KR102559333B1 - Cardiovascular disease prediction system using A.I.-based cardiovascular disease prediction model for patients with SDB - Google Patents

Abstract

The present invention relates to a cardiovascular disease prediction system that predicts future cardiovascular disease in sleep breathing disorder patients by applying the electrocardiogram and cardiovascular disease risk factors of patients with sleep breathing disorders to an artificial intelligence-based cardiovascular disease prediction model. will be.
In the present invention, in a cardiovascular disease prediction system comprising an operation processing unit for predicting the occurrence of cardiovascular disease using an electrocardiogram signal and a CVD risk factor detected during polysomnography, the operation processing unit comprises: From the electrocardiogram (ECG) signal, detecting time domain electrocardiogram parameters including ST-segment and T-wave (ST-T) changes (STTc) segments, and including low-frequency to high-frequency band intensity ratio (PLF / PHF ratio) a signal-processed electrocardiogram feature extractor for detecting frequency domain electrocardiogram parameters and storing averages and standard deviations of the detected time domain electrocardiogram parameters and frequency domain electrocardiogram parameters in a memory unit as 'signal processed electrocardiogram features'; An electrocardiogram signal for 30 seconds is input to the pre-learned CNN-based artificial intelligence model, and the mean and standard deviation of the results of each node of the flatten layer in the CNN-based artificial intelligence model are obtained, an artificial intelligence-based electrocardiogram feature extraction unit to store in a memory unit as 'artificial intelligence-based electrocardiogram feature'; a CVD risk factor loading unit that reads the CVD risk factors stored in the memory unit; Including cardiovascular disease (CVD) predictive models including servo vector machines (SVMs), including signal-processed ECG features, AI-based ECG features and CVD risk factors, to cardiovascular disease predictive models input, and from the cardiovascular disease prediction model, whether or not cardiovascular disease will occur in the future, and if a cardiovascular disease occurs, which disease among coronary heart disease (CHD), heart failure (HF), and stroke will occur and a cardiovascular disease (CVD) prediction unit that outputs the result as a cardiovascular disease prediction result.

Description

Cardiovascular disease prediction system using A.I.-based cardiovascular disease prediction model for patients with SDB}

The present invention provides a cardiovascular disease prediction model, which is an artificial intelligence (A.I.)-based algorithm using an electrocardiogram and CVD risk factors for predicting future cardiovascular disease in patients with sleep breathing disorder, and a cardiovascular disease prediction model using the same. It relates to a disease prediction system.

Sleep-disordered breathing (SDB) is partial or complete obstruction of the upper airway during sleep, resulting in apnea or hypopnea, a ventilatory disorder. Sleep breathing disorder is known to be about 2-4% of the total population, and is a very common disease known to be 3.2-4.5% in Korea. Sleep apnea-hypopnea that occurs repeatedly during sleep causes repetitive hypoxemia, reoxygenation, rapid changes in intrathoracic pressure, and arousal of the central nervous system, which act as an acute stress factor for the cardiovascular system. Therefore, if sleep breathing disorder is left untreated for a long time, the occurrence of hypertension, heart failure, myocardial infarction, arrhythmia, and stroke increases, increasing the risk of death due to cardiovascular disease.

To diagnose sleep breathing disorder (SDB), polysomnography is usually performed. Polysomnography is a test that records and analyzes sleep conditions using various instruments to diagnose various abnormal conditions that occur during sleep. In polysomnography, various examination equipments are mobilized to diagnose sleep disorders, such as electroencephalography (EEG) to check brain function status, safety level test (EOG) to check eye movements, and electromyography to check muscle conditions. (EMG), electrocardiogram (ECG) to check the heart rhythm, video shooting to see the overall condition, etc. are performed together, usually while sleeping for about one night.

The National Heart Lung & Blood Institute (NHLBI) conducted a multicenter cohort study (the Sleep Heart Health Study, SHHS) to confirm the association between sleep breathing disorders and cardiovascular disease. Gottlieb et al., in a prospective study examining the relationship between sleep breathing disorder, coronary artery disease, and heart failure through SHHS, found that patients with severe sleep breathing disorder had a 68% and 58% higher probability of developing coronary artery disease and heart failure, respectively, than normal subjects. reported as In addition, Redline et al. investigated the relationship between sleep breathing disorder and stroke through a prospective study of SHHS, and reported that patients with mild to moderate sleep breathing disorder showed a high correlation with ischemic stroke. In addition, studies on predictors of cardiovascular diseases such as cholesterol, blood pressure, obesity, smoking, and electrocardiogram have been actively conducted. Auer et al reported that ECG waveforms are related to cardiovascular disease, and ECG abnormality can be used as a predictor of coronary artery disease. However, in previous studies, only one disease was analyzed or group-level analysis was performed, such as disease incidence (morbidity) or mortality (mortality). Also, the usefulness of CVD predictors to predict future cardiovascular diseases has not been evaluated.

Therefore, a cardiovascular disease prediction system that predicts cardiovascular disease that will actually occur in the future using a cardiovascular disease predictor is desired.

Recently, interest in personalized medicine has been increasing, suggesting the possibility of personalized disease diagnosis and prediction by applying technologies such as artificial intelligence and big data in various fields.

In the present invention, an electrocardiogram and an electrocardiogram for predicting future (eg, within the next 10 years) cardiovascular diseases, for example, coronary heart disease, heart failure, and stroke of sleep breathing disorder patients We propose a cardiovascular disease prediction model, which is an artificial intelligence-based algorithm using CVD risk factors, and a cardiovascular disease prediction system that applies it.

As a prior art, Korean Patent Registration No. 10-1839910 discloses the subject's age, diabetes, hypertension, smoking, family history of coronary artery disease, blood pressure, cholesterol level, white blood cell count, creatine level, glycated hemoglobin level, and atrial fibrillation factors obtaining each measured value; assigning a risk prediction score corresponding to each obtained measurement value to each factor; and matching the total score obtained by adding the risk prediction scores of each factor to the probability of cardiovascular disease occurrence within the prediction period. In the case of Korean Patent Registration No. 10-1839910, the cardiovascular disease risk is predicted by adding the prediction scores according to the cardiovascular disease risk factors, and there is a problem with accuracy.

The present invention provides a cardiovascular disease prediction system that predicts future cardiovascular disease in sleep breathing disorder patients by applying the electrocardiogram and cardiovascular disease risk factors of sleep breathing disorder patients to an artificial intelligence-based cardiovascular disease prediction model. is to do

In order to solve the above problems, the present invention is a cardiovascular disease prediction system, including an arithmetic processing unit for predicting the occurrence of cardiovascular disease using an electrocardiogram signal and a CVD risk factor detected during polysomnography. The operation processing unit detects a time domain electrocardiogram parameter including a ST-segment and T-wave (ST-T) changes (STTc) segment from the electrocardiogram (ECG) signal, and a low-frequency to high-frequency band intensity ratio (PLF) /PHF ratio), detects the frequency domain electrocardiogram parameters, and stores the average and standard deviation of the detected time domain electrocardiogram parameters and frequency domain electrocardiogram parameters as 'signal processed electrocardiogram features' in a memory unit. electrocardiogram feature extraction unit; An electrocardiogram signal for 30 seconds is input to the pre-learned CNN-based artificial intelligence model, and the mean and standard deviation of the results of each node of the flatten layer in the CNN-based artificial intelligence model are obtained, an artificial intelligence-based electrocardiogram feature extraction unit to store in a memory unit as 'artificial intelligence-based electrocardiogram feature'; a CVD risk factor loading unit that reads the CVD risk factors stored in the memory unit; Including cardiovascular disease (CVD) predictive models including servo vector machines (SVMs), including signal-processed ECG features, AI-based ECG features and CVD risk factors, to cardiovascular disease predictive models input, and from the cardiovascular disease prediction model, whether or not cardiovascular disease will occur in the future, and if a cardiovascular disease occurs, which disease among coronary heart disease (CHD), heart failure (HF), and stroke will occur and a cardiovascular disease (CVD) prediction unit that outputs the result as a cardiovascular disease prediction result.

The cardiovascular disease (CVD) prediction unit consists of a pre-learned first servo vector machine to determine whether or not cardiovascular disease will occur in the future, and if it is determined that cardiovascular disease does not occur, a value indicating that cardiovascular disease does not occur. Cardiovascular disease (CVD) output as a predictive value, cardiovascular disease occurrence determination unit; If it is determined that future cardiovascular disease will occur in the cardiovascular disease occurrence determination unit, coronary heart disease (CHD) is classified as future cardiovascular disease using a one-to-one (OvO, One-Vs-One) multi-class classifier. ), heart failure (HF), stroke (stroke) selects one of the cardiovascular disease (CVD) predictive value, a detailed disease determination unit; includes.

The detailed disease discrimination unit selects one of coronary heart disease and heart failure with a servo vector machine pre-learned to discriminate one of coronary heart disease and heart failure, and outputs the result as a result of the second servo vector machine. boat vector machine; A third servo vector that is pre-learned to discriminate between coronary heart disease and stroke, and selects one of coronary heart disease and stroke and outputs it as a result of the third servo vector machine. machine; a fourth servo vector machine configured to select one of heart failure and stroke as a previously learned servo vector machine to discriminate between heart failure and stroke and output the result as a result of the fourth servo vector machine; a voting unit which counts and outputs the number of times each of coronary heart disease, heart failure, and stroke is selected from the result of the second servo vector machine, the result of the third servo vector machine, and the result of the fourth servo vector machine; include

The voting unit is configured to output values for coronary heart disease, heart failure, and diseases with the highest number of selected strokes as cardiovascular disease (CVD) prediction values.

The input vector of the first servo vector machine is the average of the result of the 15th node of the flatten layer, the average of the STTc segments, the BMI index, the systolic blood pressure, the AHI index, and the 11 of the flatten layer. The standard deviation of the result of the th node, the average of the low-frequency to high-frequency band intensity ratio (P _LF / P _HF ratio), the low back pain dysfunction index (ODI), the degree of smoking, and the result of the 12th node of the flatten layer Average, average of the results of the 1st node of the flatten layer, standard deviation of the results of the 7th node of the flatten layer, and the low-frequency to high-frequency band intensity ratio (P _LF / P _HF ratio) The standard deviation, the standard deviation of the STTc segment, the average of the low frequency band strength (P _LF ), and the average of the results of the third node of the flatten layer include 11 ECG features.

The input vector of the second servo vector machine is the average of the results of the 4th node of the flatten layer, the HDL cholesterol index, the average of the low frequency band strength (P _LF ), and the 10 of the flatten layer. The average of the result of the th node, the standard deviation of the low-frequency to high-frequency band intensity ratio (P _LF / P _HF ratio), the standard deviation of the result of the 12th node of the flatten layer, and the 6 electrocardiogram characteristics of the AHI index include

The third servo vector machine is the average of the results of the 10th node of the flatten layer, the average of the STTc segments, the average of the low frequency to high frequency band intensity ratio (P _LF / P _HF ratio), diastolic blood pressure, HDL cholesterol Index, mean of low frequency band intensity (P _LF ), and three of the BMI index.

The input vectors of the fourth servo vector machine are the HDL cholesterol index, the average of the results of the third node of the flatten layer, the standard deviation of the low frequency band intensity (P _LF ), the cholesterol index, the AHI index, and the platen Average of the results of the 2nd node of the flatten layer, standard deviation of systolic blood pressure, high frequency band intensity (P _HF ), average of the results of the 4th node of the flatten layer, smoking degree, platen layer It includes 7 of the standard deviations of the result of the 6th node of the (flatten layer).

In addition, the present invention is a method for driving a cardiovascular disease prediction system, including an arithmetic processing unit for predicting the occurrence of cardiovascular disease using an electrocardiogram signal and a CVD risk factor detected during polysomnography, The calculation processing unit detects a time domain electrocardiogram parameter including a ST-segment and T-wave (ST-T) changes (STTc) segment from the electrocardiogram (ECG) signal, and a low-frequency to high-frequency band intensity ratio (PLF/PHF ratio) ), detecting the frequency domain ECG parameters including ), and storing the mean and standard deviation of the detected time domain ECG parameters and the frequency domain ECG parameters in a memory unit as 'signal processed ECG features', signal-processed ECG feature extraction step; The calculation processing unit inputs the electrocardiogram signal for 30 seconds to the pre-learned CNN-based artificial intelligence model, and averages and standardizes the results of each node of the flatten layer in the CNN-based artificial intelligence model. an artificial intelligence-based electrocardiogram feature extraction step of obtaining a deviation and storing it in a memory unit as an 'artificial intelligence-based electrocardiogram feature'; The calculation processing unit may include a CVD risk factor loading step of reading the CVD risk factors stored in the memory unit; The calculation processing unit uses a cardiovascular disease (CVD) prediction model including servo vector machines (SVMs), but obtains the signal-processed ECG features obtained in the electrocardiogram feature extraction step and the artificial intelligence-based electrocardiogram feature extraction step The true AI-based electrocardiogram features and the electrocardiogram features including the CVD risk factors read in the CVD risk factor loading step are input into the cardiovascular disease prediction model, and the output results from the cardiovascular disease prediction model are used to determine future cardiovascular disease occurrence. Cardiovascular disease (which is output as a cardiovascular disease prediction result for whether or not cardiovascular disease occurs and which disease among coronary heart disease (CHD), heart failure (HF), and stroke will occur if cardiovascular disease occurs) It is characterized by including; CVD) prediction step.

In the cardiovascular disease (CVD) predicting step, the calculation processing unit inputs electrocardiogram features including the average of the STTc segments as an input vector to the pre-learned first servo vector machine, and the output value of the first servo vector machine, It outputs a value that determines whether cardiovascular disease will occur in the future, but if the output value of the first servo vector machine determines that cardiovascular disease does not occur, the value indicating that cardiovascular disease does not occur is the cardiovascular disease (CVD) prediction value. Outputting as, cardiovascular disease occurrence determination step; When the calculation processing unit determines that a cardiovascular disease will occur in the future in the cardiovascular disease occurrence determination unit, coronary heart disease (corona It includes a detailed disease determination step of selecting one of heart disease (CHD), heart failure (HF), and stroke and outputting it as a cardiovascular disease (CVD) predictive value.

In the detailed disease determination step, the operation processing unit inputs the electrocardiogram characteristics including the average of the low frequency band intensities (P _LF ) to the previously learned second servo vector machine, and from the second servo vector machine, the second a second servo vector machine calculation step of outputting one of coronary heart disease and heart failure as a result of the servo vector machine; The operation processing unit inputs the electrocardiogram characteristics including the average of the low frequency band intensities (P _LF ) to the previously learned third servo vector machine, and from the third servo vector machine, the result of the third servo vector machine a third servo vector machine calculation step, outputting one of coronary heart disease and stroke as a value; The operation processing unit inputs the electrocardiogram characteristics including the standard deviation of the low frequency band strength (P _LF ) to the previously learned fourth servo vector machine, and from the fourth servo vector machine, the fourth servo vector machine a fourth servo vector machine, outputting one of heart failure and stroke as a result value; The calculation processing unit counts the number of times each of coronary heart disease, heart failure, and stroke is selected from the result values of the second servo vector machine, the result values of the third servo vector machine, and the result values of the fourth servo vector machine. and a voting step of outputting counted values for each type of coronary heart disease, heart failure, and stroke.

In the voting step, the calculation processing unit outputs the disease value having the largest number of count values for coronary heart disease, heart failure, and stroke as a cardiovascular disease (CVD) prediction value.

In addition, the present invention is characterized by a recording medium storing a computer program source for a driving method of a cardiovascular disease prediction system.

The cardiovascular disease prediction system of the present invention applies the electrocardiogram and CVD risk factors of sleep breathing disorder patients to an artificial intelligence-based cardiovascular disease prediction model to predict future cardiovascular diseases of sleep breathing disorder patients.

Through this, it is possible to predict the occurrence of cardiovascular diseases, that is, coronary heart disease, heart failure, and stroke within the next 10 years using the electrocardiogram and CVD risk factors of patients with sleep breathing disorder. Personalized medical services can be made possible by predicting diseases that will accompany patients.

1 is a block diagram schematically illustrating the configuration of a cardiovascular disease prediction system using an artificial intelligence-based cardiovascular disease prediction model for sleep breathing disorder patients according to the present invention.
2 shows the cardiovascular disease (CVD) prediction model of the present invention.

Hereinafter, a cardiovascular disease prediction system using an artificial intelligence-based cardiovascular disease prediction model for sleep breathing disorder patients according to the present invention will be described in detail with reference to the accompanying drawings.

1 is a block diagram schematically illustrating the configuration of a cardiovascular disease prediction system using an artificial intelligence-based cardiovascular disease prediction model for sleep breathing disorder patients according to the present invention.

The data input unit 100 includes an electrocardiogram detector 110 and a cardiovascular disease risk factor data input unit 160 .

The electrocardiogram detector 110 includes an electrocardiogram sensor 120, a signal pre-processor 130, and an A/D converter 150.

The electrocardiogram sensor unit 120 detects an electrocardiogram signal of a patient with sleep breathing disorder, and the signal preprocessor 130 removes noise from the electrocardiogram signal detected by the electrocardiogram sensor unit 120 and performs signal preprocessing to amplify it, and A/ The D conversion unit 150 converts the electrocardiogram signal preprocessed by the signal preprocessing unit 130 into a digital signal.

The cardiovascular disease risk factor data input unit 160 is a means for inputting cardiovascular disease risk factor data.

The electrocardiogram detection unit 110 may be an electrocardiogram detection device and a cardiovascular disease risk factor data input device provided in the polysomnography device.

Here, the ECG signal may be an ECG signal detected during one night for polysomnography.

The calculation processing unit 200 stores the electrocardiogram data and cardiovascular disease risk factor data input from the data input unit 100 in the memory unit 310 . The calculation processing unit is composed of a computer.

Signal processed electrocardiogram (ECG) feature extraction unit 210, AI-based ECG features extraction unit 220, CVD risk factor loading unit 230, cardiovascular disease (CVD) prediction unit ( 250).

The signal-processed electrocardiogram (ECG) feature extraction unit 210 obtains QTc, STTc segment, SDNN, RMSSD, very low frequency band strength (PVLF), low frequency band strength (PLF), and high frequency band strength (PHF), and obtains a low frequency band high frequency The mean and standard deviation of each of the PLF/PHF ratios are obtained and stored in the memory unit 310 as signal processing-based ECG features (referred to as 'SP-based ECG features'). do.

The AI-based ECG features extraction unit 220 extracts the ECG signal for 30 seconds as an input signal, and the CNN-based AI model (ie, for extracting the AI-based ECG features) artificial intelligence model), and the CNN-based artificial intelligence model (ie, artificial intelligence model for extracting electrocardiogram features) obtains the result of each node of the flatten layer (smoothing layer), and the result The average and standard deviation of each of the values are obtained and stored in the memory unit 310 as AI-based ECG features.

The CVD risk factor loading unit 230 reads the CVD risk factors stored in the memory unit 310 .

The cardiovascular disease (CVD) prediction unit 250 is composed of a cardiovascular disease (CVD) prediction model.

The cardiovascular disease (CVD) prediction model, after going through the process of extracting features, can predict coronary heart disease (CHD), heart failure (heart failure) within 10 years through a support vector machine (SVM) model. Failure, HF), stroke occurrence or not is predicted, and the prediction result is output to the output unit 320 .

Hereinafter, driving of the arithmetic processing unit of the present invention will be described in detail.

In the present invention, future cardiovascular diseases are predicted using the electrocardiogram and CVD risk factors detected during polysomnography.

To develop a predictive model for cardiovascular disease, signal processing-based electrocardiogram features, artificial intelligence-based electrocardiogram features, and CVD risk factors were extracted.

During the experiment, polysomnography data of a baseline study were used as the electrocardiogram signal, and electrocardiograms during sleep from sleep onset to sleep end were analyzed.

After removing baseline fluctuation and power supply noise using a 1.5-20 Hz bandpass filter, the QRS complex and T wave were detected using an adaptive threshold algorithm and a morphological method.

In general, the QRS complex (i.e., R wave) in an electrocardiogram signal detects the maximum value during a time interval that becomes smaller than a predetermined threshold value (adaptive threshold value) from the time point exceeding a predetermined threshold value (adaptive threshold value) after the previous R peak. Peak (R point, R point), the minimum value immediately before the R peak is detected as the Q point (Q point), and the minimum value immediately after the R peak is detected as the S point (S point). In addition, the peak following the S point is detected as the T point (T point). A method for detecting the QRS complex and the T wave in the electrocardiogram signal is widely known, so a detailed description thereof will be omitted.

After detecting the feature points of P, Q, R, S, and T of the ECG (electrocardiogram), the ECG feature parameters in the time domain, QTc (corrected QT interval) and STTc segment (ST-segment and T-wave (ST-T) changes, ST segment and T-wave changes).

QTc is obtained by Equation 1.

Here, Ti, Qi, and Ri represent the starting points of the i-th (i.e., i-th period) T point (T wave), Q point (Q wave), and R point, respectively, and Ri+1 is the i+1th R point. Indicates the starting point.

The STTc segment is obtained by Equation 2.

Here, point J refers to the point where the QRS complex meets the ST segment.

The RR interval is expressed as Equation 3.

In this case, Ri-1 is the time point R of the i-1 (i.e., i-1 cycle) electrocardiogram, and fs is the sampling frequency of the electrocardiogram signal.

In order to calculate heart rate variability, the ectopic bits of RR were removed, and this signal was defined as NN (normal-to normal RR).

That is, the time interval between the R peak (ie, normal R peak) and the previous consecutive R peak (ie, consecutive previous normal R peak) is calculated, and this is taken as the NN interval.

In addition, as an ECG characteristic parameter in the time domain, the square root of the mean (RMSSD) of the square of the difference between the standard deviation (SDNN) of the entire RR intervals and the adjacent RR intervals may be further obtained.

The standard deviation (SDNN) of the entire RR interval is obtained as in Equation 4.

Here, N is the total heart rate, meanRRI refers to the average RR interval, and I(i) refers to the i-th RR interval.

The square root of the mean of the squared difference between adjacent RR intervals (RMSSD) is obtained as shown in Equation 5.

Here, I(i+1) refers to the i+1th RR interval.

For heart rate variability analysis, the calculated NN was interpolated at equal intervals and then resampled at 4 Hz. That is, when the number of samples in each RR interval is less than a predetermined number, interpolation is performed by filling in zeros, and the interpolated electrocardiogram is resampled (re-sampled) at 4 Hz.

After performing a fast Fourier transform (FFT) on the resampled signal in units of 30 seconds, a power spectrum density (PSD) was calculated by taking the square of the FFT. That is, for frequency analysis, fast Fourier transform (FFT) is performed on the interpolated and resampled ECG.

Each frequency band used to calculate the frequency domain features is very low frequency (VLF: 0 to 0.04 Hz), low frequency (LF: 0.04 to 0.15 Hz), and high frequency (HF: 0.15 to 0.15 Hz). 0.4 Hz).

In order to detect the electrocardiogram characteristic parameters in the frequency domain, the very low frequency band strength (P _VLF ), the low frequency band strength (P _LF ), and the high frequency band strength (P _HF ), which are the electrocardiographic characteristic parameters in the frequency domain, are obtained, and the low frequency band high frequency band strength Detect the ratio (P _LF /P _HF ratio). The method of obtaining the frequency band intensity (P _LF ), the high frequency band intensity (P _HF ), and the low frequency band high frequency band intensity ratio (P _LF / P _HF ratio) is widely known in Korean Patent No. 10-0493714, etc., so detailed description is omitted.

ECG characteristic parameters in the time domain of QTc, STTc segments, SDNN, and RMSSD, very low frequency band intensity (P _VLF ), low frequency band intensity (P _LF ), and high frequency band intensity (P _HF ) are obtained, and the low-frequency band intensity ratio If the average and standard deviation of each of the electrocardiogram feature parameters in the frequency domain of (P _LF / P _HF ratio) is obtained, 18 electrocardiogram feature parameters (for convenience of explanation, 'signal processing-based ECG features, SP -based ECG features)') is detected.

Table 1 shows the signal-processed ECG characteristics used as inputs of the cardiovascular disease (CVD) predictive model of the present invention.

That is, signal processing-based ECG features are extracted, and each feature (i.e., QTc, STTc segment, SDNN, RMSSD, P _VLF , P _LF , P _HF , P _LF / P _HF ratio) to obtain the average and standard deviation during the entire sleep, and store them in the memory as 'signal processed ECG features (SP-based ECG features)'. This process may be referred to as a signal-processed ECG feature extraction step.

Next, we designed a convolutional neural network (CNN)-based artificial intelligence model (i.e., an artificial intelligence model for extracting ECG features) to extract AI-based ECG features. The model receives electrocardiogram signals for 30 seconds and undergoes batch normalization. Afterwards, features are extracted through a convolution layer and a pooling layer consisting of a three-layer structure. The AI-based features extracted from the nodes of the flatten layer (smoothing layer) of the CNN model learned in this way were calculated for the average and standard deviation for each entire night's sleep time.

That is, the ECG signal for 30 seconds is input to the CNN-based artificial intelligence model (ie, the artificial intelligence-based artificial intelligence model for extracting the ECG features), and the CNN-based artificial intelligence model (ie, the artificial intelligence model for extracting the ECG features) , AI model for extracting electrocardiogram features) outputs the extraction result of each node of the flatten layer (smoothing layer). In addition, the average and standard deviation of each of these extraction results (15 pieces of data) are obtained, and the obtained data is stored in a memory unit as 'AI-based ECG features'. This process can be referred to as an AI-based electrocardiogram feature extraction step. Here, the data part of 'AI-based ECG features' is a total of 30 pieces of data.

Table 2 shows the characteristics of the artificial intelligence-based electrocardiogram used as an input for the cardiovascular disease (CVD) prediction model of the present invention.

In the present invention, in addition to the ECG features of signal-processed ECG features (SP-based ECG features) and AI-based ECG features, additionally representative cardiovascular disease (CVD) risk factors, CVD predictive models use as input

As the clinical CVD risk factors used in the present invention, the degree of smoking, the degree of hypertension (systolic blood pressure, diastolic blood pressure), the degree of cholesterol (cholesterol index, HDL cholesterol index), the degree of obesity (BMI index), age, A total of 10 data are used for gender, degree of sleep breathing disorder (SDB) (AHI index, sleep apnea index), and low back pain dysfunction index (ODI).

Table 3 shows the CVD risk factors used as inputs to the cardiovascular disease (CVD) prediction model of the present invention.

That is, the calculation processing unit reads the CVD risk factors obtained during polysomnography and stored in advance in the memory unit. This process can be referred to as the CVD risk factor stage.

In this way, the inputs used in the cardiovascular disease (CVD) prediction model in the present invention are 18 signal-processed electrocardiogram features, 30 artificial intelligence-based electrocardiogram features, and 10 CVD risk factors, for a total of 58 input feature data have

Next, these feature data were analyzed through statistical analysis, that is, two independent sample t-tests and chi-square tests, where each feature between classes had a p-value <0.05. , select feature data to be used as significantly different input data. This process can be referred to as the feature selection step.

In some cases, the feature selection step may be selected in the experiment step and determined at the time of product shipment.

In the present invention, after a process for extracting features, coronary heart disease (CHD), Predict the occurrence of heart failure (HF) and stroke.

2 shows the cardiovascular disease (CVD) prediction model of the present invention.

Prior to describing the cardiovascular disease (CVD) prediction model, electrocardiogram feature data applied as input to each servo vector machine in the cardiovascular disease (CVD) prediction model of FIG. 2 are shown.

The cardiovascular disease occurrence determination unit 710 is composed of a first servo vector machine (SVM_CVD), and inputs selected ECG characteristic data or given ECG characteristic data in the feature selection step, and determines whether or not cardiovascular disease will occur in the future ( CVD), that is, whether or not (CVD-free) is determined. If the cardiovascular disease occurrence determination unit 710 determines that cardiovascular disease does not occur, it outputs a value indicating that cardiovascular disease does not occur (ie, a CVD-free value), which is the final value of the cardiovascular disease (CVD) prediction model. output as a result.

The input vectors of the first servo vector machine (SVM_CVD) are the average of the results of the 15th node of the flatten layer, the average of the STTc segments, the BMI index, the systolic blood pressure, the AHI index, and the flatten layer (flatten layer). ), the standard deviation of the result of the 11th node, the average of the low-frequency to high-frequency band intensity ratio (P _LF / P _HF ratio), the low back pain dysfunction index (ODI), the degree of smoking, the 12th node of the flatten layer The average of the results of , the average of the results of the 1st node of the flatten layer, the standard deviation of the results of the 7th node of the flatten layer, the low-frequency to high-frequency band intensity ratio (P _LF / P _HF ratio), the standard deviation of STTc segments, the average of low frequency band strength (P _LF ), and the average of the results of the third node of the flatten layer.

The mean and standard deviation of these data are the mean and standard deviation for each total night's sleep.

In general, a support vector machine (SVM) aims to maximize its generalizing capability. Using the representative feature vectors of each class, we retrieve the optimized decision hyperplane with the maximum margin between categories. The servo vector machine is a well-known classifier, and a detailed description thereof will be omitted.

When it is determined that future cardiovascular disease will occur (CVD) in the cardiovascular disease occurrence determination unit 710, the detailed disease determination unit 720 uses a one-to-one (OvO, One-Vs-One) multi-class classifier , Select one of coronary heart disease (CHD), heart failure (HF), and stroke as the cardiovascular disease that will occur in the future and output it. The detailed disease determination unit 720 includes a second servo vector machine (SVM_C-H) 721, a third servo vector machine (SVM_C-S) 722, and a fourth servo vector machine (SVM_H-S). (723) consists of three servo vector machines.

The second servo vector machine (SVM_C-H) 721 is a means for determining one of coronary heart disease (CHD) and heart failure (HF), and is CHD), heart failure (heart failure, HF) is selected (candidate cardiovascular disease (CVD) prediction result) is output.

The input vector of the second servo vector machine (SVM_C-H) 721 is the average of the results of the 4th node of the flatten layer, the HDL cholesterol index, the average of the low frequency band strength (P _LF ), and the flatten layer. Average of the result of the 10th node of the flatten layer, standard deviation of the low frequency to high frequency band intensity ratio (P _LF / P _HF ratio), standard deviation of the result of the 12th node of the flatten layer, AHI index may be included.

The third servo vector machine (SVM_C-S) 722 is a means for discriminating between coronary heart disease (CHD) and stroke, and determines whether coronary heart disease (CHD) and stroke The result of selecting one of the strokes (candidate cardiovascular disease (CVD) prediction result) is output.

The input vector of the third servo vector machine (SVM_C-S) 722 is the average of the results of the 10th node of the flatten layer, the average of the STTc segments, and the low-frequency to high-frequency band intensity ratio (P _LF / P _HF ratio) mean, diastolic blood pressure, HDL cholesterol index, low frequency band intensity (P _LF ) mean, BMI index.

The fourth servo vector machine (SVM_H-S) 723 is a means for determining one of heart failure (HF) and stroke, and one of heart failure (HF) and stroke Outputs the selected result (candidate for cardiovascular disease (CVD) prediction result).

The input vectors of the fourth servo vector machine (SVM_H-S) 723 are the HDL cholesterol index, the average of the results of the third node of the flatten layer, the standard deviation of the low frequency band intensity (P _LF ), Cholesterol index, AHI index, average of the results of the second node of the flatten layer, systolic blood pressure, standard deviation of the high frequency band intensity (P _HF ), and the result of the fourth node of the flatten layer It may include the standard deviation of the average, smoking degree, and the result of the 6th node of the flatten layer.

With the voting unit 770, the result of the second servo vector machine (SVM_C-H) 721, the result of the third servo vector machine (SVM_C-S) 722, and the fourth servo vector machine (SVM_H- From the results of S) 723, the number of selected times of coronary heart disease (CHD), heart failure (HF), and stroke are counted and output. And the disease with the highest number of selections is output as the final result as a disease that can occur within the next 10 years.

In other words, the cardiovascular disease (CVD) prediction model, after going through the process of extracting features, coronary heart disease (CHD), Predict the occurrence of heart failure (HF) and stroke. To this end, first go through SVM (SVM_CVD), which predicts CVD and CVD-free, and then, when classified as CVD, go through a one-to-one (OvO, One-Vs-One) multi-class classifier to classify CHD, HF, and stroke. OvO is a method of selecting a class with the most discriminant values through K(K-1)/2 binary class classification by selecting two class combinations among K target classes, a well-known method. As such, a detailed description is produced.

In this specification, the detailed description of the contents that can be sufficiently recognized and inferred by those skilled in the art of the present invention is omitted, and the technical spirit or essential configuration of the present invention is changed in addition to the specific examples described in this specification. More diverse modifications are possible within the range not specified. Accordingly, the present invention may be practiced in a manner different from that specifically described and exemplified herein, which can be understood by those skilled in the art.

100: data input unit 110: electrocardiogram detection unit
120: electrocardiogram sensor unit 130: signal pre-processing unit
150: A / D conversion unit 160: cardiovascular disease risk factor data input unit
210: Signal processed electrocardiogram feature extraction unit
220: AI-based electrocardiogram feature extraction unit
230: CVD risk factor loading unit 250: cardiovascular disease prediction unit
310: memory unit 320: output unit

Claims (19)

A cardiovascular disease prediction system comprising an arithmetic processing unit for predicting the occurrence of cardiovascular disease using an electrocardiogram signal and a CVD risk factor detected during polysomnography,
calculation processing unit,
From the electrocardiogram (ECG) signal detected by the electrocardiogram detector during polysomnography, time domain electrocardiogram parameters including STTc (ST-segment and T-wave (ST-T) changes) segments are detected, and low-frequency and high-frequency band intensities Detecting frequency domain electrocardiogram parameters including the ratio (PLF/PHF ratio), and storing the mean and standard deviation of the detected time domain electrocardiogram parameters and frequency domain electrocardiogram parameters as 'signal processed electrocardiogram features' in a memory unit, a signal-processed electrocardiogram feature extraction unit;
An electrocardiogram signal for 30 seconds is input to the pre-learned CNN-based artificial intelligence model, and the mean and standard deviation of the results of each node of the flatten layer in the CNN-based artificial intelligence model are obtained, an artificial intelligence-based electrocardiogram feature extraction unit to store in a memory unit as 'artificial intelligence-based electrocardiogram feature';
a CVD risk factor loading unit that reads the CVD risk factors stored in the memory unit;
Including cardiovascular disease (CVD) predictive models including servo vector machines (SVMs), including signal-processed ECG features, AI-based ECG features and CVD risk factors, to cardiovascular disease predictive models input, and from the cardiovascular disease prediction model, whether or not cardiovascular disease will occur in the future, and if a cardiovascular disease occurs, which disease among coronary heart disease (CHD), heart failure (HF), and stroke will occur a cardiovascular disease (CVD) prediction unit that outputs the result as a cardiovascular disease prediction result;
Characterized in that it comprises a cardiovascular disease prediction system. The method of claim 1, wherein the cardiovascular disease (CVD) prediction unit,
It is composed of the pre-learned first servo vector machine to determine whether cardiovascular disease will occur in the future or not, and if it is determined that cardiovascular disease does not occur, the value indicating that cardiovascular disease does not occur is the cardiovascular disease (CVD) prediction value To output, cardiovascular disease occurrence determination unit;
If it is determined that future cardiovascular disease will occur in the cardiovascular disease occurrence determination unit, coronary heart disease (CHD) is classified as future cardiovascular disease using a one-to-one (OvO, One-Vs-One) multi-class classifier. ), heart failure (heart failure, HF), stroke (stroke) to select one of the cardiovascular disease (CVD) predictive value, a detailed disease determination unit for outputting;
Characterized in that it comprises a cardiovascular disease prediction system. The method of claim 2, wherein the detailed disease determination unit
a second servo vector machine configured to select one of coronary heart disease and heart failure as a result of the second servo vector machine;
A third servo vector that is pre-learned to discriminate between coronary heart disease and stroke, and selects one of coronary heart disease and stroke and outputs it as a result of the third servo vector machine. machine;
a fourth servo vector machine configured to select one of heart failure and stroke as a previously learned servo vector machine to discriminate between heart failure and stroke and output the result as a result of the fourth servo vector machine;
Characterized in that it comprises a cardiovascular disease prediction system. The method of claim 3, wherein the detailed disease discrimination unit
a voting unit which counts and outputs the number of times each of coronary heart disease, heart failure, and stroke is selected from the result of the second servo vector machine, the result of the third servo vector machine, and the result of the fourth servo vector machine;
Characterized in that it further comprises, cardiovascular disease prediction system. According to claim 4,
A cardiovascular disease prediction system, characterized in that the voting unit outputs values for coronary heart disease, heart failure, and diseases with the highest number of strokes selected as cardiovascular disease (CVD) prediction values. According to claim 5,
The input vector of the first servo vector machine is the average of the result of the 15th node of the flatten layer, the average of the STTc segments, the BMI index, the systolic blood pressure, the AHI index, and the 11 of the flatten layer. The standard deviation of the result of the th node, the average of the low-frequency to high-frequency band intensity ratio (P _LF / P _HF ratio), the low back pain dysfunction index (ODI), the degree of smoking, and the result of the 12th node of the flatten layer Average, average of the results of the 1st node of the flatten layer, standard deviation of the results of the 7th node of the flatten layer, and the low-frequency to high-frequency band intensity ratio (P _LF / P _HF ratio) Cardiovascular disease prediction, characterized by including 11 electrocardiogram features among the standard deviation, standard deviation of the STTc segment, average of low frequency band intensity (P _LF ), and average of the results of the 3rd node of the flatten layer system. According to claim 5,
The input vector of the second servo vector machine is the average of the results of the 4th node of the flatten layer, the HDL cholesterol index, the average of the low frequency band strength (P _LF ), and the 10 of the flatten layer. The average of the result of the th node, the standard deviation of the low-frequency to high-frequency band intensity ratio (P _LF / P _HF ratio), the standard deviation of the result of the 12th node of the flatten layer, and the 6 electrocardiogram characteristics of the AHI index Characterized in that it comprises, cardiovascular disease prediction system. According to claim 5,
The third servo vector machine is the average of the results of the 10th node of the flatten layer, the average of the STTc segments, the average of the low frequency to high frequency band intensity ratio (P _LF / P _HF ratio), diastolic blood pressure, HDL cholesterol Characterized in that it comprises three of the index, the average of the low frequency band intensity (P _LF ), and the BMI index, cardiovascular disease prediction system. According to claim 6,
The input vectors of the fourth servo vector machine are the HDL cholesterol index, the average of the results of the third node of the flatten layer, the standard deviation of the low frequency band intensity (P _LF ), the cholesterol index, the AHI index, and the platen Average of the results of the 2nd node of the flatten layer, standard deviation of systolic blood pressure, high frequency band intensity (P _HF ), average of the results of the 4th node of the flatten layer, smoking degree, platen layer Characterized in that it includes 7 of the standard deviations of the results of the 6th node of the (flatten layer), cardiovascular disease prediction system. A method for driving a cardiovascular disease prediction system, including an arithmetic processing unit for predicting the occurrence of cardiovascular disease using an electrocardiogram signal and a CVD risk factor detected during polysomnography,
The calculation processing unit detects time domain electrocardiogram parameters including ST-segment and T-wave (ST-T) changes (STTc) segments from the electrocardiogram (ECG) signal detected by the electrocardiogram detection unit during polysomnography, and Detect frequency domain electrocardiogram parameters including the frequency-to-high frequency band intensity ratio (PLF/PHF ratio), and store the mean and standard deviation of the detected time domain electrocardiogram parameters and frequency domain electrocardiogram parameters as 'signal processed electrocardiogram features' Storing in, signal-processed electrocardiogram feature extraction step;
The calculation processing unit inputs the electrocardiogram signal for 30 seconds to the pre-learned CNN-based artificial intelligence model, and averages and standardizes the results of each node of the flatten layer in the CNN-based artificial intelligence model. an artificial intelligence-based electrocardiogram feature extraction step of obtaining a deviation and storing it in a memory unit as an 'artificial intelligence-based electrocardiogram feature';
The calculation processing unit may include a CVD risk factor loading step of reading the CVD risk factors stored in the memory unit;
The calculation processing unit uses a cardiovascular disease (CVD) prediction model including servo vector machines (SVMs), but obtains the signal-processed ECG features obtained in the electrocardiogram feature extraction step and the artificial intelligence-based electrocardiogram feature extraction step The true AI-based electrocardiogram features and the electrocardiogram features including the CVD risk factors read in the CVD risk factor loading step are input into the cardiovascular disease prediction model, and the output results from the cardiovascular disease prediction model are used to determine future cardiovascular disease occurrence. Cardiovascular disease (which is output as a cardiovascular disease prediction result for whether or not cardiovascular disease occurs and which disease among coronary heart disease (CHD), heart failure (HF), and stroke will occur if cardiovascular disease occurs) CVD) prediction step;
Characterized in that it comprises a, driving method of the cardiovascular disease prediction system. The method of claim 10, wherein the cardiovascular disease (CVD) prediction step,
The calculation processing unit inputs electrocardiogram features including the average of the STTc segments as an input vector to the pre-learned first servo vector machine, and uses the output value of the first servo vector machine to determine whether cardiovascular disease will occur in the future. Cardiovascular disease occurrence determination step of outputting a value indicating that cardiovascular disease does not occur as a cardiovascular disease (CVD) predictive value if the output value of the first servo vector machine determines that cardiovascular disease does not occur ;
When the calculation processing unit determines that a cardiovascular disease will occur in the future in the cardiovascular disease occurrence determination unit, coronary heart disease (corona A detailed disease determination step of selecting one of heart disease (CHD), heart failure (HF), and stroke and outputting it as a cardiovascular disease (CVD) predictive value;
Characterized in that it comprises a, driving method of the cardiovascular disease prediction system. The method of claim 11, wherein the detailed disease determination step
The operation processing unit inputs the electrocardiogram characteristics including the average of the low frequency band strengths (P _LF ) to the previously learned second servo vector machine, and from the second servo vector machine, the result of the second servo vector machine a second servo vector machine calculation step, outputting one of coronary heart disease and heart failure as a value;
The operation processing unit inputs the electrocardiogram characteristics including the average of the low frequency band intensities (P _LF ) to the previously learned third servo vector machine, and from the third servo vector machine, the result of the third servo vector machine a third servo vector machine calculation step, outputting one of coronary heart disease and stroke as a value;
The operation processing unit inputs the electrocardiogram characteristics including the standard deviation of the low frequency band strength (P _LF ) to the previously learned fourth servo vector machine, and from the fourth servo vector machine, the fourth servo vector machine a fourth servo vector machine, outputting one of heart failure and stroke as a result value;
Characterized in that it comprises a, driving method of the cardiovascular disease prediction system. The method of claim 12, wherein the detailed disease determination step
The calculation processing unit counts the number of times each of coronary heart disease, heart failure, and stroke is selected from the result values of the second servo vector machine, the result values of the third servo vector machine, and the result values of the fourth servo vector machine. a voting step of outputting counted values for each type of coronary heart disease, heart failure, and stroke;
Characterized in that it further comprises, the driving method of the cardiovascular disease prediction system. According to claim 13,
A method of driving a cardiovascular disease prediction system, characterized in that in the voting step, the calculation processing unit outputs, as a cardiovascular disease (CVD) prediction value, a disease value with the highest count value for coronary heart disease, heart failure, and stroke. According to claim 13,
The input vector of the first servo vector machine is the average of the result of the 15th node of the flatten layer, the average of the STTc segments, the BMI index, the systolic blood pressure, the AHI index, and the 11 of the flatten layer. The standard deviation of the result of the th node, the average of the low-frequency to high-frequency band intensity ratio (P _LF / P _HF ratio), the low back pain dysfunction index (ODI), the degree of smoking, and the result of the 12th node of the flatten layer Average, average of the results of the 1st node of the flatten layer, standard deviation of the results of the 7th node of the flatten layer, and the low-frequency to high-frequency band intensity ratio (P _LF / P _HF ratio) Cardiovascular disease prediction, characterized by including 11 electrocardiogram features among the standard deviation, standard deviation of the STTc segment, average of low frequency band intensity (P _LF ), and average of the results of the 3rd node of the flatten layer How the system works. According to claim 13,
The input vector of the second servo vector machine is the average of the results of the 4th node of the flatten layer, the HDL cholesterol index, the average of the low frequency band strength (P _LF ), and the 10 of the flatten layer. The average of the result of the th node, the standard deviation of the low-frequency to high-frequency band intensity ratio (P _LF / P _HF ratio), the standard deviation of the result of the 12th node of the flatten layer, and the 6 electrocardiogram characteristics of the AHI index Characterized in that it comprises, the driving method of the cardiovascular disease prediction system. According to claim 13,
The third servo vector machine is the average of the results of the 10th node of the flatten layer, the average of the STTc segments, the average of the low frequency to high frequency band intensity ratio (P _LF / P _HF ratio), diastolic blood pressure, HDL cholesterol An index, a method for driving a cardiovascular disease prediction system, characterized in that it includes three of the average of the low frequency band intensity (P _LF ) and the BMI index. According to claim 13,
The input vectors of the fourth servo vector machine are the HDL cholesterol index, the average of the results of the third node of the flatten layer, the standard deviation of the low frequency band intensity (P _LF ), the cholesterol index, the AHI index, and the platen Average of the results of the 2nd node of the flatten layer, standard deviation of systolic blood pressure, high frequency band intensity (P _HF ), average of the results of the 4th node of the flatten layer, smoking degree, platen layer Characterized in that it includes 7 of the standard deviations of the result of the 6th node of the (flatten layer), a method of driving a cardiovascular disease prediction system. A recording medium storing a computer program source for a driving method of the cardiovascular disease prediction system according to any one of claims 10 to 18.