JP2007536645A - Real-time predictive modeling method and apparatus for patients with chronic diseases - Google Patents

Abstract

  Various embodiments of the present invention relate to methods and apparatus for improving patient care (e.g., patients with chronic illness, for example). In one example (this example is intended to be illustrative and not limiting), the present invention may provide its real-time output through a real-time output (eg, periodic, such as hourly, daily, weekly, or monthly). Can be designed to improve such care and prevent the onset of comorbidities associated with many chronic diseases. Various embodiments of the present invention include facilitating data collection, simplification of data transmission, efficient interpretation of chronically ill patient information, providing a time series to form the basis for subsequent analysis, medical practices and By expanding the ability of chronically ill patients or other users of the present invention to understand the relationships between their own lifestyles and / or reducing the workload of healthcare service providers (such as patient physicians) Can improve the care of patients with chronic diseases. In this regard, the present invention would be useful in generating conversations that would otherwise be impossible between a chronically ill patient and a health care provider.

Description

  Various embodiments of the present invention relate to methods and apparatus for improving patient care (e.g., chronically ill patients, etc.). In one embodiment (this example is intended to be illustrative and not limiting), the present invention provides this through real-time output (eg, hourly, daily, weekly, monthly, etc.) By improving such care, it can be designed to prevent the occurrence of co-morbidities associated with many chronic diseases. Various embodiments of the present invention facilitate data collection, simplify data transmission, efficiently interpret information for patients with chronic diseases, provide time series to form the basis for subsequent analysis, Or extend the capabilities of other users of the present invention to understand the relationship between medical practice and the patient's own lifestyle and / or reduce the workload of a healthcare service provider (such as a patient's physician) Can improve the care of patients with chronic diseases. In this regard, the present invention would be useful for generating conversations between chronically ill patients and healthcare providers that would otherwise be impossible.

  In one particular embodiment, the present invention records periodic data generated by users such as chronically ill patients and periodic data generated by additional external sources such as doctors and health care providers. At least one raw data collection device linked to one or more network sites for modeling, storage, analysis, and / or modeling (eg, according to various mathematical computing mechanisms described herein) Can have. In one example (this example is intended to be illustrative and not limiting), the output is for review for television receivers, telephones (fixed line, portable), wireless devices, and It can be provided to the user of the present invention on a regular basis (for example, by a predetermined method or schedule) by a computer or the like, and can be output to various sources such as a doctor (the output is Including, but not limited to, one or many recipients, including, but not limited to, chronically ill patients, patient caregivers, and / or physicians). In one example (this example is intended to be illustrative and not limiting), the output to the physician can be provided in a manner that allows the physician to access a summary of the patient's progress. In another example (this example is intended to be illustrative and not limiting), the user's / patient specification is met by formatting the output of the present invention according to the predictive modeling of the present invention. And can be used to analyze the patient's progress and / or change treatment in the event of an emergency.

  In this application, the term “predictive modeling” is intended to mean a broad mathematical category that includes (but is not limited to) trend detection, variance detection, prediction, and / or prediction. Intended.

  Also, in this application, the term “real time” refers to an essentially simultaneous or interactive process (as opposed to occurring in a relatively slow and “batch” or non-interactive manner). Is meant to mean

  Diabetes is a chronic disease that impairs the body's ability to use food. A hormone called insulin, produced in the pancreas, helps the body use food for energy. In the case of people with diabetes, the pancreas does not produce insulin or the body cannot use insulin correctly. Without insulin, glucose increases in the blood, which is the body's main energy source. About 90-95% (about 16 million) of Americans with diabetes have type 2 diabetes.

  Some of the symptoms of type 2 diabetes (1. frequent urination, 2. excessive thirst and hunger, 3. dramatic weight loss, 4. anger, 5. weakness and fatigue, and 6 Nausea and vomiting) is the same as the symptoms of type 1 diabetes.

  Further symptoms of type 2 diabetes are: 1. skin infections that repeatedly develop or are difficult to cure; 2. Gingival or bladder infection; 3. Visual impairment, 4. tingling and numbness in limbs, and 5. May include itching of the skin.

  Unlike type 1 diabetes, the symptoms of type 2 diabetes usually develop gradually over months or even years, and some people with type 2 diabetes are not aware of themselves. Some people have mild symptoms.

  The cause of diabetes remains a mystery, but researchers have discovered that overweight can trigger the onset of diabetes because the normal functioning of insulin is hindered by excess fat. Type 2 diabetes is usually treated with exercise and a separate diet plan (eg, designed to help maintain healthy weight, control blood glucose levels, and avoid complications). If diet and exercise alone do not reduce blood glucose levels, in addition to diet and exercise, diabetes drugs, insulin, or both may be required.

  Diabetes is usually not curable but can be treated. With proper treatment, daily care, and family support, patients can lead a healthy and active life.

  In this regard, this type of chronic disease has heretofore been usually treated temporarily (only at spaced apart temporary intervals). Specifically, chronically ill patients are scheduled for regular donor intervention with an interval of 3 months, 6 months, or 12 months after being identified as having a particular disease. ing. For example, in the case of diabetes, these interventions have been characterized by the use of the hemoglobin A1C test (a blood measurement that can estimate average blood glucose over the past 90 days). In addition to this study, chronic diabetics measure blood glucose at home and sometimes report the average to their health care providers. By relying on such an average measurement, a chronically ill diabetic physician will attempt to determine whether a particular treatment that he / she has prescribed is working effectively. However, in the case of many chronically ill patients, including those with diabetes, a large number of patients, even though very little change has occurred (or no change has occurred) The condition of patients with chronic diseases is getting worse.

  It is believed that the primary cause of these failures is that chronically ill patients lose their motivation to do what they need to do to control their diabetes. They take their medicines irregularly, eat inappropriate food, and do not exercise enough. A further cause of these failures is that health professionals rely on aggregated data for the development of treatments and that they are overly simple to analyze all available data This can be explained by using statistics.

  For example, relying solely on the daily average of glucose levels can have a dampening effect and can make the glucose of chronically ill diabetics appear to be better controlled than the actual condition (generally) Individual diabetics irregularly provide a single variable (ie, blood glucose) that can be used to focus on the care characteristics of that particular chronically ill patient, but about 65% of all diabetic patients Suffering from high blood pressure, mostly overweight, many not doing enough exercise, and a significant number suffering from hypoxia, and optimizing care Data on all of these important factors that can be used to provide treatments that are available and that function to improve their care (and that are cost-effective for health care systems) Have Chronic disease patients, currently, it's not almost non-existent).

  Finally, it should be noted that various patient monitoring mechanisms have been proposed so far. This example includes the mechanism described in the following patent document.

  US Pat. No. 6,852,080 relates to a system and method for providing feedback to individual patients for automatic remote patient care. More specifically, according to the disclosure of this patent, medical devices with sensors for monitoring the physiological scales of individual patients regularly record the scale sets. The remote patient treats audio feedback as a set of Quality of Life (QoL) measures related to the patient self-assessment indicator. A database collects collected measure sets, identified and collected device measure sets, and QoL measure sets as individual patient care records. The patient care record in relation to other collected device measure sets that the server periodically receives from the medical device a set of identified and collected device measures and a set of QoL measures The set of identified and collected device measures, the set of QoL measures, and the set of collected device measures are analyzed.

  US Pat. No. 6,383,136 relates to health analysis and prediction of abnormal conditions. More particularly, this patent discloses a method for tracking the health status of a patient that includes inputting a plurality of health record signals. Each signal has a record of measurements of predetermined health notification parameters that are considered to be within a normal range related to the patient's health status acquired at various times. Save this health record signal. By processing the stored health record signal, one predicts a possible tendency for a given health parameter to take a value within an abnormal range. This trend provides a notification signal for future anomalies when a predetermined parameter is expected to take a value within an anomalous range.

  In addition to the advantages and improvements disclosed above, other objects and advantages of the present invention will become apparent by reference to the following description taken in conjunction with the accompanying drawings. These accompanying drawings form part of the present specification, include exemplary embodiments of the present invention, and illustrate various objects and features of the present invention.

  DETAILED DESCRIPTION While detailed embodiments of the present invention are disclosed below, it is to be understood that the disclosed embodiments are merely illustrative of the invention that can be implemented in various forms. In addition, each of the examples given in connection with the various embodiments of the present invention is intended to be illustrative and not limiting. Moreover, the attached drawings are not necessarily drawn to scale, and some features may be exaggerated to show details of particular components. Accordingly, specific structural and functional details described herein are not to be construed as limiting, but are representative of those skilled in the art for making various uses of the present invention. It needs to be interpreted only as a basic basis.

  Referring first to FIGS. 1, 9, and 10, there are shown diagrams for systems / methods according to various embodiments of the present invention. In this regard, the present invention can be set up to collect, transfer, store, and / or analyze data. By analyzing the input, the patient and / or other users of the present invention can be notified of the latest analysis and predictive modeling results. In one example, the present invention can include an input / output data management device and a private network center. The present invention provides an Internet computer architecture that provides multiple user access (such as worldwide access) to the present invention through multiple data management devices supported by at least one central private network center site / server computer (or Other communication technologies and / or devices) may be used. Of course, it is possible to construct the invention in an intranet or other type of closed system architecture.

  In either case, the patient's initial input data (e.g., initial input data for a chronically ill patient) can be composed of a measurement record of a health related parameter related to the patient or a combination of measurement records. Data can be input to the present invention through at least one input device (eg, a blood glucose meter, blood pressure monitor, weight scale, oximeter, physical activity monitor, and / or digital camera, etc.). The data can have measurements of various parameters separated in time by a series of time intervals. Data include body fat percentage, electrocardiogram, stress test, CBC (white blood cell count, red blood cell count, hemoglobin, hematocrit, MCV, MCH, MCHC, platelet count), BMP (sodium, potassium, chloride, bicarbonate, anion gap) , Glucose, BUN, creatinine), LFT (albumin, total bilirubin and direct bilirubin, alkaline phosphatase, ALT / GPT, AST / GOT), HDL-P (cholesterol, triglycerides, HDL-cholesterol, LDL-cholesterol), Monotest, iron , TIBC,% Sat, FT4, thyroxine, TSH, total T3, T3 uptake rate, FTI, ferritin, vitamin B12, folic acid, DHEA-F, CANCER AG 125-A, luteinizing hormone, FSH, testostero , Pros Spec Ag, various eye tests such as glaucoma, and other parameter measurements such as but not limited to.

  In other embodiments, users of the present invention may include, but are not limited to, keyboards, scanners, microphones, cameras, wireless and mobile devices, other computer systems, and those through medical sensors and measurement devices. Health record data can be entered directly into the present invention using other types of input devices. Examples of such medical sensors and measuring devices include (but are not limited to) thermometers, sphygmomanometers, and EKG devices. Data signals can include (but are not limited to) results of blood tests, urine tests, and other medical tests.

  In yet another embodiment, stored data related to the measurement and other input related to the patient (eg, a chronically ill patient) includes (but is not limited to) a sheet of printed paper output by the printer. It can be output by various output devices, and can also be output by a facsimile machine, a voice synthesizer, a computer, and other devices.

  Reference will now be made to various data sources that can be used in connection with the present invention. In one example (this example is intended to be illustrative and not limiting), the user can be provided with at least one device (such as a chronically ill patient) for generating time series data. is there. Such one or more devices can include (but are not limited to) a blood glucose meter, blood pressure monitor, weight scale, oximeter, physical activity monitor, and / or digital camera. In one example, the user can obtain measurements by medical practice using one or more devices that would be essentially the same when the patient received conventional care. In another example, the acquired data (such as measurements) does not require the user to record manually or otherwise (eg, the recording can be performed automatically and transparently to the user) Is). Data can be conveyed by communication (eg, wireless technology) to the data management device of the present invention (this will be described in detail later).

  In one example, the present invention can utilize data from at least three sources. In one example (this example is intended to be illustrative and not limiting), the three sources may be as follows:

  First source: This first source can contain user input and can generate the following data (of course, the following measurements, methods, and frequencies are illustrative Intended, not intended to be limiting).

  Second source: This second source can be generated when a patient's medical history (eg, initial input) is input to the present invention from an input data source such as a physician in charge of treatment. With such a source, the information required by the present invention can be updated (eg, periodically). When such patient data is first input to the present invention, a patient history base point can be generated, past test results can be incorporated, and for historical analysis and / or additional input. In addition, patient files can be reviewed by external sources. Similarly, future medical intervention data can be entered into the present invention.

  Third source: This third source of data can have pharmacological compliance / information and can be interviewed and / or other means, eg, the user's pharmacy, health plan, and / or It can be provided by a pharmacy benefits manager (PBM) (such data can be updated (eg, periodically)).

  As described above, various embodiments of the present invention can generate time series for individual patients by using one or more data sources (these time series can be internal or external). Can be referred to). Also, various embodiments of the present invention are independent of data sources (eg, by smoothly performing patient onset and assessment and patient progress surveys in terms of control variables including control of variance). Simple analysis can be provided. Such administrative tasks could be performed through the use of the analysis and predictive modeling of the present invention.

  In another example (this example is intended to be illustrative and not limiting), a link to evaluate patient therapy with multiple time series analyzes generated by the present invention. Can be generated. The extent to which (1) sequences are correlated with each other according to the present invention or (2) sequences are not correlated with each other according to the present invention can provide useful therapeutic input. Also, using system identification methods and utilizing datasets can facilitate the contribution of specific treatment options in terms of precise control variable results (and these datasets are Resulting in a comparison of effectiveness and patient-specific cost effectiveness analysis).

  Next, reference will be made to a data management device (DMD) that manages data related to and / or provided by the user. In one example (this example is intended to be illustrative and not limiting), the DMD is small and can be placed near the user (such as the patient's home or work). It's okay. In another example (this example is intended to be illustrative and not limiting), the DMD may have a server device. The DMD can collect data about and / or generated by the user (such data collection can be performed, for example, wirelessly), and the DMD can be a user's cable, land, or cellular telephone line. The DMD can connect to a private network center (PNC) and provide data at regular intervals, for example.

  In this regard, the PNC can have a mechanism for performing data analysis algorithms, calculations, and / or predictive modeling processes. At this site, all relevant data can be recorded and stored and can be analyzed and / or modeled (for changes in transition and possible variances). By integrating data and information with lifestyle data and externally generated information about patient treatment, for example, it is possible to provide an assessment of progress using time series nonlinear analysis combined with system identification mathematics is there. In one example (this example is intended to be illustrative and not limiting), data transmission and / or storage may be compliant with HIPAA.

  In another embodiment, the flow of information from the user to the PNC and the output from the PNC to various independent and / or concurrent users (eg, periodic output such as hourly, daily, weekly, or monthly). (Access to information does not necessarily have to be limited by time or the number of concurrent users). Such output (eg, from a PNC) that can be used can be provided in a variety of forms (eg, can be printed, such as on a computer monitor, audio, visual). Furthermore, the data can be sent to the user's DMD for access via a medium transmission / communication such as a telephone line. In one example (this example is intended to be illustrative and not limiting) by providing output on a visual display such as a computer monitor or wireless device screen (eg, for activity Providing users with user notifications and feedback for providing output (including advice, recommendations for follow-up, changes in appropriate behavior, and / or interview notifications with physicians as needed).

  In another embodiment, the output can include a computer generated image to communicate with the user. In this specification, such output is often referred to as a “virtual companion” (which is generated, for example, by a computer that is visually displayed and provides a verbal presentation of regular feedback. Image). This presentation will motivate users to continue using the present invention, and the patient's regular patient requirements (including examples such as ophthalmic visits and / or other medical or healthcare related follow-ups) Useful for encouraging compliance.

  Therefore, according to the present invention, compliance and participation by the user can be promoted by appropriately changing the content of the output. It is also possible to provide at least one access point that allows the user to communicate with others (eg, so that they can understand a particular disease and / or treatment).

  Reference will now be made to exemplary models, algorithms, and statistical evaluations available in the context of various embodiments of the present invention (eg, to perform predictive modeling of a particular patient). It should be noted that such exemplary models, algorithms, and statistical evaluations are intended to be illustrative and not limiting. Also, the names applied to these exemplary models, algorithms, and statistical evaluations include the techniques described herein when these names are utilized within the claims of this application. Note also that it is intended to.

(1. Trend detection algorithm)
A trend detection algorithm is a series designed to detect whether a given time series shows a sign of a significant upward trend, a significant downward trend, or no trend within a given time window. You can have a calculation of The input required for this process is time 1. . . Data Y ₁ . . . Time series data defined as Y _T , where Y ₁ represents the first data point and Y _T represents the data point at time T.

(A. Linear model)
This time-series linear model is described as follows.

Y _t = β ₀ + β ₁ t + E _t (1)

Here _Et is auto-correlated random noise, but this fact is not important in this trend detection version. The coefficients β ₀ and β ₁ of this model are estimated using the ordinary least squares method (OLS). This model calculates the best fit to the given response data and calculates the estimated coefficients

Is significantly different from zero, the standard test statistic

Where can be tested by

Is the sample standard deviation given by the regression output. This test statistic τ is distributed according to the t distribution, and the critical value for detecting the tendency can be set to ± 1.96 corresponding to a reliability level of 95%. A value above 1.96 will represent a significant positive trend and a value below −1.96 will represent a significant negative trend. However, the reliability level can be adjusted, in which case the corresponding critical value will change accordingly.

(B. piece-wise linear model)
A further mechanism is to use an algorithm that is more stable in detecting non-linear monotonic trends. This algorithm may be that described by Fried and Imhoff in 2004. This method includes a piecewise linear model instead of the linear model of equation (1) to better characterize the non-linear trend. Moreover, autocorrelation E _t, in order to determine the autocovariance used in the test statistic, autoregressive: are modeled using (AutoRegressive AR) model. Finally, to reduce the bias in the estimator associated with piecewise regression, a “shrinkage estimate” is obtained using the procedures and formulas included in the paper. Although this method differs from the previous procedure, the output of this algorithm still follows a significant positive trend, a significant negative trend, or no trend, according to the specified probability distribution described in the article. Is a test statistic τ for detecting.

(C. Pattern recognition)
A further mechanism for detecting trends is that after explicitly designing possible “trend templates”, a given time series matches the trend shown in one of those templates. It is a judgment. These templates can also include linear or non-linear monotonic or non-monotonic trends and non-trending templates (straight lines). One algorithm for determining matching between a given time series and one template in a predefined template library is described by Haimowitz et al. (1995). The output of such an algorithm includes the most likely match of the trend observed between the trend template and the time series and the estimated time point k at which the trend is estimated to have started.

(2. Distributed detection algorithm)
By using these algorithms, changes in variance in time series data are detected. For example, an increase in data variance can be observed in readings toward the end of a time series that fluctuates strongly, with relatively early readings being confined within a relatively small range. Changes in variance can be considered separately from changes in trend, and data from two periods can have the same average level and different variance.

  Before using an algorithm that searches for changes in the variance of the time series, it may be necessary to detrend the series and perform corrections for autocorrelation. After performing the trend removal and autocorrelation corrections, the residue will be analyzed to detect one or more changes in variance.

(A. ARIMA model)
Data Z ₁ . . . Z _T time series trend removal and correlation correction can be performed using the ARIMA (p, d, q) model. This model is described by the following equation:

Y _t = α ₁ Y _t-1 +. . . _{+ Α p Y tp + e t} + β 1 e t-1 +. . . + Β _{q etq} (2)

Therefore, the “p” term represents the number of autoregressive components to be included (in other words, the amount of time lag included in the model). The “q” term defines the number of moving average components (ie, the number of innovations included). The “d” term defines the desired degree of differentiation. For example, when d = 1, Y _t = Z _t −Z _t−1 , Y _t−1 = Z _t−1 −Z _t−2 (hereinafter the same). If d = 0, {Y ₁ . . . Y _t } = {Z ₁ . . . Z _t }. Presumably, after trend removal, the residue is centered around zero, so by detecting changes in the distribution of the residue, the algorithm essentially changes in the spread of the residue (ie, in other words For example, their dispersion) will be detected.

  The ARIMA model is a representation of a time series, and its estimation method can be implemented by any standard statistical software package with time series capabilities. The parameters (p, d, q) can be set by the investigator according to prior information (or can be entered by experts in the field), for example, the minimum performed by a standard statistical software package It is also possible to recall an automated method for determining the best ARIMA model using the squared prediction error. Within the ability to analyze change in variance, the aim of the ARIMA model is zero mean and variance

A set of residuals X ₁ . . . It is to determine _Xt . These residuals are observed values Y _t and model predicted values.

Is determined by obtaining the difference between. Formally, it is as follows:

(BD statistical method)
(A. Definition of D _k statistics)
formula

Is mean 0 and variance

Uncorrelated random variables X ₁ . . . Let XT be the cumulative sum of the squares of the sequence of _T (CUUMUM) (t = 1, 2,..., T). And the following formula

D _k = C _k / C _T −k / T (4)

Let be the cumulative sum of the squares (and normalized) squared (where k = 1,..., T and D ₀ = D _t = 0).

(B. Detection of a single change in variance)
The D _k statistic can detect changes in the variance of a sequence of uncorrelated random variables. Let k ^* be the value of k realized by max _k | D _k |. The time k ^* that maximizes the D _k statistic is taken as the most likely place of change in variance. This point depends on the variables X ₁ . . . _XT is mean 0 and variance

Can be detected (t = 1... T). The logarithm of the likelihood ratio to test whether there is a change within that period is as follows:

Next, D _{k *} is calculated using the value of k that maximizes the likelihood ratio in equation (5) (ie, k ^* ). If k is fixed, the value of D _k can be described as a function of normal F statistics that tests the equivalence of variance between two independent samples. Specifically, the full set of variables X ₁ . . . Let us consider the X _T. The first sample consists of observations X _i (where i = 1,..., K ^* ),

And the second sample is X _j, where j = k ^* + 1,.

Suppose you have In this case, the null assumption

Against

Is as follows.

_{F Tk *, k * = (} (CT-C k *) / (Tk *) / (C k * / k *)

Therefore, D _{k *} can be expressed as the following equation from the viewpoint of F _{Tk * and k *} .

At a particular _{Dk *} , F _{Tk *, k *} can be calculated, and the significance of the change in variance can be determined at the desired confidence level (usually 95%).

  Other methods using CUSUM can also be utilized, and the D-statistics outlined above should be considered to encompass all of these other procedures as well.

(C. Multiple Change Detection Algorithm)
Thus, if a single point change can exist, the D _k function can be used. However, if multiple change points may exist, an iterative algorithm based on the continuous application of D _k for each segment of the sequence is performed by continuously dividing after detecting possible change points. Need to use. An algorithm for performing such a procedure can be described, for example, in “Use of Cumulative Sums of Squares for Retrospective Detection of Changes of Variation” (Inclan and Tio, Journal of Ath 94).

(C. Non-parametric method (rank statistics))
a. An alternative to using a cumulative sum of squares (CUSUM) to detect changes in a sequence of random variables is to use non-parametric rank statistics. A sequence of consecutive random variables X ₁ . . . Suppose X _t is observed and, in terms of these distributions, it is desirable to detect whether a change in these variables has occurred at time τ. For this procedure, it is not necessary to know one or more distribution functions of these variables. R ₁ . . . R _t is t observations X ₁ ... In a complete sample of T observations. . . Suppose the rank of X _t. Let us define U _{t, T} as

U _{t, T} = 2W _t −t (T + 1) (7)

Here, W _t = ΣR _j (where j = 1... T). Note that the distribution of U _{t, T} is symmetric around zero, with large negative values indicating an increase in mean and large positive values indicating a change in the positive direction.

Then, conforms to the procedures outlined by Pettitt (1971), can be defined distribution aspects of W _t, which, in the case of the null assumption no change, leading to the construction of the test statistic K _T. Let us define K _T as follows:

As a result, the significance probability related to the value k of K _T can be regarded as being given by the following equation.

  By using these formulas, the estimated point of change

And the investigator needs to calculate the input X ₁ . . . Only X _T. The rejection level of p can be set to any level, but a numerical value less than 0.05 is common for the declaration of statistical significance.

Observed variables X ₁ . . . X _T is derivable from any distribution, since it may be the distribution of these residues is not known, by using this algorithm, not correlated from the series when detrended The residue can be analyzed. However, since the residue should not show a trend with time (change in mean) after trend removal, if a change in the distribution is detected, it is a change in variance, i.e. possible It shows that an increase or decrease in the spread of the residue in the range of values has occurred.

  b. A generalization of this test for changes in variance can be performed to detect multiple changes in the distribution of variables. Such a generalization can be found in Meelis et al. (1991), which is a detection of the multiple change algorithm used by Inclan and Tiao (1995) with D statistic as the test statistic. Very similar.

(3. Prediction)
The ability to predict values in a time series involves using any or all past and current information. The exact characteristics of the processes that generate these time series are often unknown and can include complicating factors such as noise, so various models exist to achieve predictions . Predictions fall into two categories. The first category, interpolation, involves a step of predicting time-series past values. Extrapolation, the second category, involves predicting future values, and this second category is also called prediction. The model presented below can be used as appropriate to achieve both objectives.

(A. General ARMAX model)
A general ARMAX (r, m, n) model is a time series Y ₁ . . . An ARMA model (described in Section 2A) with an additional set of “external” covariates from Y _t . This model is described as:

Here, α _i is an automatic regression coefficient, β _j is a moving average coefficient, ε _t is an innovation, and Y _t is a target variable. X (t, k) is an explanatory regression matrix in which each column is a time series, and X (t, k) represents the t th row and the k th column. The input to this model is the time series Y ₁ . . . Y _t and one (or multiple) time series X (1... T, 1... K) used as covariates.

  The parameters of this model can be estimated recursively by an algorithm that uses the least squares method, and this can be done by standard statistical software packages. Next value

When estimating the data, the prediction interpolation of the data at an arbitrary time is obtained by the equation (19)

Can be calculated. Further, according to this equation, extrapolation prediction of the next data point in the time series at t + 1 is possible by using the following equation.

  Then, by using equation (20), the prediction can be extended to the next h values by a build-off of the previously predicted value.

(B. GARCH)
The GARCH (General Automatic Conditional Heterostatic) model is variable with time, and is used to model the distribution of time series data. X ₁ . . . Assume that there is data X _T (where X _t is normally distributed normally (X _{t to} N (0, h)). GARCH (p, q) model describing this data Is as follows:

There are constraints that α ₀ > 0, α _i ≧ 0, and β _j ≧ 0. The parameter “p” determines the number of lag terms of the estimated variance of observations that will be included. The parameter “q” determines the number of lag terms of the squared data observations included. Parameters (p, q) are usually set by the investigator with prior information or entered by experts in the field, but using, for example, the least square prediction error performed by a standard statistical software package It is possible to recall an automated method for determining the best ARIMA model. The only input to this model is the time series X ₁ . . . _Xt .

  The parameters of this model can be estimated recursively using standard statistical packages. value

Predictive interpolation of variances at all points in estimating

Can be calculated by the equation (21). Also, according to this equation, extrapolation prediction of the variance at the time point t + 1 is possible by using the following equation.

  Then, by using equation (22), the prediction can be extended to the next h values by a build-off of the previously predicted value.

(C. Kalman filtering)
The Kalman filter is designed to estimate the state X _t of a discrete time controlled process governed by the following linear stochastic difference equation:

X _t = AX _t−1 + Bμ _t−1 + ω _t−1 (23)

It has a measurement value _Zt that cannot be observed as follows.

Z _t = HX _t + υ _t (24)

The random variables ω _t and υ _t represent process and measurement noise, respectively, which are assumed to be independent, white, and have a normal probability distribution.

p (ω) to N (0, Q) (25)
p (υ) to N (0.R) (26)

  Here, the process noise covariance Q and the measurement noise error R matrix are assumed to be constant.

  Estimated time series true state

Must be calculated using a recursive Kalman algorithm. The recursive Kalman algorithm has two parts. The first part is called the time update stage, and the second part is called the measurement update stage. These equations are as follows.

(Time update type)

  A and B are from equation (1) and Q is from equation (3). Initial conditions (ie

And P _t-1 ), for example, from a test period that does not fall below the 30-day data

(Measurement update type)

The first task in the update phase is to calculate the “Kalman gain”, ie K _t H. The next step is from the observed _Zt and other estimated parameters

Is something to get. The last step is to obtain an estimate of the error covariance P _t. After this is complete, the process resumes with the next time slice. Q and R estimates can be performed “offline” to achieve optimal filter performance by assuming they are performed offline and remain constant.

The only input to this algorithm is the observed time series Z ₁ . . . It is Z _t. The output is an estimated filtered time series

This is because "process noise" has been removed. According to the Kalman filter, first,

Can be predicted.

  And then

Use the expression

By using

Is calculated.

  As a result, by using the equation (32 + 33), the prediction can be expanded to the next h values by build-off of a value predicted in advance.

(D. Markov model)
The time series can also be modeled as a stochastic process (ie, a process whose future is conditionally dependent on the past and present). The simplest form of a stochastic process is a random variable Markov chain. A Markov chain is defined as a set of random variables {X _t } that has the property that the future is conditional and independent of the past when given the present (where index t is 0,1 ... and so on). In other words, it is as follows.

P (X _t = j | X ₀ = i ₀ , X ₁ = i ₁ ,... X _t-1 = i _t-1 ) = P (X _t = j | X _t-1 = i _t-1 ) (34)

According to this framework, a transition matrix can be constructed only based on observation data included in a time series that summarizes the probability of transition from the current state to other possible states within the distribution of X. The ij-th entry of the transition matrix T indicates the probability that X in the state i at time t−1 transitions to the state j at t (ie, P (X _t = j | X _t−1 = i _t−1 )). Represents.

  Based on this framework and transition matrix, includes calculating the probability of reaching a specific state within a specific amount of time, calculating the expected time to reach a specific state, and building a probability distribution around these answers Predictive computer simulation can be performed (but not limited to).

(E. Random walk model)
A further method for predicting values in a time series is to use a random walk model. In the random walk model, it is assumed that the process can be described by the following equation at an arbitrary time point t.

Y _t = Y _t-1 + α (36)

  Here, Y is a random variable, and α is an additional random variable. If Y is continuous, this process is called Brownian motion. If α is non-zero (which shows a constant trend in one direction), the model is called a “random walk with drift”. The determination of whether a trend is included and the constants used to represent the trend can be generated by the investigator. However, it is easy to recall an automated routine that determines the level of tendency to include. For example, a linear trend can first be estimated using OLS.

In practice, the point-by-point prediction at Y _{t + 1} is simply Y _t or “Y _t + some additional constant α” if a trend is considered to exist. However, according to the random walk model, it is possible to obtain the ability to construct a reliability interval surrounding the predicted value. The variance of the forecast error is as follows: . . Estimated by the variance of the error in the prediction from t (ie, Var (Y ₁ ... Y _t ). One step prediction error is simply the actual value Y and its one step prediction

Is the difference between Estimated variance of prediction error

After calculating

Expected by defining as

Construct a confidence interval for. Expected value "n" steps ahead (ie

The reliability interval

Therefore, the reliability interval grows with increasing uncertainty.

(F. Multi-layer neural network)
By using a multi-layer neural network, a nonlinear model capable of classifying data can be implemented. In this process, the parameters governing the nonlinear mapping are learned simultaneously with those governing the linear discriminating means for classification. This eliminates the need to have prior knowledge about the non-linearity present in the model. The back propagation algorithm for training a three-layer neural network is outlined below. This process utilizes gradient descent in error and is a natural extension of the least mean square algorithm. Note that this analysis has been performed for special cases, but it can be easily extended to much more common networks and training protocols.

(A. Definition of 3 layer network)
A three-layer network is composed of an input layer, a concealment layer, and an output layer, which are interconnected with variable weights. Often there is a bias unit connected to a non-input unit.

  Let X be the input d-dimensional vector.

The input vector is presented to the input layer and the output of each input unit is equal to the corresponding component in the vector. Each concealment unit computes a weighted sum of its inputs to form its scalar net activation. In other words, this net activation is the inner product of the inputs with weights in the concealment unit. When the bias unit is expressed as w _0, it can be described as follows.

Subscript i indexes units in the input layer, j indexes units in concealment, and w _ji represents the weight from input to concealment layer in concealment unit j. Each concealment unit emits an output f (net _j ) that is a non-linear function of its activation. Therefore, it can be described as follows.

y _j = f (net _j )

  Similarly, each output calculates its net activation based on the concealment unit signal as follows.

  Here, the subscript k indexes the units in the output layer, and nH represents the number of concealment units.

As a result, the output unit z _k is calculated through the nonlinear function of the net, and the following expression is obtained.

zk = g (net _k )

Thus, the output z _k is a function of the input feature vector X after experiencing multiple weightings and nonlinearities.

(B. Back propagation algorithm)
As with LMS, consider that the training error on the pattern is the sum across the output unit of the squared difference between the desired output t _k given by the teacher and the actual output z _k .

  Here, T and Z are network output vectors of target and length c, and W represents all weights in the network.

  The back propagation learning rule is based on gradient descent. The weights are initialized with random values, then they are changed in a direction that will reduce the error.

ΔW = -η (∂J / ∂W)

  Here, η is a learning rate and merely indicates a relative size in the change in weight. In the form of components, this has the form:

Δw _pq = −η (∂J / ∂w _pq )

  This gradient descent equation requires treatment in a weight space that reduces the criterion function. Criterion function J cannot be negative, so learning will always eventually stop unless it is unusual. The iterative algorithm requires obtaining a weight vector at iteration m and updating it as follows.

W (m + 1) = W (m) + ΔW (m)

  Here, m indexes a specific pattern presentation.

Here, the evaluation of Δw _pq of the three-layer net involves using a chain rule for differentiation. First, consider the weight from concealment to output described as:

∂J / ∂m _kj = (∂J / ∂net _k ) (∂net _k / ∂m _kj ) = − δ _k (∂net _k / ∂m _kj )

  Here, the sensitivity of unit k is defined as follows:

δ _k = − (δJ / δnet _k )

And this describes a scheme in which the overall error changes with unit net activation. By assuming that the activation function is differentiable (this is a reasonable assumption since the activation function is selected), differentiation is performed and the following equation is obtained for the output unit:

δ _k = −∂J / ∂net _k = − (∂J / ∂z _k ) (∂z _k / ∂net _k ) = (t _z −z _k ) g ′ (net _k )

  Note that the derivative is

∂net _k / ∂m _kj = y _j

The learning rule of the weight from concealment to output is described as follows.

Δm _kj = ηδ _k y _j = η (t _k −z _k ) g ′ (net _k ) y _j

  Here we observe the unit learning rules from input to concealment, and again use the chain rules of differentiation and calculate as follows:

∂J / ∂w _ji = (∂J / ∂y _j ) (∂y _j / ∂net _j ) (∂net _j / ∂w _ji )

The first term on the right hand side is associated with all weights m _kj and is therefore written as:

  Similar to the differentiation of the unit from concealment to output, the sensitivity of the concealment unit is defined as follows:

By using the fact that the sensitivity in the concealment unit is simply weighted by the concealment-to-output weight m _kj , both being the sum of the individual sensitivities in the output unit multiplied by f ′ (net _j ). The learning rule of weight from input to concealment is described as follows.

  With the two learning rules found for each set of weights, all that remains is to define the training protocol and implement learning. Since we are only interested in supervised learning and will not handle excessive amounts of data, we will use batch training (in this case, all Pattern is presented to the network).

(C. Batch back propagation)
Dor r ← r + 1
m ← 0; Δw _ji ← 0; Δm _kj ← 0
Do m ← m + 1
Xm ← m-th input vector is selected Δw _ji ← Δw _ji + ηx _i η _j ; Δm _kj ← Δm _kj + ηδ _k y _j
Until m = nH
w _ji ← w _ji + Δw _ji ; m _kj ← m _kj + Δm _kj
Unitl ‖ ▽ J (W) ‖ <θ
Return W

(G. Extreme value analysis)
Classification of time series data into an appropriate distribution is essential for determining extreme risk. The process of determining this risk by fitting the distribution is called extreme value analysis. This is important because extreme values are often associated with critical situations. Data misclassification has the potential to lead to an underestimation of the risk of reaching critical conditions.

  Often, a normal distribution is assumed as the underlying distribution of the data. Since this has the property of being completely symmetric, this assumption leads to a potentially misleading conclusion regarding the behavior of the data. Most physical data do not fit the normal distribution symmetry and are either positively or negatively biased due to some natural bounds on the data values (ie blood glucose is not positive) Must not). The method of extreme value analysis is to fit a distribution that obtains its true shape to the data so that an appropriate assessment of risk can be performed.

  For example, in the case of blood glucose readings, there are a number of distributions that fit into the general shape of these readings (ie, they have a lower limit and a large upper tail). These include (but are not limited to) Gumbel, Gamma, log-Pearson, Weibull, and lognormal distributions. There are two stages in applying extreme value analysis to blood glucose measurements: parameter fitting and fitness testing. In parameter fitting, when all the data is applied, it finds the most promising parameters for a given distribution, which is performed separately for different distributions. In the goodness-of-fit test, it is tested how well the data fits to each of the assumed distributions using the calculated parameters.

(A. Parameter fitting)
There are many methods for parameter fitting, but as an example, focus on maximum likelihood estimation. In maximum likelihood estimation, the parameters are considered as quantities whose values are fixed but unknown. The best estimate of those values is defined as maximizing the probability of obtaining a sample that is actually observed. In contrast, in another procedure using the Bayesian method, the parameters are considered random variables with some known previous distribution. Next, this is converted into a posterior density according to the observed value of the sample, and the estimated value of the true value of the parameter is updated. However, this procedure basically generates almost the same estimated value as the maximum likelihood estimation, and therefore the description thereof is omitted here.

(Maximum likelihood)
A set of n independent vectors comes from the set D of training samples. Here it is assumed that p (x | w) (where xεD) has a known parametric form and is therefore uniquely determined by the value of the parameter vector θ. Since the samples are derived independently, the following equation is obtained.

  Let p (D | θ) be the likelihood of θ for a set of samples. Maximum likelihood estimation of θ is a value that maximizes p (D | θ) by definition.

It is. If p (D | θ) has good behavior and is differentiable, this can be found by standard methods of differentiation. For analysis, it is often easier to work with the logarithm of likelihood than the likelihood itself. Since the logarithm monotonically increases, the likelihood of the logarithm is maximized

Also maximizes the likelihood.

(B. Compatibility test)
Test whether a sample of data came from a population with a specific distribution by using the chi-square fit test. A chi-square suitability test is applied to binned data (ie, data sorted into multiple classifications). In the case of non-binned data, this is not really a limitation, since a histogram or frequency table can simply be calculated before generating the chi-square test. However, the value of the chi-square test statistic depends on the method of binning the data. Therefore, first, by using the Lloyd algorithm, an optimum binning is found based on the number of samples and the sample value. For reference, the Lloyd algorithm is referred to “Least-Squares Quantization in PCM” (Lloyd, IEEE Transactions on Information Theory 28 (2): 129-137, 1982).

(Chi-square compatibility)
The chi-square test is defined under the assumption that the data conform to the specified distribution. For the calculation of test statistics, the data is divided into k bins and the test statistics are defined as follows:

Here, O _i is the observation frequency in bin i, and E _i is the expected frequency in bin i. Expected frequency is calculated by _{E i = n (F (Y} u) -F (Y l)), where, F is a cumulative distribution function of the distribution of the test object, Y _u is the upper limit in the classification i , Y _l is the lower limit in category i, and n is the sample size.

  The test statistic is approximately compliant with a chi-square distribution with degrees of freedom (k−c), where k is the number of bins and c is (including location and scale parameters and shape parameters) ) Expected distribution parameter +1. For example (this example is intended to be illustrative and not limiting), in the case of a three parameter Weibull distribution, c = 4.

  Therefore, the assumption that the data is from a population with a specified distribution is rejected if:

X ² > X ² _{(α, kc)}

Here, X ² _{(α, kc)} is a chi-square percent point function having degrees of freedom (k−c), and the significance level is α.

  In another embodiment of the invention, the mathematical analysis results in descriptive output based on advanced analysis of patient data (eg, chronically ill patient data). In one example, the predictive modeling descriptive output is forwarded to the DMD via media transmission / communication means for access by the user. In yet another embodiment, the output generated by the predictive modeling of the present invention can be comprised of a visual display that provides feedback to the patient, such as advice for physical activity, follow-up therapy. Recommendations, changes in appropriate behavior, and / or contact with medical service providers as needed (eg, ophthalmic visits and medical follow-ups), including but not limited to ).

  Predictive modeling can provide output for a number of sources such as patients and / or physicians. Interactive conversations are possible between various input and output sources. Data input for the present invention can be organized and analyzed based on advanced analysis.

  Various embodiments of the present invention may be patient-independent, provide data transmission, promote patient compliance and / or for real-time intervention (eg, in the care of patients with chronic diseases) Real-time data can be provided. The present invention can provide periodic feedback that is accessible to all, multiple output locations, based on the results of dynamic algorithms and predictive modeling.

  By improving the care of patients with chronic illnesses, the data for patients with chronic illnesses can be broken down into components and identified through predictive modeling to identify specific forms of chronic illnesses that individually characterize patients with chronic illnesses. The complexity associated with the treatment of can be reduced. The present invention can generate decision making tools that are used quantitatively to improve the problems of chronically ill patients or other users of the present invention by utilizing individual patient data. The present invention can identify the basis of analysis appropriate for an individual patient and compare the most appropriate method for that chronically ill patient with conventional statistical methods. Such a comparison can identify the amount of benefit associated with this correct tool. Such a comparison can provide a unique output for several sources, such as a physician, who can determine the care available to a particular chronically ill patient.

  In another embodiment, the present invention can perform statistical analytical tests on a general basis for prediction. The predictive model of the present invention may be based on a suitable generator function and by using it can control the potential for tragic medical events. After a certain period, the present invention can generate sufficient data to improve the statistical basis and update the decision instructions. The estimate can be updated continuously by using the data in essence. Since a plurality of time series can be used, an external analysis (that is, a stage relating one time series to another) is possible, and the effectiveness of the treatment can be confirmed.

  As described above, in one embodiment of the present invention, the technology has the ability to generate data without the need for wired (or intrusive) devices, Data generation can be performed at the patient's home, office, or travel without the need for manual data entry by the caregiver, freeing patients with chronic illness from the duties associated with data transmission . The generated data can be input (eg, wirelessly to the DMD and / or PNC) without requiring a direct physical connection. For example, within the PCN, a series of mathematical, statistical, and / or modeling procedures can be utilized to collect information about the current and predicted health status of the patient from the patient data.

  In this regard, reference will now be made to the application of the various models, algorithms, and statistical evaluations described above for patient data collection and monitoring. It should be noted that such application examples are intended to be illustrative and not limiting.

(1. Trend analysis)
In this example, trend analysis can be performed on four measurements: blood glucose, blood pressure, physical activity, and weight. This analysis can be realized by modeling the trend as linear, as in Section 1A above. This trend detection can be performed, for example, based on t consecutive days of blood glucose data collected before breakfast. This data is converted to X ₁ . . . Let us call the X _t. This data is the only input to the model described in equation (1). The output from the algorithm is a test statistic τ, which is distributed according to the t probability distribution, a well-known studied statistical distribution. If τ is above the reference value of 1.96 (which corresponds to a confidence level of 95%), within the past t days in the patient's blood glucose level at breakfast the “significant” "There seems to be some evidence." This significance test can be performed automatically according to a predefined confidence threshold. As a result, the output is finally a numerical value such as (-1, 0, 1), which indicates a significant negative trend, no significant trend, or a significant positive trend, respectively. Yes.

Further (or alternatively), the trend uses the algorithm described in Section 1B above (ie, that used by Fried and Imhoff) to model the model as piecewise linear. It is also possible to detect by. This trend detection can be performed, for example, based on t consecutive days of blood glucose data collected before breakfast. This data is converted to X ₁ . . . Let us call the X _t. This data is the only input to the model. The number k of “sections” in the model will be determined by the investigator. The output from this algorithm is the test statistic τ, which is distributed according to the probability table described in the article by Fried and Imhoff. If τ is above a critical value corresponding to a 95% confidence level, there is likely “significant” evidence of a trend in the past t days of the patient's blood glucose level at breakfast . This significance test can be performed automatically according to a predefined confidence threshold. As a result, the output finally becomes a numerical value such as (−1, 0, 1), which indicates a significant negative tendency, no significant tendency, or a significant positive tendency, respectively.

Further (or alternatively), the trend can be detected by using the pattern recognition algorithm described in Section 1C above. This trend detection can be performed, for example, based on t consecutive days of blood glucose data collected before breakfast. This data is converted to X ₁ . . . Let us call the X _t. This data is the only input to the model. A professional investigator has pre-generated trend templates that describe possible linear monotonic, non-linear monotonic, linear non-monotonic, non-linear non-monotonic trends, and no trend, each assigned a number. It is done. The output from this algorithm is a number that represents the most promising template. If the selected template is not a template that means no trend, it is likely that there is “significant” evidence of a trend in the past t days of the patient's blood glucose level at breakfast. This significance test can be performed automatically according to a predefined confidence threshold. As a result, the output finally becomes a numerical value such as (−1, 0, 1), which indicates a significant negative tendency, no significant tendency, or a significant positive tendency, respectively.

  In either case (i.e., regardless of the trend detection algorithm used), the output of the algorithm will then be used to generate results for the patient. Continuing with the blood glucose example, if the algorithm's output over the past t days of data on blood glucose before breakfast shows a significant positive trend for this patient in the system ( For example, the flag will automatically rise (on the day the trend is reported).

  This process depends on blood glucose measured before or after lunch, before or after dinner, at bedtime, or on a daily average; blood pressure (systole and / or diastole); physical activity; and / or It can be used for any input including (but not limited to) body weight. Also, possible outputs showing a significant negative trend, no significant trend, or a significant positive trend may be the same for some or all of the aforementioned inputs, and the same test statistics and probabilities It can be determined by the distribution. Furthermore, the time “window” . . T can also encompass any length of time (eg, greater than 30 data points).

  An analysis of any or all of the listed inputs can be used to report or otherwise communicate reports or communications to one or more of the parties (patient physician, patient caregiver, and / or patient: but not limited to) It can be used in the form.

(A. Physical activity)
Physician report: the results of physical activity trend analysis are in charts, graphs and / or text reports generated by the system for review by the physician to better inform the physician about the patient's condition It can be used. It is also possible to generate a warning that suggests the physician to take immediate action on the patient by using consecutive periods of negative trends.

  Caregiver report: This content may be similar to a physician report and includes charts, graphs, and / or text reports, but especially to improve caregiver interaction with the patient. Adapted to caregiver preference. It is also possible to send a warning to the caregiver based on a continuous period of negative trends.

  Patient report: The patient report can include all of the content given to the doctor and caregiver when requested by the patient. Furthermore, the analysis can also provide automatic generation of motivational or celebration content, suggestions, and / or queries that can eventually include interactive components.

(B. Weight)
Physician Report: The results of weight trend analysis can be used in charts, graphs, and / or text reports generated by the system. In one example, the alert can be generated by using a continuous period of positive trends (ie, weight gain).

  Caregiver report: This content may be similar to a physician report and includes charts, graphs, and / or text reports, but is specifically adapted to caregiver preferences. In one example, it is also possible to send a warning to the caregiver based on a continuous period of positive trends (ie weight gain).

  Patient report: The patient report can include all of the content given to the doctor and caregiver when requested by the patient. Furthermore, the analysis can also provide automatic generation of motivational and celebration content, suggestions, and / or queries that can eventually include interactive components.

(C. Blood pressure)
Physician report: Trend analysis results in blood pressure can be used in charts, graphs and / or text reports generated by the system. Trend analysis may be able to drive content and alerts based on various permutations of cardiac contraction and diastole that have an upward trend, a downward trend, or no change. Such analysis can also contribute to a range of knowledge that drives changes in patient medication and / or other treatment by the physician.

  Caregiver report: This content may be similar to a physician report and includes charts, graphs, and / or text reports, but is specifically adapted to caregiver preferences. Trend analysis may be able to drive content and alerts based on various permutations of cardiac contraction and diastole that have an upward trend, a downward trend, or no change.

  Patient report: The patient report can include all of the content given to the doctor and caregiver when requested by the patient. Furthermore, the analysis can also provide automatic generation of motivational or celebration content, suggestions, and / or queries that can eventually include interactive components.

(Glucose)
Physician report: The results of trend analysis on glucose can be used in charts, graphs, and / or text reports generated by the system. Trend analysis may be able to drive content and alerts based on various permutations of measurements such as before and after breakfast, before and after lunch, before and after dinner, at bedtime, and daily average . Such analysis can also contribute to a range of knowledge that drives changes in patient medication and / or other treatment by the physician.

  Caregiver report: This content is similar to a physician report and includes charts, graphs, and / or text reports, but is specifically adapted to caregiver preferences. Trend analysis may be able to drive content and alerts based on various permutations of measurements such as before and after breakfast, before and after lunch, before and after dinner, at bedtime, and daily average .

  Patient report: The patient report can include all of the content given to the doctor and caregiver when requested by the patient. Furthermore, the analysis can also provide automatic generation of motivational or celebration content, suggestions, and / or queries that can eventually include interactive components.

(2. Detection of changes in dispersion)
In this example, changes in variance can be detected in the time series of blood glucose or blood pressure. Correction for time series trend removal and autocorrelation of any measurement can be performed using the ARIMA model described in Section 2A above. The residue from this model (which is described in equation (3)) is then used as input to the algorithm to detect changes in variance. This modeling can be performed, for example, based on t consecutive days of blood glucose data collected before breakfast. This data is represented by Z ₁ . . . Let us call the Z _t. This data is the only input to the ARIMA (p, d, q) model. The parameters (p, d, q) are determined based on an analysis based on a prior population. In another embodiment, the parameters (p, d, q) are determined by experts in the field of diabetes and / or hypertension. In another embodiment, the parameters (p, d, q) are automatically determined by a search algorithm that finds the most effective value. The output from the algorithm is a series of data X ₁ . . . _Xt .

The data X ₁ . . . _Xt is used as input to the distributed detection algorithm. In one embodiment, the analysis to detect single or multiple changes is performed by a change in the distributed detection algorithm described in Section 2B above. The output from the algorithm is a test statistic T, which is distributed according to the F probability distribution, a well-known studied statistical distribution. If T is above a critical value corresponding to a 95% confidence level and an appropriate degree of freedom, “significant” evidence of a change in variance over the past t days in the patient's blood glucose level at breakfast Is considered to exist. This significance test can be performed automatically according to a predefined confidence threshold. This ultimately results in one aspect of the output being a numerical value such as, for example, (-1, 0, 1), which is a significant decrease in variance, no significant change in variance, or significant variance. Each indicates an increase. The second aspect of the output will be a value k ^* that indicates the most likely time of change of variance in the period between 1 and t.

In addition (or alternatively), it is possible to detect single or multiple changes in variance using the algorithm described in Section 2C above, which utilizes non-parametric statistics. ing. The output from this algorithm is the test statistic _KT . The significance of this statistic given by the value p can be calculated using the formula given by equation (9). If p is below 0.05, it is considered that there is “significant” evidence of a change in variance in the past t days of the patient's blood glucose level at breakfast. This significance test can be performed automatically according to a predefined confidence threshold. This ultimately results in one aspect of the result, for example, a numerical value such as (-1, 0, 1), which can be a significant decrease in variance, no significant change in variance, or significant in variance. Each indicates an increase. The second aspect of the output is a value indicating the most likely time of change of variance within the period 1 to t

Will.

  In either case (i.e., regardless of the distributed detection algorithm used), the output of the algorithm will then be used to generate results for the patient. Continuing with the blood glucose example, if the algorithm results over the last t days of data on blood glucose before breakfast show a significant change in the variance, there will be a change in variance within the system. The reported date and / or the date on which the variance change is estimated to have occurred (ie k * or

) Will automatically be flagged for this patient.

  This process includes (but is not limited to) glycemic glucose measured before or after lunch, before or after dinner, at bedtime, or according to daily average; or blood pressure (cardiac contraction and / or diastole) Can be used for any input). Also, the possible outputs indicating a significant decrease in variance, no change in variance, or a significant increase in variance may be the same for some or all of the aforementioned inputs, and these are Will be determined by the same test statistics and probability distribution. Furthermore, the time “window” . . T can include any length of time (eg, greater than 30 data points).

  Detection of a change in any or all of the listed inputs can be a report to one or more of the parties (patient physician, patient caregiver, and / or patient: but not limited to) or other Note that it can be used in the form of communication. Such reports can include (but are not limited to) the following:

(A. Blood pressure)
Physician Report: The results of the variance change detection algorithm in blood pressure can be used in charts, graphs, and / or text reports generated by the system. Variance change analysis may be able to drive content and alerts based on various permutations of systolic and / or diastole readings that change upward, downward, or unchanged. Such analysis can also contribute to a range of knowledge that drives changes in patient dosing and / or other therapies by the physician.

  Caregiver report: This content is similar to a physician report and includes charts, graphs, and / or text reports, but is specifically tailored to caregiver preferences. Variance change analysis may be able to drive content and alerts based on various permutations of systolic and / or diastole readings that change upward, downward, or unchanged.

  Patient report: The patient report can include all of the content given to the doctor and caregiver when requested by the patient. Furthermore, the analysis can also provide automatic generation of motivational and celebration content, suggestions, and / or queries that can eventually include interactive components.

(B. Glucose)
Physician report: The results of the variance change detection algorithm in glucose can be used in charts, graphs and / or text reports generated by the system. Variance change analysis could drive content and alerts based on various permutations of measurement results before and after breakfast, before and after lunch, before and after dinner, at bedtime, and / or daily averages. It's okay. Such analysis can also contribute to a range of knowledge that drives changes in patient dosing and / or other therapies by the physician.

  Caregiver report: This content is similar to a physician report and includes charts, graphs, and / or text reports, but is specifically adapted to caregiver preferences. Variance change analysis drives content and alerts based on various permutations of average measurement results before and after breakfast, before and after lunch, before and after dinner, at bedtime, and / or daily average measurements It may be possible.

  Patient report: The patient report can include all of the content given to the doctor and caregiver when requested by the patient. Furthermore, the analysis can also provide automatic generation of motivational or congratulatory content, suggestions, and / or queries that can eventually include interactive components.

(3. Prediction)
In this example, predictions of future values can be performed in the time series of blood glucose or blood pressure. Any time series of measurements can be predicted using the method of Section 3A above. This prediction can be performed, for example, based on t consecutive days of blood glucose data collected before breakfast. This data is converted to Y ₁ . . . Let's call it Y _t . There are also data on blood glucose measured before or after lunch, before or after dinner, and at bedtime; blood pressure in both cardiac contraction and diastole; physical activity; and other measurements including weight Let's assume that This is generated in the data matrix X (t, k) (in this example, k = 9 since there are nine additional time series). This data is input to the ARMAX (r, m, n) model. In one example, the parameters (r, m, n) are determined based on an analysis based on a prior population. In another example, the parameters (r, m, n) are determined by experts in the field of diabetes and / or hypertension. In another example, the parameters (r, m, n) are automatically determined by a search algorithm that finds the most effective value. The output from the algorithm is a series of data representing the first h expected values

It is.

In addition (or alternatively), any time series of measurements can be predicted using the method of Section 3B above. However, before using the GARCH model, the inputs X ₁ . . . A X _t, it has substantially zero mean to be dispersed to the normal state with a temporally possibly different dispersion. In one example, this type of series is generated from blood glucose or blood pressure measurements by performing corrections for trend removal and autocorrelation using the ARIMA model described in Section 2A above. Is possible. The residue from this model described in equation (3) is then used as input to the GARCH (p, q) model. This prediction can be performed, for example, based on t consecutive days of blood glucose data collected before breakfast. This data is represented by Z ₁ . . . Let us call the Z _t. This data is the only input to the ARIMA (p, d, q) model. The parameters (p, d, q) are determined based on an analysis based on a prior population. In another example, the parameters (p, d, q) are determined by experts in the field of diabetes and / or hypertension. In another example, the parameters (p, d, q) are automatically determined by a search algorithm that finds the most effective value. The output from the algorithm is a series of data series X ₁ . . . X _t , which represents an uncorrelated trend-removed residue that should be normally distributed with zero mean. Then, data X ₁ . . . GARCH the X _t is described in Section 3B described above (p, q) is used as input to the model (the parameters p and q, ARIMA (p, d, q ) to differ from those used in the model May be). The output from the algorithm is a series of data representing the first h expected values of the trended series

It is. These values can be added to the pre-modeled trend to produce an estimate of Z _{t + h} that represents the reading for the next h days of blood glucose reading before breakfast. .

Further (or alternatively) the prediction can be performed using the Kalman filter equation contained in Section 3C above. This prediction can be performed, for example, based on t consecutive days of blood glucose data collected before breakfast. This data is converted to Y ₁ . . . Let's call it Y _t . There are also data on blood glucose measured before or after lunch, before or after dinner, and at bedtime; blood pressure in both cardiac contraction and diastole; physical activity; and other measurements including weight Let's assume that This data can be contained in a matrix □ _t , which takes into account possible external factors. This data is input to the Kalman filter model. The output from the algorithm is a series of data representing the expected value of the first h days of the patient's blood glucose reading before breakfast

It is.

Additionally (or alternatively), the prediction can be performed using a computer simulation procedure based on the Markov process model described in Section 3D above. This prediction can be performed, for example, based on t consecutive days of blood glucose data collected before breakfast. This data is converted to Y ₁ . . . Let's call it Y _t . This data can then be classified into “bins” representing different levels of blood glucose. There can be k defined bins, each of which contains an equal range of possible blood glucose values. These bins can represent the blood glucose status of the patient. The estimated probability of transition from state j to state j + 1, j to j + 2, etc. can be calculated, and a transition matrix can be constructed from the observation data. This transition matrix can then be used in computer simulations to calculate relevant results. One possible outcome is the probability that the patient will reach the blood glucose threshold after time t within h days. This probability is a defined cutoff (for example,

) Result in

It is considered that the reliability level is expected.

Further (or alternatively), the prediction can be performed using the random walk model described in Section 3E above. This prediction can be performed, for example, based on t consecutive days of blood glucose data collected before breakfast. This data is converted to Y ₁ . . . Let's call it Y _t . This data is input to the random walk model. If a certain trend term is included, the investigator can use an estimate of the term □ from the OLS model described in Section 1A above. The output from the algorithm is a series of data representing the expected value of the first h days of the patient's blood glucose reading before breakfast

It is.

  In addition (or alternatively), the three-layer network described in Section 3F above (for example) is used (for example) to predict hyperglycemia or hypoglycemia events. The input to the network may be any combination of measurements taken (eg, the previous day), including but not limited to blood glucose, blood pressure, physical activity, and weight measurements. By adding a label to each day reading, the above-described managed learning procedure can be implemented. The measurement that precedes the day when a blood glucose reduction event occurs can be labeled (for example) a-1, and the measurement that precedes the day when neither a hypoglycemia event nor an excessive blood glucose event occurs For example, a label of 0 can be given, and a label of (for example) a + 1 can be added to the measurement one day before the hyperglycemia event occurs. The number of concealment units used in this system may be in the range 1 to n / 10, where n represents the number of training points available. Use of a greater number of concealment units can lead to unacceptably large test errors, and training errors generally decrease as the number of concealment units increases. The activation function used may fit many of the positive features that have value in such a function. The function must be centered on zero and have antisymmetry, so in this example we choose:

f (net) = atanh (b * net)

  The overall range and slope are not so important, but for convenience we choose a = 1.716 and b = 2/3. this is,

And guarantees that the linear range is -1 <net <+1. In the case of the above parameter set, the learning rate uses η = 0.1 as a starting point. The learning rate needs to be reduced if the reference function diverges during learning, and increased if learning seems to be too slow. As a result, the output of this network is a single number similar to the label used in the learning process. Then, by generating general rules, the possible directions of action can be outlined. For example (this example is intended to be illustrative and not limiting), it is considered that an hyperglycemic or hypoglycemic event can occur when the network output is | z | ≧ 0.5. , You can choose intervention for the patient.

  These processes include blood glucose measured at bedtime, before or after lunch, before or after dinner, or according to daily average; or blood pressure (cardiac contraction and / or diastole) (however, (But not limited to) can be used for any target input. The expected value can be compared against a predefined threshold in the system described below or described above, where the patient can anticipate a health problem. The final result could be encoded as (1, 0, -1), meaning an expected value above the upper limit, an expected value within the limit, and an expected value below the lower limit, respectively. The number of expected periods, or “h”, can be any number greater than or equal to 1, but the uncertainty of the prediction increases with distance from the current time. Also, the time “window” used in training the model. . . T can include any length of time (eg, greater than 30 data points).

  Any or all of the enumerated values in the form of a report or other communication to one or more of the parties (patient physician, patient caregiver, and / or patient: but not limited to) It can be used. Such reports can include (but are not limited to) the following:

(A. Blood pressure)
Physician report: Blood pressure prediction results can be used in charts, graphs, and / or text reports generated by the system. Based on whether the expected blood pressure (diastole and / or systole) is expected to be above or below the limit defined by the physician, the expected value may be able to drive content and alerts. Such analysis can also contribute to a range of knowledge that drives changes in patient medication and / or other treatment by the physician.

  Caregiver report: This content is similar to a physician report and includes charts, graphs, and / or text reports, but is specifically adapted to caregiver preferences. For example, a caregiver may desire to have a different threshold that indicates a dangerous expectation. Based on whether the expected blood pressure (diastolic and / or systolic) is expected to be above or below the limit defined by the caregiver, the expected value may be able to drive content and alerts.

  Patient report: The patient can receive an alert if he is in an imminent danger state with a dangerous level of systolic and / or diastolic blood pressure. In order to avoid critical situations, these warnings can be accompanied by content describing how the patient can reduce the risk of having dangerous high or low blood pressure.

(B. Glucose)
Physician Report: Blood glucose prediction results can be used in charts, graphs, and / or text reports generated by the system. Expected blood glucose to be above or below the limits defined by the physician before and after breakfast, before and after lunch, before and after dinner, at bedtime, and daily average measurements Depending on whether or not, it may be possible to drive content and alerts according to expected values. Such an analysis can also contribute to a range of knowledge that drives changes in patient medication and / or other therapies by the physician.

  Caregiver report: This content is similar to a physician report and includes charts, graphs, and / or text reports, but is specifically adapted to caregiver preferences. For example, a caregiver may wish to have a different threshold for notifying about dangerous expectations. Expected blood glucose above and below the limits defined by the caregiver before and after breakfast, before and after lunch, before and after dinner, at bedtime, and daily average measurements Content and alerts may be driven by the expected value based on whether or not

  Patient Report: Patients are in imminent danger before and after breakfast, before and after lunch, before and after dinner, at bedtime, and at dangerous levels of average blood glucose daily. A warning can be received. In order to avoid dangerous situations such as hypoglycemia or hyperglycemia, respectively, these warnings contain content describing how patients can reduce the risk of having dangerous high blood glucose or low blood glucose. It can be attached.

  Reference is now made to another example of data and data processing according to one embodiment of the present invention (which is, of course, intended to be illustrative and not limiting). In this example, a type 2 diabetic patient (in this example, this is referred to as a “subject”) has generated data for six months. The data is provided as follows (this is the subject's data sent to the PNC in one day).

(1) Blood glucose reading value Reading time Reading value (mg / dcl)
Average (0600-1000) 250
Average (1000-1400) 200
Average (1400-1800) 300
Average (1800-2200) 220
Average (2200-0600) 180
(2) Blood pressure reading value Reading time Reading value (mmHg)
0900 160/82
1230 170/84
1800 180/90
(3) Physical activity Reading time Reading (cumulative combustion calories)
(Cumulative) 2200 1800
(4) Weight Reading time Reading value (kg)
0700 330

  An example of subject data analysis performed according to one embodiment of the present invention using predictive modeling is as follows.

(1) Generated warning signal A. The dinner reading has exceeded 280 for more than three times in the last 90 days, so a warning or “flag” regarding hyperglycemia at dinner is raised. Physical activity is below the minimum value prescribed for the patient (1900 for the object).

(2) Prediction (using the GARCH model described later)
A. In this model, it is as follows.
T is today.
・ 1. . . t-1 is all days prior to today.
T + 1 is tomorrow.

B. All data are averages for the day. For example, in the case of data given with respect to the target person, the data of “today” is as follows.
○ Blood sugar = 230mg / dcl
○ Blood pressure = 170 / 85mmHg
○ Physical activity = 1800 cals
○ Weight = 330kg

C. The inputs to the GARCH model are as follows:
· X ₁ is a full-time series of the systolic blood pressure of the subject (X ₁₁ ... X _1t).
· X ₂ is a full-time series of diastolic blood pressure of the subject (X ₂₁ ... X _2t).
X ₃ is the full time sequence (X ₃₁ ... X _3t ) of the subject's physical activities.
X ₄ is a full-time series of subject weights (X ₄₁ ... X _4t ).
Y is the subject's blood glucose level (y ₁ ... Y _t ).

D. Predict y _{t + 1} according to the GARCH model algorithm just described in the equation section. All other variables and coefficients appearing in the model are calculated using a maximum likelihood algorithm.
1. Tomorrow's blood glucose prediction value by predictive modeling is 215. This is within the “normal” range of subjects who are chronically ill patients (within one standard deviation of the daily average), but not within the normal range of healthy people.
2. The blood pressure in the last 30 days has increased when measured by OLS.
3. Physical activity in the last 60 days has decreased as measured by OLS.
4). The distribution of blood glucose over the past 30 days has increased by using the D-statistic method.
5). Body weight in the past 60 days has been reduced by using OLS.

  (3) The results from the mathematical analysis of the present invention show that this chronically ill target is poorly managed in diabetes, high in continuous blood sugar rise and extreme dispersion. Prove that you are at risk. A decrease in physical activity contributes to this high risk situation with increased blood pressure. Data analysis therefore reveals an extremely high medical risk profile. Without active intervention, this patient develops persistently high average glucose-induced diabetes microvascular complications (kidney disorders, blindness, peripheral nerve dysfunction, and infections that lead to amputation) Very high risk. Patients are also at a very high risk state of developing diabetic macrovascular complications (heart disease and paralysis) due to elevated glucose dispersion. These complications can lead to serious and chronic medical events that require hospitalization and other expensive interventions. Intervention detects and characterizes these risk factors as well as helps set the appropriate interventions up to the point in time when glucose and blood pressure are controlled, evaluate their effects, and reduce cycle time in improving the performance of the intervention. Being useful can prevent the aforementioned complications, thereby avoiding the need for expensive hospitalization, amputation, and dialysis.

  (4) The above example provides an early warning that occurs prior to the next medical appointment scheduled for the subject. This example also illustrates a predictive modeling tool for smoothly performing various interventions designed to reduce blood glucose in a patient subject in addition to reducing blood pressure and increasing physical activity. Is also provided. The subject will likely accept non-emergency intervention and receive treatment in advance, without any intervention in the emergency room (this saves emergency-related spending on the health care system) . Plans can be introduced to motivate patients who are focused on avoiding kidney damage. Such plans can include (but are not limited to) counseling, dietary changes, nutritional input, and suggestions for changes in pharmacological routines and physical activities. In addition (or alternatively), the tool of the present invention can also smoothly perform interventions designed to solve problems that hinder the level of physical activity that a subject may have. is there.

  In yet another embodiment, by generating time series using appropriate variables, the present invention provides the basis for integrated analysis of appropriate factors and (for example) in the case of diabetic patients. In this way, the development of care for patients with individualized chronic diseases has been realized by promoting the separation of causes and effects.

  Furthermore, the use of such care that can be implemented by the present invention can improve the care of patients with chronic illness and provide enhancement to the patient in the form of regular output. In this regard, for example, it is possible to provide patients with chronic diseases such as diabetic patients the ability to participate daily in their therapy. The improved care of chronically ill patients through essentially continuous care that is possible through the use of the present invention has oversimplified users of the present invention, such as chronically ill patients and / or health care providers. It will be useful to free you from the negative effects of statistics and large group averaging.

  In one particular example (this example is intended to be illustrative and not limiting), the output information can range from 2 to 24 hours for patients, caregivers, and / or health care providers. Can be transmitted within. Such output information can be transmitted (for example, to a patient) via, for example, a “virtual companion” (a computer image representing a nurse). The virtual companion recommends (a) paying attention to what to eat for dinner, (b) increasing exercise to control blood pressure (referring to lack of exercise yesterday), and (c) medication Feedback can be provided to the patient, such as but not limited to notification of the need for immediate medical appointments for adjustment.

  In another specific example (this example is intended to be illustrative and not limiting), a Diabetes Care System (DCS) can be established, which is described in real time as described above. Can be used to modify and improve the care of patients with chronic diseases, as in the example of care for diabetic patients. For example, instead of monthly updates, caregivers can receive regular updates (eg, hourly, daily, weekly, etc.) regarding testing, food intake, exercise, weight, and / or blood oxygen saturation It is. Based on the algorithm and predictive modeling of the present invention, patients can receive regular feedback regarding actions to be taken to maintain and / or improve their health.

  Furthermore, this can reduce the possibility of disastrous medical events such as the discovery of hyperglycemia or hypoglycemia symptoms in diabetic patients through the use of advanced analysis of patient data and predictive models, each of which, for example, May rely solely on the patient's own data (reducing the possibility of a catastrophe by implementing an accurate process that predicts when such an event can occur, and (Note that alerts can be provided to the patient and their caregiver / medical service provider with sufficient time to smoothly perform the intervention).

  Another embodiment of the present invention is characterized by (a) confirmation of the disease type of a patient with chronic disease and (b) subsequent appropriate identification of the patient with chronic disease, and predictive modeling (which is a tragedy (C) providing information that can be used to structure treatments by time series evaluation; (d) A method that works for the patient with that particular chronic disease through the use of subsequent mathematical techniques (which can directly improve the condition of the patient with chronic disease while avoiding the use of ineffective resources Assessment of the effectiveness of those methods focused on interventions).

  In another embodiment of the invention, the PNC can record, store, analyze, and / or predictively model data and recognize changes in patient data. Changes can include, for example, variance in disease or health transitions and conditions. In addition, the progress of chronically ill patients can be achieved by integrating measured data with additional data (eg, lifestyle data and / or physician-generated information) and using time-series nonlinear analysis combined with system identification mathematics. It can provide an essentially continuous assessment of the situation.

  It should be noted that the present invention can be implemented using any suitable computer hardware and / or computer software. In this regard, those skilled in the art will be aware of the types of computer hardware that can be used (eg, mainframe, minicomputer, personal computer (PC), network (eg, intranet and / or Internet), types of computer programming methods that can be used. Familiarity with (eg, object-oriented programming) and the types of computer programming languages / structures available (eg, C ++, Basic HTML, ASP).

  While a number of embodiments of the present invention have been described above, these embodiments are merely intended to be illustrative and not limiting and many variations will be apparent to those skilled in the art. Please understand that. For example, this specification describes a particular method as being “computer-implemented” or “computer-implementable”. In this regard, it should be noted that although the methods can be implemented using a computer, the methods do not necessarily have to be implemented using a computer. Also, not all steps need necessarily be implemented using a computer in terms of being implemented using a computer. Furthermore, while the foregoing has described the present invention primarily in the context of diabetes, it will be appreciated that the present invention applies to one or more of any other desired medical condition. It will be possible. For example, diabetes related data can be replaced (or expanded) by other input from the patient / user, such as an electrocardiogram signal for health notification parameters related to heart and / or tissue status. Further, the various time periods (eg, regular time periods) related to the present invention are (a) seconds, (b) minutes, (c) days, (d) weeks, (e) months. It is also possible to select from a group including (but not limited to). Also, the various steps can be performed in any desired order.

3 is a flowchart illustrating a method according to an embodiment of the present invention. A graphical representation of a statistical summary of complications or related disease occurrences in diabetic patients. A graphical representation of a statistical summary of costs associated with common complications of diabetes. It is a graphical overview showing the proportion of costs incurred due to diabetic patients depending on the type of care required. 1 is a graphical overview of the results of numerous clinical experiments that have demonstrated the potential to reduce diabetic complications. FIG. 6B shows the relationship between FIG. 6A and FIG. 6B. A graphical overview summarizing the impact of various interventions on health events, including avoided costs (first half). A graphical overview summarizing the impact of various interventions on health events, including avoided costs (second half). A graphical summary set of studies demonstrating the relationship between improved control of diabetes, improved care management, and cost savings. The relationship between FIG. 8A and FIG. 8B is shown. Characteristics and timing of clinical improvements and cost savings that can be attributed to the application of one embodiment of the present invention based on the clinical experiments shown in FIGS. 3, 4, 5, 6, and 7. (First half). Characteristics and timing of clinical improvements and cost savings that can be attributed to the application of one embodiment of the present invention based on the clinical experiments shown in FIGS. 3, 4, 5, 6, and 7. It is a set of tables showing (second half). FIG. 6 is a flow chart illustrating a system and method according to an embodiment of the present invention (depicting the transfer of information between at least one measurement device and a patient “data management device”; Or from a plurality of measuring devices to a data management device and / or from a data management device to one or more measuring devices). FIG. 4 is a flow chart illustrating a system and method according to one embodiment of the present invention (depicting the transfer of information between at least one “virtual private network” device and an ISP server, where the transfer of information is one; Or from a plurality of virtual private network devices to an ISP server and / or from an ISP server to one or more virtual private network devices).

Claims (36)

In a computer-implemented method for predictive modeling of patient status,
Performing a plurality of measurements of physical parameters presented by the patient;
Organizing the plurality of measurements into time-series data;
Applying a distributed detection algorithm to the time series data to predict the future state of the patient;
Having a method.   The method of claim 1, wherein the variance detection algorithm comprises an algorithm selected from the group comprising (a) an ARIMA model, (b) a D statistical method, and (c) a non-parametric method. 3. The method of claim 2, wherein the ARIMA model has an expression specified at least in part as
Y _t = α ₁ Y _t-1 +. . . _{+ Α p Y tp + e t} + β 1 e t-1 +. . . + Β _q e _tq 3. The method of claim 2, wherein the D statistic has an expression specified at least in part as
D _k = (C _k / C _t ) − (k / T) The method of claim 2, wherein the non-parametric method has an expression at least partially designated as:
U _{t, T} = 2W _t −t (T + 1)   The method of claim 1, wherein the prediction of the future state of the patient is performed in real time as further measurements are performed and organized into time series data.   The prediction of the future state of the patient is provided to at least one of (a) the patient, (b) a caregiver associated with the patient, and (c) a health care provider associated with the patient. The method of claim 6.   8. The method of claim 7, wherein the medical service provider is selected from the group comprising (a) a doctor and (b) a nurse.   The physical parameters include (a) blood glucose level, (b) blood pressure, (c) blood oxygen saturation, (d) electrical activity, (e) body weight, and (f) physical activity. The method of claim 1 selected from.   The method of claim 1, wherein a plurality of measurements consists of each of a plurality of physical parameters presented by the patient.   The method of claim 10, wherein the patient is a chronically ill patient. In a computer-implemented method for predictive modeling of patient status,
Performing a plurality of measurements of physical parameters presented by the patient;
Organizing the plurality of measurements into time-series data;
Applying a prediction algorithm to the time series data to predict the future state of the patient;
Have
The prediction algorithm includes: (a) a general ARMAX model; (b) a GARCH model; (c) a Kalman filtering; (d) a Markov model; (e) a random walk model; (f) a multi-layer neural network; ) A method having an algorithm selected from the group including extreme value analysis. The method of claim 12, wherein the general ARMAX model has an expression specified at least in part as:

13. The method of claim 12, wherein the GARCH model has an expression specified at least in part as

The method of claim 12, wherein the Kalman filtering comprises an expression specified at least in part as
X _t = AX _t−1 + Bμ _t−1 + ω _t−1 13. The method of claim 12, wherein the Markov model has an expression specified at least in part as
P (X _t = j | X ₀ = i ₀ , X ₁ = i ₁ ,... X _t-1 = i _t-1 ) = P (X _t = j | X _t-1 = i _t-1 ) The method of claim 12, wherein the random walk model has an expression at least partially specified as:
Y _t = Y _t-1 + α The method of claim 12, wherein the multi-layer neural network has an expression specified at least in part as

13. The method of claim 12, wherein the extreme value analysis includes parameter fitting and has an expression specified at least in part by

The method of claim 12, wherein the extreme value analysis includes a suitability test and has an expression at least partially designated as:

  The method of claim 12, wherein the prediction of the future state of the patient is performed in real time as further measurements are performed and organized into time series data.   The prediction of the future state of the patient is provided to at least one of (a) the patient, (b) a caregiver associated with the patient, and (c) a health care provider associated with the patient. The method of claim 21.   23. The method of claim 22, wherein the medical service provider is selected from the group comprising (a) a doctor and (b) a nurse.   The physical parameters include (a) blood glucose level, (b) blood pressure, (c) blood oxygen saturation, (d) electrical activity, (e) body weight, and (f) physical activity. The method of claim 12 selected from.   The method of claim 12, wherein a plurality of measurements is comprised of each of a plurality of physical parameters presented by the patient.   26. The method of claim 25, wherein the patient is a chronically ill patient. In a computer-implemented method for predictive modeling of patient status,
Performing a plurality of measurements of physical parameters presented by the patient;
Organizing the plurality of measurements into time-series data;
Applying a trend detection algorithm to the time series data to predict the future state of the patient;
Have
The trend analysis algorithm comprises an algorithm selected from the group comprising (a) a piecewise linear model and (b) pattern recognition.   28. The method of claim 27, wherein the prediction of the future state of the patient is performed in real time as further measurements are performed and organized into time series data.   The prediction of the future state of the patient is provided to at least one of (a) the patient, (b) a caregiver associated with the patient, and (c) a health care provider associated with the patient. 30. The method of claim 28, wherein:   30. The method of claim 29, wherein the medical service provider is selected from the group comprising (a) a doctor and (b) a nurse.   The physical parameters include (a) blood glucose level, (b) blood pressure, (c) blood oxygen saturation, (d) electrical activity, (e) body weight, and (f) physical activity. 28. The method of claim 27, selected from:   28. The method of claim 27, wherein a plurality of measurements is comprised of each of a plurality of physical parameters presented by the patient.   33. The method of claim 32, wherein the patient is a chronically ill patient.   The method of claim 1, wherein the steps are performed in the order cited.   The method of claim 12, wherein the steps are performed in the order cited.   28. The method of claim 27, wherein the steps are performed in the order cited.

Images