Evaluation system for heart condition

Abstract

The invention provides a heart condition evaluation system which comprises an electrocardiogram sensor, a light volume change sensor and a processing circuit. The electrocardiogram sensor is used for acquiring an electrocardiogram signal about a user. The light volume change sensor is used for obtaining a light volume change signal about a user. The processing circuit generates a cardiac evaluation result based on the light volume change signal sensed for a first time and the electrocardiogram signal sensed for a second time, wherein the first time is longer than the second time. The system can assess the heart condition for a longer time.

Description

Technical Field

[0001] The present invention relates to a system for evaluating cardiac status using a machine learning model. Background Art

[0002] Conventional technology must combine the patient's medical history, medical interviews, and related examinations to evaluate heart failure, but these practices are very costly. In order to solve this problem, there are already many devices on the market that can be worn on patients to obtain electrocardiogram (electrocardiogram) signals. In addition, machine learning technology has gradually become popular, and machine learning technology based on electrocardiograms can already provide good evaluation results. However, this approach usually obtains relatively short (e.g., tens of seconds) electrocardiogram signals for identification, which lacks a longer comprehensive assessment of the heart's condition. If you want to obtain a long-term electrocardiogram signal, you will face many problems, such as the electrodes used to sense the electrocardiogram cannot contact the patient's skin for a long time, and there are battery problems, etc. Summary of the invention

[0003] The present disclosure provides a cardiac condition assessment system, comprising an electrocardiogram sensor, a light volume variation sensor, and a processing circuit. The electrocardiogram sensor is used to obtain an electrocardiogram signal related to a user. The light volume variation sensor is used to obtain a light volume variation signal related to a user. The processing circuit is electrically connected to the electrocardiogram sensor and the light volume variation sensor, and is used to generate a cardiac assessment result according to the light volume variation signal sensed within a first time and the electrocardiogram signal sensed within a second time, wherein the first time is longer than the second time.

[0004] In some embodiments, the processing circuit is used to input the electrocardiogram signal into the first machine learning model to generate a first feature vector. The processing circuit identifies multiple heartbeats from the photoplethysmographic signal, calculates multiple heartbeat intervals based on the heartbeats, and inputs the heartbeat intervals into the second machine learning model to generate a second feature vector. The processing circuit is used to input the first feature vector and the second feature vector into a third machine learning model to generate a cardiac assessment result.

[0005] In some embodiments, the first time includes a plurality of sub-time ranges, and the processing circuit is used to select a heartbeat sensed within a time segment from each of the sub-time ranges, and calculate the heartbeat interval based on the selected heartbeat.

[0006] In some embodiments, the first machine learning model is a convolutional neural network, the second machine learning model is a converter, and the third machine learning model is a multi-layer perceptron.

[0007] In some embodiments, the processing circuit is used to identify multiple R waves in the electrocardiogram signal, calculate multiple RR intervals between the R waves, and calculate multiple electrocardiogram feature values ​​based on the RR intervals. The processing circuit is used to identify multiple heartbeats from the light volume change signal, calculate multiple heartbeat intervals based on the heartbeats, and calculate multiple light volume change feature values ​​based on the heartbeat intervals. The processing circuit is used to input the electrocardiogram feature values ​​and the light volume change feature values ​​into the machine learning model to generate a heart assessment result.

[0008] In some embodiments, the processing circuit is used to identify multiple heartbeat cycles in the electrocardiogram signal and align the heartbeat cycles according to the R wave to generate a sample cycle. The processing circuit is used to identify the sample P wave, sample Q wave, sample R wave, sample S wave, and sample T wave in the sample cycle. The processing circuit is used to calculate the time difference between the sample R wave and the sample P wave, sample Q wave, sample S wave, or sample T wave as the electrocardiogram characteristic value.

[0009] In some embodiments, the processing circuit is used to calculate the difference between a baseline amplitude and a sample P wave, a sample Q wave, a sample R wave, a sample S wave or a sample T wave as another electrocardiogram characteristic value. The baseline amplitude is defined, for example, as the median of all amplitudes in the sample period.

[0010] In some embodiments, the cardiac assessment results include a survival curve and a risk score, wherein the survival curve includes a plurality of probability values ​​of not developing heart failure over a time axis.

[0011] In some embodiments, the processing circuit is used to generate a risk score according to the probability value at a preset time on the time axis.

[0012] In order to make the above features and advantages of the present invention more clearly understood, embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 is a schematic diagram of a system for evaluating cardiac conditions according to an embodiment.

[0014] Figure 2 FIG. 4 is a flow chart showing the method of generating a cardiac assessment result according to the first embodiment.

[0015] Figure 3 is a schematic diagram illustrating processing a PPG signal according to an embodiment.

[0016] Figure 4 is a schematic diagram illustrating a heart assessment result according to the first embodiment.

[0017] Figure 5 FIG. 4 is a flow chart showing a method for generating a cardiac assessment result according to the second embodiment.

[0018] Figure 6 is a schematic diagram illustrating the calculation of ECG characteristic values ​​according to the second embodiment.

[0019] Component number description

[0020] 100 Evaluation System

[0021] 110 ECG sensor

[0022] 120 Optical Volume Change Sensor

[0023] 130 Processing circuit

[0024] 140 users

[0025] Steps 201-206, 501-503

[0026] 310 PPG signal

[0027] 320 Time Segments

[0028] 321 Contraction Peak

[0029] 330 sequence data

[0030] 410, 420 curve

[0031] 421 Preset time

[0032] 422 probability value

[0033] 610 R wave

[0034] Steps 620, 630

[0035] 640 Sample Cycles

[0036] 641 Sample P wave

[0037] 642 Sample Q Wave

[0038] 643 Sample R wave

[0039] 644 Sample S wave

[0040] 645 Sample T wave DETAILED DESCRIPTION

[0041] The terms “first”, “second”, etc. used in this document do not particularly refer to an order or sequence, but are only used to distinguish components or operations described with the same technical terms.

[0042] Figure 1 is a schematic diagram of a cardiac condition assessment system according to an embodiment. Figure 1In this embodiment, the evaluation system 100 is a wearable device, such as a smart watch or a wristband, which is worn on the user 140. The evaluation system 100 includes an electrocardiogram sensor 110, a photoplethysmography sensor 120, and a processing circuit 130. The electrocardiogram sensor 110 is used to obtain an electrocardiogram (ECG) signal about the user 140, and the photoplethysmography sensor 120 is used to obtain a photoplethysmography (PPG) signal of the user 140. The PPG signal is based on a light source and a light sensor, does not need to directly contact the user 140, and has the advantage of relatively low power consumption, so it is suitable for long-term continuous sensing. In this embodiment, the PPG signal sensed in the first time and the ECG signal sensed in the second time are obtained. The first time is longer than the second time, for example, the first time is several hours, one or more days, etc., and the second time is, for example, tens of seconds. In this way, the PPG signal can provide a longer time and more comprehensive information, while the ECG provides relatively accurate heart information. Next, the processing circuit 130 executes one or more machine learning algorithms based on the PPG signal and the ECG signal to generate a cardiac assessment result. This cardiac assessment result, for example, includes a risk score, which can be provided to the user 140 in the form of images, voice, text, etc., to indicate the possibility of heart failure. In other embodiments, the cardiac assessment result may also include a segment where an ECG abnormality occurs, the time point of an abnormal heart rate, or an assessment report may be generated based on this information. The present invention does not limit the content and presentation form of the cardiac assessment result (which may be images, voice, text, or digital data, etc.). The following is divided into two embodiments to illustrate how to use a machine learning model to generate a cardiac assessment result.

[0043] First embodiment

[0044] Figure 2 is a flow chart showing the generation of cardiac assessment results according to the first embodiment. Figure 2 In step 201, the ECG signal sensed within the second time is obtained and pre-processed. In this embodiment, the second time is 30 seconds, and the pre-processing may include bandpass filtering, noise removal, etc. In step 202, the ECG signal is input into a first machine learning model to generate a first feature vector. This first machine learning model is, for example, a convolutional neural network, and its architecture can adopt LeNet, AlexNet, VGG, GoogLeNet, ResNet, DenseNet or YOLO (You Only Look Once), etc., and the present invention is not limited to this.

[0045] On the other hand, in step 203, the PPG signal sensed within the first time is obtained and the PPG signal is pre-processed. In this embodiment, the first time is 24 hours, and the pre-processing may include filtering, denoising, etc. In step 204, multiple heartbeats are identified from the PPG signal, and the heartbeat interval is calculated. Specifically, the PPG signal reflects the changes in blood flow in the blood vessels. When the heart contracts, blood is sent to various parts, which causes the blood flow and pressure of the blood vessels to increase temporarily, thereby affecting the absorption or scattering of light passing through the blood vessels. The systolic peak of the blood vessel can be detected based on the PPG signal. Each systolic peak represents a heartbeat, and then the time distance between two adjacent systolic peaks is calculated as the heartbeat interval. In some embodiments, the first time can be divided into multiple sub-time ranges, and a heartbeat sensed within a time segment is selected from each sub-time range, and then the heartbeat interval is calculated based on the selected heartbeat. For example, Figure 3 is a schematic diagram illustrating processing of a PPG signal according to an embodiment. The PPG signal 310 is obtained by continuous sensing within 24 hours. In this embodiment, each sub-time range is 1 hour, and the time segment is, for example, 5 minutes. Therefore, the PPG signal 310 is divided into 24 sub-signals, and then a 5-minute time segment is selected from each sub-signal. In some embodiments, signal quality indices (SQI) can be used to select a 5-minute signal with better quality from each hour's PPG signal. Therefore, a total of 24 time segments 320 are generated, and each time segment 320 includes multiple contraction peaks 321. The time distance between these contraction peaks 321 is the heartbeat interval. The heartbeat intervals calculated from these time segments 320 are connected together to form sequence data 330. Figure 3 The horizontal axis represents time, and the vertical axis represents the value of the heartbeat interval. It is worth noting that since different time segments are not continuous, the contraction peak 321 between two time segments is not used to calculate the heartbeat interval. In other embodiments, the above sub-time range can also be 30 minutes, two hours, etc., and the length of the above time segment can also be 1, 2, 3, 4 minutes, etc.

[0046] Next, step 205 is performed to input the calculated heartbeat interval (i.e., sequence data 330) into a second machine learning model to generate a second feature vector. The second machine learning model is, for example, a transformer, but in other embodiments, it may also be a recurrent neural network, etc. In some embodiments, the PPG signal 310 may be directly input into the second machine learning model to generate a second feature vector instead of dividing the PPG signal into a plurality of sub-signals.

[0047] In step 206, the first feature vector generated in step 202 and the second feature vector generated in step 205 are connected together and then input into a third machine learning model to generate a cardiac assessment result. The third machine learning model may be a multilayer perceptron (MLP), a fully connected layer, etc. In some embodiments, the third machine learning model may also include a softmax function, etc. In some embodiments, personal information about the user 140 may also be input into the third machine learning model. The personal information may include age, gender, weight, height, medical history, etc., which is not limited by the present invention.

[0048] Figure 4 is a schematic diagram illustrating the cardiac assessment result according to the first embodiment. In this embodiment, the output of the third machine learning model is a vector, each element of which represents the probability of heart failure occurring at a specific time in the future, such as Figure 4 As shown in the curve 410, the horizontal axis is time, and the vertical axis is the probability of heart failure. Subtracting 1 from the probability of heart failure can get the probability of not having heart failure, and after adding them up, we can get the probability of not having heart failure in the future, which is called the survival curve 420. The horizontal axis is time, and the vertical axis is the cumulative probability. In other words, the survival curve 420 contains multiple probability values ​​of not having heart failure on the time axis. If a preset time 421 (for example, 5 years) is taken on the time axis, the corresponding probability value 422 indicates the probability that the user can survive for 5 years without developing heart failure. A risk score can be generated based on the probability value 422, for example Figure 4 The risk score is "95", where the probability value 422 is positively correlated with the risk score. Any linear or nonlinear function can be used here to convert the probability value 422 into a risk score. Alternatively, the risk score can be set to four categories, namely good, ordinary, mild risk, and severe risk. When the probability value 422 is greater than 0.8, the corresponding risk score is good; when the probability value 422 is between 0.64 and 0.8, the corresponding risk score is ordinary; when the probability value 422 is between 0.44 and 0.64, the corresponding risk score is mild risk; when the probability value 422 is less than 0.44, the corresponding risk score is severe risk. The above method is only an example. The above risk score can be a numerical value, or text, a specific pattern or color, and the present invention is not limited to this.

[0049] Second embodiment

[0050] In the first embodiment, the machine learning model is used to extract the feature vector in the signal. However, in the second embodiment, the feature vector is first calculated and then input into the machine learning model.

[0051] Figure 5 is a flow chart showing the generation of cardiac assessment results according to the second embodiment. Figure 5 , wherein steps 201, 203, and 204 are the same as those in the first embodiment and are not repeated here. In step 501, the ECG characteristic value is calculated. Figure 6 : is a schematic diagram illustrating the calculation of ECG characteristic values ​​according to the second embodiment. In the ECG signal, a heartbeat includes a P wave, a Q wave, an R wave, an S wave, and a T wave. First, multiple R waves 610 in the ECG signal are identified. This can be achieved by any known algorithm such as determining whether the amplitude of the ECG signal is greater than a critical value, and the present invention is not limited to this. Next, the time interval between these R waves is calculated, hereinafter referred to as the RR interval. Since one R wave represents one heartbeat, the RR interval also represents the time of one heartbeat. Next, multiple ECG characteristic values ​​can be calculated based on these RR intervals. For example, the average heart rate can be calculated according to the following mathematical formula 1.

[0052] Mathematical formula 1

[0053]

[0054] Where MEAN_HR represents the average heart rate, i is a positive integer, RRI[i] represents the i-th RR interval, and K is the number of RR intervals. In addition, the standard deviation of the RR interval can be calculated according to the following mathematical formula 2.

[0055] Mathematical formula 2

[0056]

[0057] Where SDNN is the standard deviation, is the expected value of all RR intervals. In some embodiments, the difference between two adjacent RR intervals can also be calculated, expressed as ΔRRI[i], and then two standard deviations SD1 and SD2 can be calculated according to the following mathematical formulas 3-5, where is the expected value of the difference ΔRRI[i] between RR intervals. The ratio SD1 / SD2 can be regarded as an ECG feature value.

[0058] Mathematical formula 3

[0059]

[0060] Mathematical formula 4

[0061]

[0062] Mathematical formula 5

[0063]

[0064] Next, in step 620, the cardiac cycle in the ECG signal is identified. For example, the middle point between two adjacent R waves can be used as the boundary between two heartbeats. Alternatively, a cardiac cycle can be obtained by extending a distance forward and backward from the R wave, and the present invention is not limited to this. Here, each cardiac cycle is padded so that all cardiac cycles have the same length.

[0065] In step 630, the above-mentioned heartbeat cycle is aligned according to the R wave, and then the median is taken to generate a sample cycle 640. For example, the time points of the maximum amplitude in the R wave can be aligned, and the median of the amplitude of these R waves is taken at each time point. In this way, the sample cycle 640 can be obtained. Next, the sample P wave 641, the sample Q wave 642, the sample R wave 643, the sample S wave 644, and the sample T wave 645 of the sample cycle 640 are identified. This can be done by any conventional method, and the characteristics of the shape can be calculated based on this information. Specifically, the time difference between the sample R wave and the sample P wave, the sample Q wave, the sample S wave or the sample T wave can be calculated according to the following mathematical formula 6 as an ECG characteristic value.

[0066] Mathematical formula 6

[0067] Timing j =Rj,wherejisP,Q,S,T

[0068] Where R represents the time of the sample R wave, and j represents the time of the sample P wave, sample Q wave, sample S wave or sample T wave. j Represents the time difference.

[0069] On the other hand, the amplitude difference between the baseline amplitude and the sample P wave, sample Q wave, sample R wave, sample S wave or sample T wave may be calculated as the ECG characteristic value according to the following mathematical formula 7.

[0070] Mathematical formula 7

[0071] Amplitude j =template[j]-baseline,wherejisP,Q,R,S,T

[0072] Amplitude j is the amplitude difference. baseline, is the baseline amplitude. template[j] represents the amplitude of a template P wave, template Q wave, template R wave, template S wave, or template T wave. The definition of the baseline amplitude is, for example, the median of all amplitudes in the template period.

[0073] In summary, the above average heart rate MEAN_HR, standard deviation SDNN, ratio SD1 / SD2, time difference Timing j , Amplitude j In other embodiments, other ECG feature values ​​may be added according to other algorithms, and the present invention does not limit this.

[0074] Please refer to the reply Figure 5 In step 502, the light volume change characteristic value (called PPG characteristic value) is calculated according to the heartbeat interval. Similar to the first embodiment, the heartbeat interval is generated according to the PPG signal in step 204. These heartbeat intervals can be substituted into the RRI[i] in the above mathematical formulas 1 and 2, and in this way, the MEAN_HR and SDNN corresponding to the PPG signal can be calculated as the PPG characteristic value. In addition, the difference between adjacent heartbeat intervals is expressed as ΔBB[i]. The root mean square error RMSSD can be calculated as the PPG characteristic value by the following mathematical formula 8, where K represents the number of heartbeats in the PPG signal.

[0075] Mathematical formula 8

[0076]

[0077] Similarly, by substituting ΔBB[i] into ΔRRI[i] in the above mathematical formula 3-5, the ratio SD1 / SD2 of the PPG signal can be calculated as a PPG feature value. In some embodiments, π.SD1.SD2 can also be calculated as the PPG feature value. In some embodiments, the ratio between several consecutive differences ΔBB[i] greater than 50 microseconds in the PPG signal and the value K can also be calculated, expressed as PNN50 as the PPG feature value. In some embodiments, the sample entropy of the PPG signal can also be calculated as the PPG feature value.

[0078] In step 503, the ECG feature values ​​calculated in step 501 and the PPG feature values ​​calculated in step 502 are connected together (to form a vector) and input into a machine learning model to generate a cardiac assessment result. The machine learning model may be a decision tree, a random forest, a multi-level neural network, a convolutional neural network, a support vector machine, etc., and the present invention is not limited thereto. In addition, the cardiac assessment result in the second embodiment may be the same as that in the first embodiment, both including a survival curve and a risk score. In some embodiments, the machine learning model in step 503 directly outputs a survival curve instead of Figure 4 Curve 410.

[0079] Through the above content, the PPG signal can be used to obtain the user's status for the whole day, providing more comprehensive heart information, while the ECG signal provides relatively short and more accurate heart information. Combining these two information can better evaluate the heart status.

[0080] Although the present invention has been disclosed as above by way of embodiments, it is not intended to limit the present invention. Any person having ordinary knowledge in the technical field may make some changes and modifications without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the scope defined by the attached patent application scope.

Claims

1. A system for evaluating heart condition, characterized in that: include: An electrocardiogram sensor for obtaining an electrocardiogram signal of a user; a photovolume variation sensor for obtaining a photovolume variation signal related to the user; as well as A processing circuit is electrically connected to the electrocardiogram sensor and the photovolume variation sensor, and is used to generate a cardiac assessment result based on the photovolume variation signal sensed within a first time and the electrocardiogram signal sensed within a second time, wherein the first time is longer than the second time.

2. The evaluation system according to claim 1, characterized in that The processing circuit is used to input the electrocardiogram signal into a first machine learning model to generate a first feature vector. wherein the processing circuit identifies a plurality of heartbeats from the light volume variation signal, calculates a plurality of heartbeat intervals according to the plurality of heartbeats, and inputs the plurality of heartbeat intervals into a second machine learning model to generate a second feature vector, The processing circuit is used to input the first feature vector and the second feature vector into a third machine learning model to generate the heart assessment result.

3. The evaluation system according to claim 2, characterized in that The first time includes a plurality of sub-time ranges, and the processing circuit is used to select the plurality of heartbeats sensed within a time segment from each of the plurality of sub-time ranges, and calculate the plurality of heartbeat intervals according to the selected plurality of heartbeats.

4. The evaluation system according to claim 2, characterized in that The first machine learning model is a convolutional neural network, the second machine learning model is a converter, and the third machine learning model is a multilayer perceptron.

5. The evaluation system according to claim 1, characterized in that: The processing circuit is used to identify a plurality of R waves in the electrocardiogram signal, calculate a plurality of RR intervals between the plurality of R waves, and calculate a plurality of electrocardiogram characteristic values ​​according to the plurality of RR intervals. wherein the processing circuit is used to identify a plurality of heartbeats from the light volume change signal, calculate a plurality of heartbeat intervals according to the plurality of heartbeats, and calculate a plurality of light volume change characteristic values ​​according to the plurality of heartbeat intervals, The processing circuit is used to input the multiple electrocardiogram feature values ​​and the multiple photovolume variation feature values ​​into a machine learning model to generate the heart assessment result.

6. The evaluation system according to claim 5, characterized in that The processing circuit is used to identify a plurality of heartbeat cycles in the electrocardiogram signal, and align the plurality of heartbeat cycles according to the plurality of R waves to generate a sample cycle, wherein the processing circuit is used to identify a sample P wave, a sample Q wave, a sample R wave, a sample S wave, and a sample T wave in the sample cycle, The processing circuit is used to calculate the time difference between the sample R wave and the sample P wave, the sample Q wave, the sample S wave or the sample T wave as one of the multiple electrocardiogram characteristic values.

7. The evaluation system according to claim 6, characterized in that: The processing circuit is used to calculate the amplitude difference between a reference amplitude and the sample P wave, the sample Q wave, the sample R wave, the sample S wave or the sample T wave as another electrocardiogram characteristic value among the multiple electrocardiogram characteristic values.

8. The evaluation system according to claim 7, characterized in that The first time includes a plurality of sub-time ranges, and the processing circuit is used to select the plurality of heartbeats sensed within a time segment from each of the plurality of sub-time ranges, and connect the selected plurality of heartbeats together to calculate the plurality of heartbeat intervals.

9. The evaluation system according to claim 1, characterized in that: The heart assessment result includes a survival curve and a risk score. The survival curve includes a plurality of probability values ​​of not having heart failure on a time axis.

10. The evaluation system according to claim 9, characterized in that The processing circuit is used for generating the risk score according to the probability value at a preset time on the time axis.

Images