TWI716083B - System and method for predicting types of pathogens in patients with septicemia - Google Patents

Abstract

A system for predicting types of pathogens in patients with septicemia is provided. The system includes at least one sensor and a processor. The sensor is used to sense current physiological data including at least one of body temperature, blood pressure, and pulse. The processor is configured to calculate at least one feature value according to the current physiological data, and input the feature value into a machine learning model to determine one of categories including at least two of uninfected, fungal infection, contaminated bacteria infection, Gram-negative infection, and Gram-positive infection.

Description

Prediction system and method of bacterial species for sepsis

The invention relates to a strain prediction system and method for sepsis, which can predict the type of pathogen before the result of pathogen culture is released.

Sepsis is the main cause of death in hospitalized patients. Immediate administration of effective antibiotics can reduce the mortality of patients with sepsis. However, there is currently no test method to correctly predict the pathogen infection before the results of the pathogen culture are released, so clinicians usually base it on the basis of no basis. The patient is given antibiotics based on personal judgment, so how to determine whether the patient is infected and what kind of pathogen is infected before the pathogen culture result is released is a topic of concern to those skilled in the art.

The embodiment of the present invention provides a strain prediction system for sepsis, including a sensor and a processor. The sensor is used to sense current physiological data, and the current physiological data includes at least one of body temperature, blood pressure, and pulse. The processor is used to calculate the characteristic value according to the current physiological data and input the characteristic value To the machine learning model to determine one of multiple categories, these categories include at least two of uninfected, fungal infection, contaminated bacteria infection, Gram-negative bacteria infection and Gram-positive bacteria infection.

In some embodiments, for each current physiological data, the processor is used to perform multiple steps: obtain a healthy physiological data that changes with time; calculate the average value of the healthy physiological data as a healthy average; The variance is used as a health variance; the variance of the current physiological data is calculated as a current variance; the variance of the current physiological data relative to the health average is calculated as a reference variance; the reference variance is divided by the health variance As the first characteristic value; and dividing the current variance by the health variance as the second characteristic value.

In some embodiments, the reference variance is calculated according to the following equation (1), where X _current is the current physiological data, μ _health is the average health, and # current is the sampling number of the current physiological data.

In some embodiments, the aforementioned sensor includes a gravity sensor, and the processor is used to determine whether the user is stationary according to the signal sensed by the gravity sensor, and obtain at least current physiological data when the user is stationary.

In some embodiments, the aforementioned machine learning model is a random forest algorithm.

From another perspective, the embodiment of the present invention provides a method for predicting sepsis bacteria, including: sensing current physiological data through a sensor, and the current physiological data includes at least one of body temperature, blood pressure, and pulse; Calculate the feature value from the current physiological data and input the feature value into the machine Use the model to determine one of multiple categories, including at least two of non-infected, fungal infection, contaminated bacterial infection, Gram-negative bacterial infection, and Gram-positive bacterial infection.

In order to make the above-mentioned features and advantages of the present invention more comprehensible, the following specific embodiments are described in detail in conjunction with the accompanying drawings.

100‧‧‧Bacteria Prediction System

110‧‧‧Sensor

120‧‧‧Processor

130‧‧‧Communication Module

140‧‧‧Display

201~205、301~303‧‧‧Step

[Fig. 1] is a schematic diagram illustrating a strain prediction system for sepsis according to an embodiment.

[Fig. 2] is a classification flowchart according to an embodiment.

[Fig. 3] is a flowchart of a method for predicting bacterial species of sepsis according to an embodiment.

Regarding the "first", "second", etc. used in this text, it does not specifically refer to the order or sequence, but only to distinguish elements or operations described in the same technical terms.

Fig. 1 is a schematic diagram illustrating a strain prediction system for sepsis according to an embodiment. Please refer to FIG. 1, the strain prediction system 100 includes a plurality of sensors 110, a processor 120, a communication module 130 and a display 140. The sensor 110 can be used to sense physiological data such as body temperature, blood pressure (including diastolic blood pressure and systolic blood pressure), pulse, heart rhythm, etc. Those with ordinary knowledge in the art should use a suitable sensor, such as an infrared thermometer. Body temperature etc. The processor 120 may be Central processing units, microprocessors, microcontrollers, signal processors, integrated circuits for special applications, etc. The communication module 130 may be a wired or wireless communication module for communicating with other devices. For example, the communication module 130 may be equipped with a universal serial bus (USB), the Internet, a local area network, a wide area network, Circuits with communication functions such as cellular telephone network, near field communication, infrared communication, Bluetooth, and WiFi. The display 140 may be a liquid crystal display, an organic light emitting diode display, or other suitable displays. In this embodiment, the sensor 110 is used to sense at least one current physiological data, and the processor 120 is used to calculate feature values based on the current physiological data, and input the feature values into a machine learning model to determine multiple One of the categories, these categories may include viral infections, uninfected, fungal infections, contaminated bacteria infections, Gram-negative bacteria infections, and Gram-positive bacteria infections. In some embodiments, the strain prediction system 100 can be implemented as a wristband to be worn on the patient's hand, but in other embodiments, the strain prediction system 100 can also be implemented as any form of computer or mobile device. The present invention Not limited to this. In other embodiments, the strain prediction system 100 may also be provided with other suitable devices, or the communication module 130 and the display 140 may also be omitted.

Here will explain in detail how to determine the type of infection. First of all, the aforementioned physiological data such as body temperature, blood pressure, pulse, heart rate, etc. are signals that change over time. The processor 120 can obtain the physiological data for a period of time (for example, several seconds, the present invention does not limit the length of time) through the sensor 110. . For example, if the sampling frequency is 60 Hz, the physiological data for 5 seconds will include 60×5=300 values, but the present invention does not limit the sampling frequency. For the sake of clarity, the physiological data obtained through the sensor 110 is referred to as Current physiological data.

In addition, the processor 120 can also obtain physiological data (also called healthy physiological data) corresponding to the health state such as body temperature, blood pressure, pulse, and heart rhythm from the database (not shown). These health physiological data are when a person is in a healthy state. (Such as not infected) measured through a sensor. These healthy physiological data are also signals that change over time, but the present invention does not limit the length of these healthy physiological data, that is, it does not limit each piece of healthy physiological data to contain several values. In other words, the length of healthy physiological data can be different from the length of current physiological data.

For each type of physiological data (ie body temperature, blood pressure, pulse or heart rhythm), the processor 120 will calculate two characteristic values. Here, the numerical value in the healthy physiological data represents X _health , and #health represents the length of the healthy physiological data (ie, the number of values X _health ). The value in the current physiological data is expressed as X _current , and #current represents the length of the current physiological data (that is, the number of values X _current ), also known as the number of samples. The processor 120 calculates the average value of the healthy physiological data as a healthy average value, which is denoted as μ _{health below} , and the average value of the current physiological data is denoted as μ _current . In addition, according to the following equation (1), the variance of health physiological data can be calculated as the health variance σ _health ; according to the following equation (2), the variance of the current physiological data can be calculated as the current variance σ _sick-sick ; According to the following equation (3), the variance of the current physiological data relative to the average health can be calculated as a reference variance σ _{current-health} .

The first feature value f1 can be obtained by dividing the reference variance by the health variance, as shown in the following equation (4). In addition, dividing the current variance by the health variance can obtain the second characteristic value f2, as shown in the following equation (5).

In this embodiment, there are four physiological data such as body temperature, blood pressure, pulse, and heart rhythm. Therefore, there are at least four of the above-mentioned first feature values f1 and four second feature values f2, a total of eight feature values (or diastolic blood pressure corresponding The two characteristic values of systolic blood pressure also have two corresponding characteristic values, so there are 10 characteristic values in total). In other embodiments, all of the above-mentioned first feature value f1 and second feature value f2 will form a feature vector, and this feature vector will be input to a machine learning model. The machine learning model can be a random forest algorithm, a support vector machine (support vector machine), a neural network, etc., and the present invention is not limited thereto. This machine learning model is trained to determine whether a patient is infected and the type of pathogenic bacteria infected. In some embodiments, the categories output by the machine learning model include at least two of virus infection, non-infection, fungal infection, contaminating bacteria infection, Gram-negative bacteria infection, and Gram-positive bacteria infection. Contaminating bacteria infection means that the pathogen in the patient's body is caused by some pollution source, not sepsis.

Please refer to FIG. 2. In some embodiments, the sequence of judgment is to perform step 201 first to judge whether it is infected. If the result of step 201 is No, it means there is no infection. If the result of step 201 is yes, then step 202 is performed again to determine the bacterial species and determine whether it is a bacterial infection, a fungal infection, or a virus infection. If judged If it is judged to be a bacterial infection, it is judged in step 203 whether it is a Gram-positive bacteria. According to the result of step 203, it can be determined as Gram-negative bacteria infection (step 204) or Gram-positive bacteria infection (step 205). In some embodiments, a total of 3 classifiers can be trained according to the process of FIG. 2, corresponding to steps 201 to 203 respectively. In other embodiments, only one classifier needs to be trained. The output results of this classifier include uninfected, fungal infection, virus infection, Gram-negative bacteria infection and Gram-positive bacteria infection, and the present invention is not limited to this. It is worth noting that the process in FIG. 2 is only an example, and one or more determination steps may be added or deleted in other embodiments. For example, in step 202, it can also be determined whether it is a contaminating bacteria infection.

In the above physiological data, body temperature is important information used to determine whether the patient is infected. However, since the patient may get up and walk, which will affect the value of body temperature, the sensor 110 of FIG. 1 may include gravity sensing in some embodiments. The gravity sensor is, for example, an acceleration sensor. According to the signal of the gravity sensor, it can be determined whether the user is in a static state, for example, when the acceleration in each direction is less than a critical value, it is determined to be static. In addition, the current physiological data can be obtained only when the user is stationary, that is, the processor 120 will ignore the physiological data sensed by the sensor 110 when the user is not stationary. In this way, it can be avoided that when the user gets up to move or perform other actions, getting an inappropriate body temperature, which will affect the judgment result.

It should be noted that the aforementioned eigenvalues f1 and f2 may only be part of the eigenvector, and the eigenvector may also include other information. For example, the feature vector can also include information such as the user's age, gender, and medical history, which will be digitized as part of the feature vector. Or it can be based on The signal detected by the sensor 110 calculates other characteristic values to form a characteristic vector, and the invention is not limited thereto.

In some embodiments, the strain prediction system 100 is implemented as a wearable device, which is carried by the patient, so the patient can be in any position. The strain prediction system 100 can determine whether the patient is infected from time to time or regularly. The strain prediction system 100 can also send the collected physiological data or the determined classification result to a server or doctor through the communication module 130 On your mobile phone, the hospital or doctor can notify the patient to seek effective medical treatment immediately.

In some embodiments, the above-mentioned physiological data can be converted into an image, which is input to a convolutional neural network for classification. For example, for each type of physiological data, an image can be generated based on the calculation of the covariance between the current physiological data and healthy physiology. The pixel p _i in the i-th row (column) and j-th row (row) of this image _{,j is} expressed as the following equation (6), where X _current,i represents the i-th value in the current physiological data, X _health,j is the j-th value in the healthy physiological data, and i and j are positive integers.

p _i,j =(X _current,i -μ _current )×(X _health,j -μ _health ) (6)

Since each type of physiological data can be used to generate an image, a total of 4 images will be generated. These 4 images will be merged into a 2D image with 4 channels. This 2D image will be input to the convolutional nerve Classification in the network. From another perspective, the above-mentioned pixels p _i,j can also be called feature values.

In some embodiments, the image can also be generated according to the following equation (7).

p _i,j = ( x _i - x _j ) ² (7)

x _i is the i-th value in the current physiological data or healthy physiological data. It is worth noting that both physiological data and healthy physiological data can be applied to equation (7). Therefore, two images can be generated for each type of physiological data. In the above example, a total of 8 images will be generated, and these 8 images will be Merge together into a two-dimensional image with 8 channels, and this two-dimensional image will be input into a convolutional neural network for classification.

Fig. 3 is a flowchart illustrating a method for predicting bacterial species of sepsis according to an embodiment. 3, in step 301, the current physiological data is sensed, and the current physiological data includes at least one of body temperature, blood pressure, and pulse. In step 302, the characteristic value is calculated based on the current physiological data. In step 303, the feature value is input to the machine learning model to determine one of multiple categories, including non-infected, fungal infection, contaminated bacteria infection, Gram-negative bacteria infection and Gram-positive bacteria At least two of them are infected. However, each step in FIG. 3 has been described in detail as above, and will not be repeated here. It is worth noting that each step in FIG. 3 can be implemented as multiple program codes or circuits, and the present invention is not limited thereto. In addition, the method in FIG. 3 can be used in conjunction with the above embodiments, or can be used alone. In other words, other steps can also be added between the steps in FIG. 3.

In the above-mentioned system and method, since the infection and the type of pathogen can be predicted, there is no need to wait until the blood culture result is released. In addition, clinicians can refer to the predicted results to prescribe appropriate antibiotics to treat patients with sepsis, thereby improving the survival rate of patients with sepsis. In addition, the above prediction method is a non-invasive examination and does not require additional blood tests.

Although the present invention has been disclosed in the above embodiments, it is not intended to limit the present invention. Anyone with ordinary knowledge in the technical field can make some changes and modifications without departing from the spirit and scope of the present invention. The scope of protection of the present invention shall be determined by the scope of the attached patent application.

301~303‧‧‧Step

Claims (8)

A strain prediction system for sepsis, comprising: at least one sensor for sensing at least one current physiological data, wherein the at least one current physiological data includes at least one of body temperature, blood pressure, and pulse: a processor, At least one feature value is calculated based on each of the at least one current physiological data, and the at least one feature value is input to a machine learning model to determine one of a plurality of categories, where the categories include uninfected, At least two of fungal infections, contaminating bacteria infections, Gram-negative bacteria infections, and Gram-positive bacteria infections. For each of the current physiological data, the processor is used to perform multiple steps: obtaining time-varying A healthy physiological data; calculate the average value of the healthy physiological data as a healthy average; calculate the variance of the healthy physiological data as a health variance; calculate the variance of the current physiological data as a current variance; Calculate a variance of the current physiological data relative to the average health value as a reference variance; divide the reference variance by the health variance as a first characteristic value; and Divide the current variance by the health variance as a second characteristic value. For the strain prediction system described in item 1 of the scope of patent application, the reference variance is calculated according to the following equation (1):

Where X _current is the value in the current physiological data, μ _{health is} the average health value, and # current is the sampling number of the current physiological data. The strain prediction system described in claim 1, wherein the at least one sensor includes a gravity sensor, and the processor is used to determine whether a user is stationary according to the signal sensed by the gravity sensor , And obtain the at least one current physiological data when the user is stationary. Such as the strain prediction system described in item 1 of the scope of patent application, wherein the machine learning model is a random forest algorithm. A method for predicting bacterial strains of sepsis is applicable to a processor, wherein the predicting method of bacterial strains includes: sensing at least one current physiological data through at least one sensor, wherein the at least one current physiological data includes temperature, blood pressure, and pulse At least one of them: calculating at least one characteristic value based on each of the at least one current physiological data; and The at least one feature value is input into a machine learning model to determine one of multiple categories, where the categories include uninfected, fungal infection, contaminated bacteria infection, Gram-negative bacteria infection, and Gram-positive At least two of the bacterial infections, wherein the step of calculating at least one characteristic value based on each of the at least one current physiological data includes: obtaining a healthy physiological data that changes over time; calculating the average value of the healthy physiological data as a healthy Average; calculate the variance of the health physiological data as a health variance; calculate the variance of the current physiological data as a current variance; calculate the variance of the current physiological data relative to the health average as a Reference variance; dividing the reference variance by the health variance as a first characteristic value; and dividing the current variance by the health variance as a second characteristic value. For the bacterial species prediction method described in item 5 of the scope of patent application, the reference variance is calculated according to the following equation (1):

Where X _{current is} the current physiological data, μ _{health is} the average health value, and # current is the sampling number of the current physiological data. For example, the bacterial species prediction method described in item 5 of the scope of patent application further includes: determining whether a user is stationary according to a signal sensed by a gravity sensor, and obtaining the at least one current when the user is stationary Physiological information. In the strain prediction method described in item 5 of the scope of patent application, the machine learning model is a random forest algorithm.

Images