Prediction device

Abstract

To provide a technique that improves the accuracy of predicting an abnormality by suppressing false positive detection.SOLUTION: A prediction device includes: an acquisition unit 312 for acquiring RRI data; a generation unit 314 for generating heartbeat fluctuation index data based on the RRI data; a first processing unit 316 for detecting a possibility of an abnormality occurring in a subject by inputting the RRI data to an auto-encoder; and a second processing unit 320 for predicting occurrence of an abnormality in the subject based on the heartbeat fluctuation parameter data when the possibility of an abnormality occurring in the subject is detected in the first processing unit 316.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present disclosure relates to a prognostic technique, and in particular to a prognostic device that uses RR interval (RRI) data. [Background technology]

[0002] Epilepsy is one of the most common neurological disorders in the world, affecting an estimated 60 million people. The problem with epilepsy is unpredictable seizures. Seizures can cause loss of consciousness or convulsions. These can lead to death in traffic accidents, drowning, or asphyxiation during sleep. A seizure can also lead to cardiac arrest, a condition known as sudden death after a seizure (SUDEP).

[0003] Antiepileptic drugs (AEDs) are increasingly being developed and prescribed to control seizures. However, approximately 30% of epilepsy patients do not respond to drug treatment, and serious side effects may occur. It is also important to note that AEDs cannot be used in women who may become pregnant. Meanwhile, epilepsy surgery is not always feasible. Identifying the cause of seizures and surgically removing it is difficult, and only 47% of patients remain seizure-free after 10 years.

[0004] Electroencephalography (EEG) measurements, which monitor brain waves using electrodes attached to the patient's scalp, have been shown to predict seizures with a 100% success rate and a false alarm rate of 0.11 ± 0.02 per hour using various methods, including long short-term memory (LSTM) models. However, the lack of widely available portable EEG equipment means that EEG measurements are limited to hospital-based testing.

[0005] It has been reported that epileptic seizures may be predicted by changes in the autonomic nervous system, and interest in changes in cardiac-related signals has been growing. Attention has primarily focused on changes in heart rate variability (HRV) and RR intervals (RRI). For example, seizures can be predicted from abnormalities in HRV data using multivariate statistical process control (MSPC), which is commonly used in process control (see, for example, Non-Patent Document 1). [Prior art documents] [Non-patent literature]

[0006] [Non-Patent Document 1] K. Fujiwara et al., "Epileptic seizure prediction based on multivariate statistical process control of heart rate variability features", IEEE Transactions on Biomedical Engineering, 2016, vol. 63, No 6, p.1321-1332 Summary of the Invention [Problem to be solved by the invention]

[0007] A self-attention autoencoder (SA-AE) can be used to predict epileptic seizures from RRI data. This model reconstructs RRI data and uses the reconstruction error (RE) to predict seizures. However, the SA-AE model also generates many false positives (FPs), which reduces the usability of the prediction system. Therefore, it is necessary to further suppress FPs while maintaining sensitivity.

[0008] The present disclosure has been made in light of these circumstances, and aims to provide a technology that improves the accuracy of predicting abnormalities by suppressing false positive detection. [Means for solving the problem]

[0009] In order to solve the above problems, a prediction device according to one embodiment of the present disclosure includes an acquisition unit that acquires RRI data indicating the interval between R waves in the electrocardiogram signal of a subject; a generation unit that generates heart rate variability index data including index values ​​for each of multiple types of heart rate-related indexes based on the RRI data acquired by the acquisition unit; a first processing unit that inputs the RRI data acquired by the acquisition unit as input data to an autoencoder, and then acquires the input data reconstructed by the autoencoder as output data, and detects the possibility of an abnormality occurring in the subject if the error between the input data and the output data exceeds a threshold value for a certain period of time; and a second processing unit that, if the first processing unit detects the possibility of an abnormality occurring in the subject, predicts the occurrence of an abnormality in the subject based on the heart rate variability index data generated by the generation unit.

[0010] Any combination of the above components, and conversion of the present disclosure into a method, device, system, recording medium, computer program, etc., are also valid aspects of the present disclosure. [Effects of the Invention]

[0011] According to the present disclosure, the accuracy of predicting abnormalities can be improved by suppressing false positive detection. [Brief explanation of the drawings]

[0012] [Figure 1] FIG. 1 is a diagram illustrating the configuration of a prediction system according to an embodiment of the present invention. [Figure 2] FIG. 2 is a diagram illustrating the configuration of the prediction device of FIG. 1. [Figure 3] 3 is a diagram showing an R wave contained in a signal received by the communication unit of FIG. 2. FIG. [Figure 4] 3 is a diagram illustrating the configuration of an autoencoder in the first processing unit in FIG. 2. FIG. [Figure 5] 3 is a diagram showing the independent characteristics of a binary classifier in the second processing unit of FIG. 2. FIG. [Figure 6] 6(a)-(b) are diagrams showing prediction results obtained by the prediction device of FIG. [Figure 7]2 is a flowchart showing a prediction procedure by the prediction system of FIG. 1. DETAILED DESCRIPTION OF THE INVENTION

[0013] Before describing the present disclosure in detail, an overview will be provided. This embodiment relates to a prediction system for predicting abnormal occurrences in a subject, such as epileptic seizures. Epilepsy is a brain disorder characterized by recurrent epileptic seizures due to excessive neural activity. Epileptic seizures affect cardiac autonomic nervous activity before a seizure. Because epileptic patients are at risk of accidents and injuries due to seizures, it is important to develop an epileptic seizure prediction algorithm to prevent accidents caused by seizures.

[0014] The prediction system according to this embodiment predicts epileptic seizures through a two-stage process. In the first stage, the subject's RRI data is input into a self-attention autoencoder (SA-AE) to detect the possibility of an epileptic seizure. In the second stage, a heart rate variability (HRV) index generated from the RRI data is input into a binary classifier to determine whether the possibility of an epileptic seizure detected in the first stage is a true seizure prediction or a false positive. The prediction system predicts an epileptic seizure when it determines that the possibility of an epileptic seizure is a true seizure prediction.

[0015] 1 shows the configuration of a prediction system 1000. The prediction system 1000 includes a plurality of electrodes 100, a heart rate sensor 200, and a prediction device 300.

[0016] The subject 10 is a person for whom epileptic seizures are to be predicted. Three electrodes 100 are attached to the body surface of the subject 10. The three electrodes 100 are composed of, for example, a positive electrode, a negative electrode, and a ground electrode. The three electrodes 100 are connected to a heartbeat sensor 200. The heartbeat sensor 200 extracts R waves from an electrocardiogram signal measured for the subject 10 via the three electrodes 100. The heartbeat sensor 200 has a wireless communication function such as Bluetooth (registered trademark) LE, and wirelessly transmits a signal indicating the extracted R waves.

[0017] The prediction device 300 is, for example, a computer or a smartphone. The prediction device 300 has a wireless communication function such as Bluetooth (registered trademark) LE, and receives a signal indicating an R wave wirelessly transmitted from the heart rate sensor 200. The prediction device 300 predicts an epileptic seizure in the subject. When the prediction device 300 predicts an epileptic seizure, it outputs an alarm sound.

[0018] Figure 2 shows the configuration of the prediction device 300. The prediction device 300 includes a communication unit 310, an acquisition unit 312, a generation unit 314, a first processing unit 316, a memory unit 318, a second processing unit 320, and an output unit 322. The communication unit 310 receives a signal indicating an R wave wirelessly transmitted from the heartbeat sensor 200. Figure 3 shows the R wave included in the signal received by the communication unit 310. The potential distribution generated by the electrical activity of the heart is detected as a potential difference between two points of electrodes 100 placed on the body surface or inside the body. The waveform representing the potential difference between two points that fluctuates over time is called an ECG. The ECG shows periodic potential changes and several peaks, the highest of which is called an R wave. Return to Figure 2.

[0019] The communication unit 310 generates data indicating R waves (hereinafter referred to as "R wave data") from the received signal and outputs the data to the acquisition unit 312. The R wave data is digital data that is, for example, "1" when the amplitude of the signal indicating the R wave is equal to or greater than a threshold voltage, and is "0" when the amplitude is less than the threshold voltage. In other words, the R wave data is a rectangular pulse train in which the period corresponding to the R wave in the electrocardiogram signal is set to "1" and the other periods are set to "0."

[0020] The acquiring unit 312 receives R-wave data from the communication unit 310. Based on the R-wave data, the acquiring unit 312 acquires RRI data, which is time-series data of RRI variables indicating the interval between R waves. The RRI data indicates the interval between R waves in the electrocardiogram signal of the subject 10. In this case, the acquiring unit 312 calculates, from the R-wave data, the time interval between the falling edges of two temporally adjacent rectangular pulses as an RRI variable, and generates the RRI data by arranging the calculated RRI variables in time series.

[0021] Specifically, the acquisition unit 312 first calculates, for the R-wave data, the time interval between a first time when an R wave is detected and a second time when the immediately preceding R wave is detected, as the RRI value at the first time. This generates data indicating a change in the RRI value over time. The acquisition unit 312 then performs spline interpolation on the data indicating the change in the RRI value over time and samples the data at equal intervals, thereby generating RRI data indicating the change in the RRI value over time, the RRI data being arranged at equal intervals in time. That is, the acquisition unit 312 performs spline interpolation (interpolation processing) on ​​time-varying data obtained from an electrocardiogram signal at different time intervals and samples the data at equal intervals, thereby generating RRI data being arranged at equal intervals in time. The acquisition unit 312 outputs the RRI data to the generation unit 314 and the first processing unit 316.

[0022] The generation unit 314 generates heart rate variability (HRV) index data including index values ​​for multiple types of heart rate-related indexes based on the RRI data acquired by the acquisition unit 312. The HRV index data reflects autonomic nervous activity, and HRV refers to fluctuations in the RRI of an ECG on the order of milliseconds. When the heart beats quickly, the RRI becomes shorter. The HRV index data is classified into time domain indexes found from the RRI on the time axis and frequency domain indexes found from the frequency density of the RRI.

[0023] Examples of time domain indices include RMSSD, RRI standard deviation, SDSD, NN50, pNN50, RRI average, HR average, and total power. Time domain HRV index data is calculated directly from RRI data. Among the time domain HRV index data, RMSSD is the square root of the mean square of the difference between two temporally adjacent RRI data for RRI data included within a time window. RRI standard deviation is the standard deviation of the RRI data included within the time window. SDSD is the fdsa. NN50 is the number of pairs of two temporally adjacent RRI data for RRI data included within the time window whose difference exceeds 50 msec. pNN50 is a value obtained by dividing NN50 by the total number of RRI data within the time window. RRI average is the average RRI value within the time window, and HR average is the average heart rate within the time window. Total power is the sum of the squares of the RRI data included within the time window.

[0024] The frequency domain indices include HF, LF, VLF, LF / HF ratio, HF norm, and LF norm. The frequency domain HRV index data is calculated from the power spectrum density (PSD) obtained from the RRI data. The PSD is calculated using a Fourier transform or an autoregression (AR) model. The generation unit 314 generates, for example, the integral value of 0.15 Hz to 0.4 Hz (first frequency band) in the power spectrum as the HF. The generation unit 314 generates, for example, the integral value of 0.04 Hz to 0.15 Hz (second frequency band) in the power spectrum as the LF. The LF / HF ratio is a value obtained by dividing LF by HF. Here, for example, 13 indices, including the time domain indices and the frequency domain indices, are used as the HRV index data. The generation unit 314 outputs the HRV index data to the second processing unit 320.

[0025] The first processing unit 316 inputs the RRI data acquired by the acquisition unit 312 into an autoencoder as input data. An autoencoder is a deep learning model trained using training data to reconstruct input data and output values ​​close to the input data. This is a neural network model commonly used in unsupervised learning, and it can be said that the output approximates the input via a latent function in the hidden layer. One example of an autoencoder is the self-attention autoencoder (SA-AE). Self-attention autoencoders are particularly useful for time series analysis and are AE models that extend the self-attention mechanism (SA), the basis of the Transformer architecture, to the hidden layer. The self-attention mechanism learns important dependencies within the input time series data, enabling the AE model to improve its predictive performance. Hereinafter, self-attention autoencoders may be simply referred to as "autoencoders." When data with trends different from the training data is input to such an autoencoder, the error between the output data and the input data increases. In other words, when abnormal data is input to an autoencoder trained using normal data as training data, the error increases.

[0026] The first processing unit 316 pre-trains the autoencoder using RRI data acquired from electrocardiogram signals during normal times (interictal periods) as training data. In other words, the training data includes RRI data during normal times, but does not include RRI data during epileptic seizures. Figure 4 shows the configuration of the autoencoder in the first processing unit 316. The autoencoder is a neural network including multiple nodes 408, which are classified into an input layer 400, a hidden layer 402, and an output layer 404. The autoencoder shown in Figure 4 is an example, and an autoencoder with four or more layers may be used. Here, the input data input to the input layer 400 is x1, x2, ..., x N and the output data output from the output layer 404 is X1, X2, . . . , X N x1, x2, . . ., x Nindicates each of the multiple types of HRV indices included in the HRV index data. In the learning stage, the first processing unit 316 derives the coefficients of each node 408 by setting the multiple types of HRV indices included in the HRV index data as input data and output data. In other words, the autoencoder is trained so that the output data obtained by inputting the HRV index data becomes closer to the HRV index data. The first processing unit 316 stores the derived coefficients of each node 408 in the storage unit 318. Return to FIG. 2.

[0027] The first processing unit 316 inputs the RRI data as input data to the autoencoder, and then obtains the input data reconstructed in the autoencoder as output data. The first processing unit 316 calculates the reconstruction error RE between the input data and the output data as follows:

number

[0028] When the first processing unit 316 detects a possibility of an epileptic seizure in the subject 10, the second processing unit 320 inputs the HRV index data generated by the generation unit 314 to the binary classifier. On the other hand, when the first processing unit 316 does not detect a possibility of an epileptic seizure in the subject 10, the second processing unit 320 does not input the HRV index data generated by the generation unit 314 to the binary classifier. Therefore, in this case, the second processing unit 320 does not operate.

[0029] The binary classifier is used to exclude false positives from the possibility of an epileptic seizure detected by the first processing unit 316. Examples of binary classifiers that can be used include random forest, XGBoost, logistic regression (LogReg), k-nearest neighbor (KNN), and gradient boosting (GBM). FIG. 5 shows the individual characteristics of the binary classifier in the second processing unit 320. According to this, LogReg exhibits the highest values ​​for sensitivity, specificity, and AUC. Specifically, these values ​​are 0.982, 0.547, and 0.954. Therefore, LogReg may be adopted as the binary classifier. Returning to FIG. 2, the binary classifier is also trained in advance. The binary classifier determines whether the possibility of the subject 10 having an epileptic seizure is true or false. If the binary classifier determines that the subject 10 has a possibility of an epileptic seizure, the second processing unit 320 predicts the subject 10 having an epileptic seizure. When the second processing unit 320 predicts an epileptic seizure, the output unit 322 outputs an alarm sound.

[0030] The following describes the evaluation of the prediction characteristics of the prediction system 1000 according to this embodiment. Data from a total of 126 focal epilepsy patients was used. The data collected from these patients consisted of both clinical video EEG and ECG. The training data for the autoencoder consisted of only seizure-free recordings, totaling 2,387 hours including 1,985 interictal episodes. The validation data, including both seizure and non-seizure recordings, totaled 164 episodes and 185 hours. The test data totaled 1,132 episodes and 1,278 hours. The types of focal epilepsy varied from patient to patient, including focal impaired consciousness seizures and focal to bilateral tonic-clonic seizures (FBTCS).

[0031] The first processing unit 316 set the threshold to 99% and the fixed period to 8 seconds, determined by grid search. The autoencoder detected 541 possible seizures, of which 110 were true positives and 433 were false positives. The performance of the autoencoder alone was 0.71 sensitivity, 0.56 specificity, and 0.67 FPR / hour.

[0032] Using the 541 seizure probabilities detected by the first processing unit 316, the data was divided into training and testing sets in a ratio of 8:2, and the second processing unit 320 performed learning and testing of a binary classifier. Hyperparameters were adjusted using a grid search method. To maintain the sensitivity of seizure prediction while suppressing the occurrence of false positives, the classification threshold of each model was adjusted to maximize the sensitivity of the binary classifier, with a target specificity of approximately 0.50.

[0033] The performance of the combined autoencoder and binary classifier was evaluated using a test dataset. Here, sensitivity and false positive rate (FPR) [times / hour] were used as performance indicators. The sensitivity and FPR of this example were 0.67 and 0.24 times / hour, respectively, which was approximately one-third of the FPR of the autoencoder alone (0.67 times / hour).

[0034] 6(a)-(b) show prediction results obtained by the prediction device 300. The horizontal axis in FIG. 6(a)-(b) represents time, and the vertical axis represents reconstruction error. A first epileptic attack candidate 410, a second epileptic attack candidate 412, a third epileptic attack candidate 414, a fourth epileptic attack candidate 416, a fifth epileptic attack candidate 418, and a sixth epileptic attack candidate 420 are sections of possible epileptic attacks detected by the first processing unit 316. Of these, the first epileptic attack candidate 410 and the third epileptic attack candidate 414 were determined to be false positives by the second processing unit 320.

[0035] This configuration can be realized in hardware terms by any computer's CPU (Central Processing Unit), memory, and other LSIs (Large Scale Integration), and in software terms by programs loaded into memory, but here we depict functional blocks realized by the cooperation of these. Therefore, those skilled in the art will understand that these functional blocks can be realized in various forms by hardware alone or a combination of hardware and software.

[0036] The operation of the prediction system 1000 configured as described above will be described. FIG. 7 is a flowchart showing the prediction procedure by the prediction system 1000. The heart rate sensor 200 acquires an electrocardiogram signal (S10). The acquisition unit 312 acquires RRI data (S12). The generation unit 314 generates HRV index data (S14). The first processing unit 316 inputs the RRI data to a self-attention autoencoder (S16). If the reconstruction error RE exceeds a threshold value for a certain period of time (Y in S18), the second processing unit 320 inputs the HRV index data to a binary classifier (S20). If the binary classifier outputs true (Y in S22), the second processing unit 320 predicts an epileptic seizure (S24). If the reconstruction error RE does not exceed the threshold value for a certain period of time (N in S18) or if the binary classifier does not output true (N in S22), the process ends.

[0037] According to this embodiment, when the autoencoder detects the possibility of an epileptic seizure, the binary classifier determines whether the result is true or false, thereby suppressing false positive detection and improving the accuracy of epileptic seizure prediction. Furthermore, since the autoencoder is trained using RRI data acquired from normal electrocardiogram signals, the possibility of an epileptic seizure can be detected. Furthermore, if the possibility of an epileptic seizure is not detected, the HRV index data is not input to the binary classifier, enabling classification using the binary classifier.

[0038] An overview of one embodiment of the present disclosure is as follows: A prediction device of one embodiment of the present disclosure includes an acquisition unit that acquires RRI data indicating the intervals between R waves in an electrocardiogram signal of a subject, a generation unit that generates heart rate variability index data including index values ​​for multiple types of heart rate-related indexes based on the RRI data acquired by the acquisition unit, a first processing unit that inputs the RRI data acquired by the acquisition unit as input data to an autoencoder, and then acquires the input data reconstructed by the autoencoder as output data, and detects the possibility of an abnormality occurring in the subject if an error between the input data and the output data exceeds a threshold value for a certain period of time, and a second processing unit that predicts the occurrence of an abnormality in the subject based on the heart rate variability index data generated by the generation unit if the first processing unit detects the possibility of an abnormality occurring in the subject.

[0039] According to this aspect, when the autoencoder detects the possibility of an abnormality occurring, the occurrence of the abnormality in the subject is predicted based on the heart rate variability index data, thereby improving the accuracy of abnormality prediction by suppressing false positive detection.

[0040] The autoencoder in the first processing unit may be trained using RRI data acquired from electrocardiogram signals under normal conditions. In this case, since the autoencoder has been trained using RRI data acquired from electrocardiogram signals under normal conditions, it is possible to detect abnormalities.

[0041] The second processing unit does not execute the process if the first processing unit has not detected the possibility of an abnormality occurring in the subject. Since the second processing unit does not execute the process if the possibility of an abnormality occurring in the subject has not been detected, unnecessary processing can be omitted.

[0042] The present disclosure has been described above based on examples. These examples are merely illustrative, and it will be understood by those skilled in the art that various modifications are possible in the combination of the components and processing steps, and that such modifications are also within the scope of the present disclosure.

[0043] The prediction system 1000 in this embodiment predicts epileptic seizures as abnormal occurrences based on RRI data. However, this is not limited to this. For example, the prediction system 1000 may predict the onset of heat stroke as abnormal occurrences based on RRI data. In this case, data suitable for heat stroke is used as data for training the autoencoder and binary classifier. This modification expands the scope of application of this embodiment. [Explanation of symbols]

[0044] 10 Subject, 100 Electrode, 200 Heart rate sensor, 300 Prediction device, 310 Communication unit, 312 Acquisition unit, 314 Generation unit, 316 First processing unit, 318 Memory unit, 320 Second processing unit, 322 Output unit, 1000 Prediction system.

Claims

1. an acquisition unit that acquires RRI data indicating the interval between R waves in an electrocardiogram signal of a subject; a generating unit that generates heart rate variability index data including index values ​​for each of a plurality of types of indexes related to heart rate based on the RRI data acquired by the acquiring unit; a first processing unit that inputs the RRI data acquired by the acquisition unit into an autoencoder as input data, and then acquires the input data reconstructed by the autoencoder as output data, and detects the possibility of an abnormality occurring in the subject when an error between the input data and the output data exceeds a threshold value for a certain period of time; a second processing unit that predicts the occurrence of an abnormality in the subject based on the heart rate variability index data generated by the generating unit when the first processing unit detects a possibility of the occurrence of an abnormality in the subject; A prediction device comprising:

2. The prediction device according to claim 1 , wherein the autoencoder in the first processing unit is trained using RRI data obtained from an electrocardiogram signal under normal conditions.

3. The prediction device according to claim 1 or 2, wherein the second processing unit does not execute processing if the first processing unit does not detect the possibility of an abnormality occurring in the subject.

Images