diagnosis for the user based on the measured low-fidelity health indicator data (and optionally corresponding (in time) other factor data). Step 1008 asks whether the identification or diagnosis is normal, and if so, the process is started over. If the identification or diagnosis is not normal, step 1010 notifies the user of the problem or detection. Optionally, the system, method, and platform may be configured to notify any combination of the user, family, friends, medical professionals, emergency 911, etc. Which of these people to notify may depend on the identification, detection, or diagnosis. If the identification, detection, or diagnosis is life-threatening, certain people may be contacted or notified who may not be notified if the diagnosis is not life-threatening. Additionally, in some embodiments, the measured health indicator data sequence is input into a trained high-fidelity machine learning model, and the amount of time the user is experiencing an abnormal event (e.g., the difference between the start and stop of the predicted abnormal event) is calculated, allowing for a better understanding of the user's abnormal burden. Understanding AF burden may be very important, especially in preventing strokes and other serious conditions. Thus, some embodiments allow for continuous monitoring of abnormal events using a mobile computing device, wearable computing device, or other portable device that can acquire only low-fidelity health factor data and, optionally, other factor data.

FIG. 11 illustrates example data 1100 analyzed based on low fidelity data to generate a high fidelity output prediction or detection, according to some embodiments described herein. Although described with reference to detection of atrial fibrillation, similar data may be generated for additional prediction of a high fidelity diagnosis based on low fidelity measurements. A first chart 1110 illustrates a heart rate calculation over time for a user. The heart rate may be determined based on PPG data or other heart rate sensors. A second chart 1120 illustrates activity data for the user during the same time period. For example, the activity data may be determined based on steps or other measurements of the user's movement. A third chart 1130 illustrates a classifier output from a machine learning model and a horizontal threshold at which a notification is generated. The machine learning model may generate a prediction based on the input of the low fidelity measurements. For example, the data in the first chart 1110 and the second chart 1120 may be analyzed by a machine learning system, as further described above. The results of the machine learning system analysis may be provided as a probability of atrial fibrillation, as shown in chart 1130. If the probability exceeds a threshold, in this case shown as a confidence level of greater than 0.6, the health monitoring system may trigger a notification or other alert to the user, a physician, or others associated with the user.

In some embodiments, the data in charts 1110 and 1120 may be provided to the machine learning system as continuous measurements. For example, heart rate and activity level may be generated as measurements every 5 seconds for accurate measurements. A segment of time with multiple measurements may then be input to the machine learning model. For example, the previous hour of data may be used as input to the machine learning model. In some embodiments, a shorter or longer time period may be provided rather than one hour. As shown in FIG. 11, the output chart 1130 provides an indication of the time period during which the user is experiencing an abnormal health event. For example, the period during which the prediction is above a certain confidence level may be used by the health monitoring system to determine atrial fibrillation. This value may then be used to determine the atrial fibrillation burden in the user during the measured time period.

In some embodiments, the machine learning model for generating the predictive output in the chart 1130 may be trained based on the labeled user data. For example, the labeled user data may be provided based on high fidelity data (such as ECG readings) taken during a time period in which low fidelity data (e.g., PPG, heart rate) and other data (e.g., activity level or step count) are also available. In some embodiments, the machine learning model is designed to determine whether atrial fibrillation was likely present during a previous time period. For example, the machine learning model may take an hour of low fidelity data as input and provide a likelihood that an event was present. Thus, the training data may include several hours of recorded data for a population of individuals. The data may be a health event labeled time when a condition was diagnosed based on the high fidelity data. Thus, if there was a health event labeled time based on the high fidelity data, the machine learning model may determine that any one hour window of low fidelity data with that event input to the untrained machine learning model should provide a prediction of the health event. The untrained machine learning model may then be updated based on comparing the prediction to the label. After several iterations and determining that the machine learning model has converged, the machine learning model may be used by the health monitoring system to monitor the user for atrial fibrillation based on the low-fidelity data. In various embodiments, conditions other than atrial fibrillation may be detected using the low-fidelity data.

12 is a schematic representation of a machine in an exemplary form of a computer system 1200 in which a set of instructions may be executed to cause the machine to perform any one or more of the methodologies discussed herein. In alternative embodiments, the machine may be connected (e.g., networked) to other machines in a local area network (LAN), an intranet, an extranet, or the Internet. The machine may operate as a server or a client machine in a client-server network environment, or as a peer machine in a peer-to-peer (or distributed) network environment. The machine may be a personal computer (PC), a tablet PC, a set-top box (STB), a personal digital assistant (PDA), a mobile phone, a web appliance, a server, a network router, a switch or bridge, a hub, an access point, a network access control device, or any machine capable of executing a set of instructions (sequential or otherwise) that specify actions to be taken by the machine. Moreover, although only a single machine is illustrated, the term "machine" should also be construed to include any collection of machines that individually or collectively execute a set (or sets) of instructions to perform any one or more of the methodologies discussed herein. In one embodiment, computer system 1200 may be representative of a server, mobile computing device, wearable, etc., configured to perform health monitoring as described herein.

The exemplary computer system 1200 includes a processing device 1202, a main memory 1204 (e.g., read only memory (ROM), flash memory, dynamic random access memory (DRAM)), a static memory 1206 (e.g., flash memory, static random access memory (SRAM), etc.), and a data storage device 1218, which communicate with each other via a bus 1230. Any of the signals provided via the various buses described herein may be time-division multiplexed with other signals and provided via one or more common buses. In addition, the interconnections between circuit components or blocks may be depicted as buses or single signal lines. Each of the buses may alternatively be one or more single signal lines, and each of the single signal lines may alternatively be a bus.

Processing device 1202 represents one or more general-purpose processing devices, such as a microprocessor, a central processing unit, or other processing device. More specifically, the processing device may be a complex instruction set computing (CISC) microprocessor, a reduced instruction set computing (RISC) microprocessor, a very long instruction word (VLIW) microprocessor, or a processor implementing other instruction sets, or a processor implementing a combination of instruction sets. Processing device 1202 may also be one or more special-purpose processing devices, such as an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), a digital signal processor (DSP), a network processor, or the like. Processing device 1202 is configured to execute processing logic 1226, which may be an example of a health monitor 1250 and associated systems for performing the operations and steps discussed herein.

The data storage device 1218 may include a machine-readable storage medium 1228 on which one or more sets of instructions 1222 (e.g., software) are stored that embody any one or more of the functional methodologies described herein, including instructions for causing the processing device 1202 to execute the health monitor 1250 and related processes as described herein. The instructions 1222 may reside, completely or at least partially, in the main memory 1204 or in the processing device 1202 during its execution by the computer system 1200, with the main memory 1204 and the processing device 1202 also constituting machine-readable storage media. The instructions 1222 may further be transmitted or received over the network 1220 via the network interface device 1208.

The machine-readable storage medium 1228 may also be used to store instructions for performing methods for monitoring a user's health, as described herein. Although the machine-readable storage medium 1228 is shown in the exemplary embodiment as being a single medium, the term "machine-readable storage medium" should be interpreted to include a single medium or multiple media (e.g., centralized or distributed databases, or associated caches and servers) that store one or more sets of instructions. A machine-readable medium includes any mechanism for storing information in a form (e.g., software, processing application) readable by a machine (e.g., a computer). A machine-readable medium may include, but is not limited to, magnetic storage media (e.g., floppy diskettes), optical storage media (e.g., CD-ROMs), magneto-optical storage media, read-only memory (ROM), random access memory (RAM), erasable programmable memory (e.g., EPROMs and EEPROMs), flash memory, or another type of medium suitable for storing electronic instructions.

The foregoing description sets forth numerous specific details, such as examples of specific systems, components, methods, and the like, to provide a better understanding of some embodiments of the present disclosure. However, it will be apparent to one of ordinary skill in the art that at least some embodiments of the present disclosure may be practiced without these specific details. In other instances, well-known components or methods have not been described in detail or have been presented in simple block diagram form to avoid unnecessarily obscuring the present disclosure. Thus, the specific details described are merely illustrative. Specific embodiments may vary from these illustrative details and still be considered to be within the scope of the present disclosure.

In addition, some embodiments may be practiced in a distributed computing environment in which the machine-readable medium is stored on and/or executed by two or more computer systems. In addition, information transferred between computer systems may be pulled or pushed over a communications medium connecting the computer systems.

Embodiments of the claimed subject matter include, but are not limited to, various operations described herein. These operations may be performed by hardware components, software, firmware, or combinations thereof.

Although the operations of the methods herein are shown and described in a particular order, the order of operations of each method may be changed such that certain operations may be performed in reverse order or such that certain operations may be performed, at least in part, simultaneously with other operations. In alternative embodiments, instructions of separate operations or sub-operations may be intermittent or alternating.

The above description of illustrated implementations of the invention, including those described in the Abstract, are not intended to be exhaustive or to limit the invention to the precise forms disclosed. Specific implementations of the invention and examples thereof have been described herein for illustrative purposes, however, as one skilled in the art will recognize, various equivalent modifications are possible within the scope of the invention. The word "example" or "exemplary" is used herein to mean serving as an example, illustration, or illustration. Any aspect or design described herein as an "example" or "exemplary" should not necessarily be construed as preferred or advantageous over other aspects or designs. Rather, use of the word "example" or "exemplary" is intended to present a concept in a concrete manner. As used in this application, the term "or" is intended to mean an inclusive "or" rather than an exclusive "or". That is, unless otherwise stated or clear from the context, "X includes A or B" is intended to mean any of the natural inclusive permutations. That is, if X includes A, if X includes B, or if X includes both A and B, then "X includes A or B" is satisfied under any of the preceding examples. In addition, the articles "a" and "an" used in this application and the appended claims should generally be construed to mean "one or more" unless otherwise specified or unless the context clearly directs to the singular form. Furthermore, the use of the terms "embodiment" or "one embodiment" or "implementation" or "one implementation" throughout is not intended to refer to the same embodiment or implementation unless so described. Furthermore, the terms "first", "second", "third", "fourth", etc., used herein are intended as labels to distinguish different elements and may not necessarily have a sequential meaning according to their numerical designations.
Variations of the above-disclosed and other features and functions, or alternatives thereof, may be combined into many other different systems or applications. Various presently unforeseen or unanticipated alternatives, modifications, variations, or improvements therein may subsequently be made by those skilled in the art, which are also intended to be encompassed by the following claims. The claims may encompass embodiments in hardware, software, or combinations thereof.

In addition to the embodiments described above, the present disclosure includes, but is not limited to, the following exemplary implementations:

Some example implementations provide a method for monitoring a user's cardiac health. The method may include receiving measured health index data and other factor data of a user at a first time; inputting, by a processing device, the health index data and other factor data into a machine learning model, where the machine learning model generates predicted health index data for a next time step; receiving the user's data for the next time step; determining, by the processing device, a loss for the next time step, where the loss is a measurement between the predicted health index data for the next time step and the user's measured health index data for the next time step; determining whether the loss exceeds a threshold; and outputting a notification to the user in response to determining that the loss exceeds the threshold.

In some example implementations of the method of any example implementation, the trained machine learning model is a trained generative neural network. In some example implementations of the method of any example implementation, the trained machine learning model is a feedforward network. In some example implementations of the method of any example implementation, the trained machine learning model is an RNN. In some example implementations of the method of any example implementation, the trained machine learning model is a CNN.

In some example implementations of the method of any example implementation, the trained machine learning model is trained on training examples from one or more of a healthy population, a cardiac population, and a user.

In some example implementations of any example implementation of the method, the loss at the next time step is the absolute value of the difference between the predicted health index data at the next time step and the user's measured health index at the next time step.

In some example implementations of the method of any example implementation, the predicted health index data is a probability distribution, and the predicted health index data at the next time step is sampled from the probability distribution.

In some example implementations of the method of any example implementation, the predicted health index data at the next time step is sampled according to a sampling technique selected from the group consisting of: predicted health index data at maximum probability, and randomly sampling predicted health index data from a probability distribution.

In some exemplary implementations of the method of any exemplary implementation, the predicted health index data is a probability distribution (β) and the loss is determined based on the negative logarithm of the probability distribution at the next time step evaluated using the user's measured health index at the next time step. In some exemplary implementations of the method of any exemplary implementation, the method further includes self-sampling of the probability distribution.

In some exemplary implementations of the method of any exemplary implementation, the method further includes averaging the predicted health index data over the time step period, averaging the user's measured health index data over the time step period, and determining a loss based on an absolute difference between the predicted health index data and the measured health index data.

In some example implementations of the method of any example implementation, the measured health indicator data includes PPG data. In some example implementations of the method of any example implementation, the measured health indicator data includes heart rate data.

In some example implementations of any of the example implementations of the method, the method further includes resampling the irregularly spaced heart rate data to a regularly spaced grid, where the heart rate data is sampled from the regularly spaced grid.

In some example implementations of the method of any example implementation, the measured health indicator data is one or more health indicator data selected from the group consisting of PPG data, heart rate data, pulse oximeter data, ECG data, and blood pressure data.

Some exemplary limitations provide an apparatus comprising a mobile computing device comprising a processing device, a display, a health indicator data sensor, and a memory having stored instructions, which, when executed by the processing device, cause the processing device to perform the following operations: receiving measured health indicator data from the health indicator data sensor at a time and other factor data at a first time; inputting the health indicator data and other factor data into a trained machine learning model, where the trained machine learning model generates predicted health indicator data at a next time step; receiving the measured health indicator data and other factor data at the next time step; determining a loss at the next time step, where the loss is a measurement between the predicted health indicator data at the next time step and the measured health indicator data at the next time step; and outputting a notification if the loss at the next time step exceeds a threshold.

In some example implementations of any of the example devices, the trained machine learning model includes a trained generative neural network. In some example implementations of any of the example devices, the trained machine learning model includes a feedforward network. In some example implementations of any of the example devices, the trained machine learning model is an RNN. In some example implementations of the method of any of the example implementations, the trained machine learning model is a CNN.

In some example implementations of any of the example devices, the trained machine learning model is trained on training examples from one of a group consisting of a healthy population, a cardiac population, and a user.

In some exemplary implementations of any exemplary device, the predicted health index data is a point prediction of the user's health index at the next time step and is the absolute value of the difference between the predicted health index data and the measured health index data at the next time step.

In some example implementations of any of the example devices, the predicted health indicator data is sampled from a probability distribution generated from a machine learning model.

In some example implementations of any of the example devices, the predicted health index data is sampled according to a sampling technique selected from the group consisting of maximum probability and random sampling from a probability distribution.

In some example implementations of any of the example devices, the predicted health index data is a probability distribution (β) and the loss is determined based on the negative logarithm of β evaluated using the user's measured health index at the next time step.

In some example implementations of any of the example apparatus, the processing device further defines a function α ranging from 0 to 1 and forms a linear combination of the user's measured health indicator data and the predicted health indicator data as a function of α.

In some example implementations of any of the example apparatus, the processing device further performs self-sampling of the probability distribution.

In some exemplary implementations of any of the exemplary apparatuses, the processing device further averages the predicted health index data sampled from the probability distribution over the time step period using an averaging method, averages the measured health index data of the user over the time step period using the averaging method, and defines a loss based on an absolute difference between the averaged predicted health index data and the measured health index data.

In some exemplary implementations of any of the exemplary devices, the averaging method includes one or more methods selected from the group consisting of calculating the mean, calculating the arithmetic mean, calculating the median, and calculating the mode.

In some exemplary implementations of any exemplary device, the measured health indicator data includes PPG data from a PPG signal. In some exemplary implementations of any exemplary device, the measured health indicator data is heart rate data. In some exemplary implementations of any exemplary device, the heart rate data is collected by resampling irregularly spaced heart rate data onto a regularly spaced grid, and the heart rate data is sampled from the regularly spaced grid. In some exemplary implementations of any exemplary device, the measured health indicator data is one or more health indicator data selected from the group consisting of PPG data, heart rate data, pulse oximeter data, ECG data, and blood pressure data.

In some exemplary implementations of any of the exemplary devices, the mobile device is selected from the group consisting of a smart watch, a fitness band, a computer tablet, and a laptop computer.

In some example implementations of any of the example devices, the mobile device further comprises a user high-fidelity sensor, the notification requests the user to obtain high-fidelity measurement data, and the processing device further receives an analysis of the high-fidelity measurement data, labels the user's measured health indicator data with the analysis to generate labeled user health indicator data, and uses the labeled user health indicator data as training examples to train the trained personalized high-fidelity machine learning model.

In some exemplary implementations of any exemplary device, the trained machine learning model is stored on a memory. In some exemplary implementations of any exemplary device, the trained machine learning model is stored on a remote memory, the remote memory is separate from the computing device, and the mobile computing device is a wearable computing device. In some exemplary implementations of any exemplary device, the trained personalized high-fidelity machine learning model is stored on a memory. In some exemplary implementations of any exemplary device, the trained personalized high-fidelity machine learning model is stored on a remote memory, the remote memory is separate from the computing device, and the mobile computing device is a wearable computing device.

In some exemplary implementations of any of the exemplary apparatus, the processing device further predicts that the user is experiencing atrial fibrillation and determines the user's atrial fibrillation burden.

Some example implementations provide a method for monitoring a user's cardiac health. The method may include receiving measured low-fidelity user health indicator data and other factor data at a first time; inputting data including the user health indicator data and other factor data at the first time into a personalized high-fidelity trained machine learning model, where the personalized high-fidelity trained machine learning model predicts whether the user's health indicator data is abnormal; and, if the prediction is abnormal, sending a notification that the user's health is abnormal.

In some example implementations of any example implementation of the method, the trained personalized high-fidelity machine learning model is trained on measured low-fidelity user health indicator data that has been labeled with analysis of the high-fidelity measurement data.

In some example implementations of the method of any example implementation, the analysis of the high fidelity measurement data is based on user-specific high fidelity measurement data.

In some example implementations of the method of any example implementation, the personalized high-fidelity machine learning model outputs a probability distribution, and predictions are sampled from the probability distribution.

In some example implementations of the method of any example implementation, the predictions are sampled according to a sampling technique selected from the group consisting of: predicting at maximum probability, and randomly sampling the predictions from a probability distribution.

In some example implementations of the method of any example implementation, the averaged prediction is determined by averaging the predictions over the time step period using an averaging method, and the averaged prediction is used to determine whether the user's health indicator data is normal or abnormal.

In some example implementations of the method of any example implementation, the averaging method includes one or more methods selected from the group consisting of calculating the mean, calculating the arithmetic mean, calculating the median, and calculating the mode.

In some example implementations of the method of any example implementation, the personalized high-fidelity trained machine learning model is stored in memory of the user's wearable device. In some example implementations of the method of any example implementation, the measured health indicator data and other factor data are time segments of data over a time period.

In some example implementations of the method of any example implementation, the personalized high-fidelity trained machine learning model is stored in a remote memory, which is located remotely from the user's wearable computing device.

In some example implementations, the health monitoring device may include a mobile computing device having a microprocessor, a display, a user health indicator data sensor, and a memory having stored instructions that, when executed by the microprocessor, cause the processing device to perform the following operations: receive measured low-fidelity health indicator data and other factor data at a first time, where the measured health indicator data is acquired by the user health indicator data sensor; input the data including the health indicator data and other factor data at the first time into a trained high-fidelity machine learning model, where the trained high-fidelity machine learning model predicts whether the user's health indicator data is normal or abnormal; and, in response to the prediction being abnormal, send a notification to the user that the user's health is abnormal.

In some exemplary implementations of the health monitoring device of any exemplary implementation, the trained high-fidelity machine learning model is a trained high-fidelity generative neural network. In some exemplary implementations of the health monitoring device of any exemplary implementation, the trained high-fidelity machine learning model is a trained recurrent neural network (RNN). In some exemplary implementations of the health monitoring device of any exemplary implementation, the trained high-fidelity machine learning model is a trained feedforward neural network. In some exemplary implementations of the health monitoring device of any exemplary implementation, the trained high-fidelity machine learning model is a CNN.

In some example implementations of any of the example implementations of the health monitoring device, the trained high-fidelity machine learning model is trained on measured user health indicator data that is labeled based on user-specific high-fidelity measurement data.

In some example implementations of the health monitoring device of any of the example implementations, the trained high-fidelity machine learning model is trained on low-fidelity health indicator data that is labeled based on the high-fidelity measurement data, and the low-fidelity health indicator data and the high-fidelity measurement data are from a population of subjects.

In some example implementations of the health monitoring device of any of the example implementations, the high-fidelity machine learning model outputs a probability distribution and predictions are sampled from the probability distribution.

In some example implementations of the health monitoring device of any example implementation, the predictions are sampled according to a sampling technique selected from the group consisting of: predicting at maximum probability, and randomly sampling the predictions from a probability distribution.

In some example implementations of the health monitoring device of any of the example implementations, the averaged prediction is determined by averaging the predictions over the time step period using an averaging method, and the averaged prediction is used to determine whether the user's health indicator data is normal or abnormal.

In some exemplary implementations of any of the exemplary implementations of the health monitoring device, the measured health indicator data and other factor data are time segments of data over a period of time.

In some exemplary implementations of the health monitoring device of any exemplary implementation, the averaging method includes one or more methods selected from the group consisting of calculating the mean, calculating the arithmetic mean, calculating the median, and calculating the mode.

In some exemplary implementations of the health monitoring device of any exemplary implementation, the personalized high-fidelity trained machine learning model is stored in a memory. In some exemplary implementations of the health monitoring device of any exemplary implementation, the personalized high-fidelity trained machine learning model is stored in a remote memory, the remote memory being located remotely from the wearable computing device. In some exemplary implementations of the health monitoring device of any exemplary implementation, the mobile device is selected from the group consisting of a smart watch, a fitness band, a computer tablet, and a laptop computer.

402 Predicted or likely output data
404 Solid Line
500 Trained Recurrent Neural Networks, Trained Neural Networks, RNN, Deployed Trained RNN
502 Status
504 Predicted PPG
506 Steps
508 Steps
510 Steps
512 steps
513 Input Data
513' Input data
513'' input data
514 State at time t
516 Trained Weights
518 Cells ( _Ct )
520 Prediction at time t+1
522 Updated Status
523 P ^* _t+2
524 C _t+1
530 S _n
532 Predictions
600 RNN
602 cells
604 Initial state S ₀
606 Weighting matrix W, weight
608 Updated state at time step 1
610 Cell
612 Updated state at time step 2
614 Cells
616 State at time step n
700 System
702 Machine Learning Models, Trained Machine Learning Models
704 Health Detector
706 Input Generator, User Input Generator
708 Predictions
710 Input Data
712 Smart Watch
714 Watches
716 Display
718 User health index data
720 Health Index Band
722 Watch Bands
724 High Fidelity Sensor
726 Low Fidelity Sensors
900 Ways
902 Virtual plot, plot
920 Abnormal
1100 Data
1110 First Chart, Chart
1120 Second Chart, Chart
1130 3rd chart, chart, output chart
1200 Computer systems, computing systems
1202 Processing device
1204 Main memory
1206 Static Memory
1208 Network Interface Device
1218 Data storage device
1220 Network
1222 Command
1226 Processing Logic
1228 Machine-readable storage medium
1230 Bus
1250 Health Monitor

Claims (20)

A processing device;
a health indicator data sensor operably coupled to the processing device;
and a memory having instructions stored thereon, the instructions, when executed by the processing device,
receiving measured low fidelity health indicator data and other factor data of a user at a time, the measured low fidelity health indicator data being acquired by the health indicator data sensor;
an operation of inputting a set of data including the low fidelity health indicator data and the other factor data into a trained high fidelity machine learning model, the trained high fidelity measurement machine learning model being trained using the low fidelity health indicator data labeled with high fidelity data and the other factor data, and generating a prediction of whether the user's high fidelity health indicator output is normal or abnormal by comparing a difference between the measured low fidelity health indicator data and predicted low fidelity health indicator data based on the input to a threshold ;
and in response to the prediction being abnormal, transmitting a notification that the user's health condition is abnormal.
Device. The apparatus of claim 1, wherein the trained high-fidelity machine learning model comprises one or more of a trained high-fidelity generative neural network, a trained recurrent neural network (RNN), a trained feedforward neural network, or a trained convolutional neural network (CNN). The device of claim 1, wherein the trained high-fidelity machine learning model is trained on measured user health indicator data labeled with user-specific high-fidelity measurement data. The apparatus of claim 1, wherein the trained high-fidelity machine learning model is trained on low-fidelity health indicator data labeled with high-fidelity measurement data, and the low-fidelity health indicator data and the high-fidelity measurement data are from a population of subjects. The apparatus of claim 1, wherein the high fidelity machine learning model outputs a probability distribution and the prediction is sampled from the probability distribution. The apparatus of claim 5, wherein the predictions are sampled according to a sampling technique selected from the group consisting of: predicting at maximum probability, and randomly sampling the predictions from the probability distribution. The apparatus of claim 5, wherein an averaged prediction is determined by averaging the predictions over a period of time steps using an averaging method, and the averaged prediction is used to determine whether the low fidelity health indicator data for the user is normal or abnormal. The device of claim 1, wherein the device is selected from the group consisting of a smart watch, a fitness band, a computer tablet, and a laptop computer. The apparatus of claim 1, wherein each of the low fidelity health indicator data and other factor data is a time segment of data spanning a time period. The device of claim 1, wherein the low fidelity health indicator data includes a record of a heart rate before the time, the other factor data includes a record of an activity level, and the prediction includes a prediction that the user experienced atrial fibrillation during the record of a heart rate before the time. The processing device further comprises:
receiving a set of training data, the training data including labeled low fidelity health indicator data from a population of individuals and corresponding other factor data from the population of individuals, the labeled low fidelity health indicator data being labeled with corresponding high fidelity data from the population of individuals;
inputting the intervals of the labeled low-fidelity health indicator data and the corresponding other factor data into the trained high-fidelity machine learning model;
and updating the labeled high-fidelity machine learning model by comparing output from the trained high-fidelity machine learning model with the labeled high-fidelity data. receiving, by a processing device, measured low fidelity health indicator data and other factor data of a user at a time, the measured low fidelity health indicator data being obtained by a user health indicator data sensor;
inputting data including the low fidelity health indicator data and other factor data at the time into a trained high fidelity machine learning model by the processing device, the trained high fidelity machine learning model being trained using the low fidelity health indicator data and other factor data labeled with high fidelity data, and generating a prediction of whether the user's high fidelity health indicator output is normal or abnormal by comparing a difference between the measured low fidelity health indicator data and predicted low fidelity health indicator data based on the input to a threshold;
and in response to the prediction being abnormal, sending a notification that the user's health condition is abnormal. The method of claim 12, wherein the trained high-fidelity machine learning model comprises one or more of a trained high-fidelity generative neural network, a trained recurrent neural network (RNN), a trained feedforward neural network, or a trained convolutional neural network (CNN). The method of claim 12, wherein each of the low fidelity health indicator data and the other factor data is a time segment of data spanning a period of time. The method of claim 12, wherein the low fidelity health indicator data includes a record of a heart rate before the time, the other factor data includes a record of an activity level, the prediction includes a prediction that the user experienced atrial fibrillation during the record of a heart rate before the time, and the notification includes instructions to obtain an ECG. The method of claim 12, wherein the trained high-fidelity machine learning model is trained on low-fidelity health indicator data labeled with high-fidelity measurement data, and the low-fidelity health indicator data and the high-fidelity measurement data are from a population of subjects. The method of claim 12, wherein the high fidelity machine learning model outputs a probability distribution and the prediction is sampled from the probability distribution. The method of claim 17, wherein the predictions are sampled according to a sampling technique selected from the group consisting of: predicting at maximum probability, and randomly sampling the predictions from the probability distribution. 18. The method of claim 17, wherein an averaged prediction is determined by averaging the predictions over a time step period using an averaging method, and the averaged prediction is used to determine whether the low fidelity health indicator data for the user is normal or abnormal. receiving a set of training data, the training data including labeled low-fidelity health indicator data from a population of individuals and corresponding other factor data from the population of individuals, the labeled low-fidelity health indicator data being labeled with corresponding high-fidelity data from the population of individuals;
inputting the intervals of the labeled low-fidelity health indicator data and the corresponding other factor data into the trained high-fidelity machine learning model;
and updating the labeled high-fidelity machine learning model by comparing output from the trained high-fidelity machine learning model with the labeled high-fidelity data.

Images