A pulse wave-based coronary artery disease risk prediction system based on deep features

The deep feature-based pulse wave coronary heart disease risk prediction system utilizes a contact pulse acquisition device and a deep convolutional neural network combined with traditional machine learning to achieve non-invasive, simple, and rapid screening and prediction of coronary heart disease. It solves the problems of high cost and low accuracy in existing technologies, provides fine-grained risk assessment, and reduces hardware costs and power consumption.

CN123058698APending Publication Date: 2026-11-10TIANZE ZHIHUI TECHNOLOGY (TIANJIN) CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611260817.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-19
Publication Date
2026-11-10

Smart Images

  • Figure CN123058698A_ABST
    Figure CN123058698A_ABST
Patent Text Reader

Abstract

This invention relates to a pulse wave-based coronary heart disease risk prediction system based on deep features, comprising: a contact pulse wave acquisition device for acquiring pulse wave photoplethysmography (PPG) of a target; a preprocessing module for processing the PPG to obtain a one-dimensional representation time-series matrix; a mapping and transformation module for mapping the one-dimensional representation time-series matrix to a two-dimensional representation time-frequency matrix according to a pre-constructed two-dimensional time-frequency mapping relationship; a feature extraction module for extracting coronary heart disease features from the two-dimensional representation time-frequency matrix according to a pre-set two-dimensional convolutional neural network fused with CAPM; and a model analysis module configured with a disease analysis model, which generates corresponding diagnostic results based on the coronary heart disease features. The contact pulse wave PPG is used for coronary heart disease risk prediction, converting one-dimensional PPG data into a two-dimensional time-frequency matrix, and integrating deep learning and machine learning to improve prediction accuracy. By advancing the prevention timeline for coronary heart disease, a simple, rapid, economical, and widespread coronary heart disease screening and early warning system can be achieved, significantly increasing the coverage of coronary heart disease screening and protecting human health.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of coronary heart disease prediction technology, and more specifically, to a pulse wave-based coronary heart disease risk prediction system based on deep features. Background Technology

[0002] Coronary heart disease is characterized by "clear precipitating factors, typical symptoms, high-risk population concentration, and the ability to be diagnosed and controlled but requiring long-term management." The key lies in "early detection and early intervention" to avoid delaying treatment until myocardial infarction. Photoplethysmography is a technique based on photoelectric conversion to measure changes in blood volume.

[0003] Currently, the diagnosis and prediction of coronary heart disease mainly rely on clinical testing methods: electrocardiogram (ECG), blood tests, coronary CTA, or coronary angiography, etc. Computer-aided diagnosis and prediction algorithms for coronary heart disease involve further computer analysis of clinical test images. Their main drawbacks include high testing costs, cumbersome procedures, limited prediction accuracy using a single algorithm, and difficulty in simple, rapid, and large-scale early screening, diagnosis, and prognostic assessment and follow-up. Cardiovascular detection technology using PPG signals is non-invasive, portable, and low-cost, but current technologies have shortcomings. First, feature extraction methods are simplistic and rely on prior human knowledge. Only one-dimensional time-domain processing of PPG signals is performed, or further extraction of manually designed statistical features such as heart rate variability (HRV) and power spectral density is performed. These features are highly dependent on the prior knowledge of clinicians, have long development cycles, and are difficult to comprehensively capture nonlinear, high-dimensional, and hidden pathological details such as peak notching and dicrotic wave disappearance; essentially, it is a "shallow" signal analysis method. Furthermore, these methods typically use a single regression model (such as random forest regression), and their prediction accuracy is limited by the quality and completeness of the manually crafted features. Secondly, some multi-sensor fusion solutions increase system complexity and cost. They rely on ECG and PPG sensors simultaneously, increasing hardware costs and power consumption, requiring a high level of expertise, and limiting their widespread application in large-scale, everyday scenarios. Finally, the lack of an integrated, high-precision end-to-end prediction process, relying solely on binary risk assessment (disease-positive / disease-free), fails to provide fine-grained risk levels (low / medium / high); real-time prediction is difficult to achieve on low-power portable devices; and valuable frequency-domain phase information in PPG signals is lost, leading to insensitivity to early, minor hemodynamic abnormalities. In summary, one-dimensional time-domain processing of PPG easily loses key pathological details such as peak notching and dicrotic wave disappearance; using a single deep learning model or traditional machine learning model results in insufficient prediction accuracy and robustness; and there is a lack of integrated systems with high-precision acquisition, anti-interference denoising, time-frequency feature enhancement, and integrated acquisition and decision-making mechanisms.

[0004] To address the aforementioned issues, this invention proposes a method and system for predicting coronary heart disease risk based on deep features. The core of this method is to avoid manually designing features and instead utilize deep convolutional neural networks to automatically mine deep pathological features from two-dimensional time-frequency spectra. Furthermore, it constructs a heterogeneous serial architecture that combines a deep feature extractor with a traditional machine learning classifier to achieve simple, accurate, fast, and low-cost large-scale screening for coronary heart disease. Summary of the Invention

[0005] In view of this, the objective of this invention is to provide a pulse wave-based coronary heart disease risk prediction system based on deep features. To solve the above technical problems, the technical solution of this invention is: a pulse wave-based coronary heart disease risk prediction system based on deep features, comprising: A contact pulse acquisition device is used to acquire pulse wave photoplethysmography of a target. The preprocessing module is used to process the pulse wave photoplethysmogram to obtain a one-dimensional representation time series matrix; The mapping and transformation module maps the one-dimensional time-series matrix to a two-dimensional time-frequency matrix based on a pre-constructed two-dimensional time-frequency mapping relationship. The feature extraction module extracts coronary heart disease features from the two-dimensional time-frequency matrix representing the disease based on a preset two-dimensional convolutional neural network. The model analysis module is equipped with a disease analysis model, which is used to generate corresponding diagnostic results information based on the characteristics of coronary heart disease.

[0006] Furthermore, the preprocessing module is equipped with a bandpass filter processing unit, which is used to process noise data in the pulse wave photoplethysmography, and the bandpass filter processing unit is equipped with a Butterworth bandpass filter.

[0007] Furthermore, the mapping and transformation module is configured with a time-frequency transformation algorithm, which is configured to use a 3-level wavelet decomposition algorithm based on the sym2 wavelet basis.

[0008] Furthermore, the feature extraction layer of the two-dimensional convolutional neural network includes three execution sub-layers and one CAPM module. Each execution sub-layer includes a convolution module, a batch normalization module, a ReLU activation module, and a max pooling module. The CAPM module consists of a time-frequency morphological convolution branch and a time-frequency attention branch for cardiac cycle perception.

[0009] Furthermore: the convolution module is configured with an execution expression, which is: ; in, The input is a two-dimensional time-frequency feature map; Let be the convolution kernel weight matrix; (i, j) be the coordinates of the output feature map; (m, n) be the traversal index within the convolution kernel; Ii+m,j+n K represents the pixel value at the corresponding location in the input feature map. m,n The weights of the convolution kernel at the corresponding positions are represented by '*'; '*' indicates a two-dimensional convolution operation.

[0010] Furthermore, the time-frequency morphological convolution branch includes a main peak rounded convolution kernel along the time axis, an ascending branch rate convolution kernel, a diphtheria wave microstructure convolution kernel, and a frequency band energy attenuation convolution kernel.

[0011] Furthermore: the cardiac cycle perception time-frequency attention branch includes time gating and frequency attention; time gating generates time weights by global pooling along the frequency axis + temporal CNN + sigmoid; frequency attention generates frequency weights by global pooling along the time axis + two fully connected layers + sigmoid; the outer product of the two weights forms an attention mask to enhance the features of key time-frequency regions.

[0012] Furthermore, it also includes a model training module, which is used to train the disease analysis model. The model training module is configured with a training strategy, which includes a cross-validation sub-strategy.

[0013] Furthermore, the disease analysis model is equipped with a nonlinear random forest classifier.

[0014] The main technical effects of this invention are reflected in the following aspects: (1) By using contact pulse wave photoplethysmography, non-invasive, simple, and large-area early warning screening and prognostic assessment of coronary heart disease can be achieved; the acceptance of patients is improved, the burden of medical resources is reduced, and the disadvantages of high cost and untimely detection of traditional methods are avoided. Through the detection system proposed in this invention, a two-dimensional time-frequency matrix is ​​obtained by using the simple and easy-to-collect pulse wave, and a model combining a two-dimensional convolutional neural network with CAPM and a machine learning classifier is constructed, which improves the sensitivity and accuracy of early screening of coronary heart disease. When patients do not have obvious physical symptoms, the prediction of coronary heart disease is realized, thereby intervening to protect the life and health of patients. (2) In this technical solution, the one-dimensional time-series matrix of pulse wave is converted into a two-dimensional time-frequency matrix by using the time-frequency transformation method. The time-frequency transformation can extend the data of the original time-series pulse wave to the time-frequency domain, increase the feature information at different scales, and by fusing the CAPM two-dimensional convolutional neural network, the feature relationship between the time domain and the frequency domain can be focused on simultaneously, capturing the pathological characteristics of coronary heart disease on PPG in detail, thereby improving the accuracy of the neural network in extracting the features of coronary heart disease in pulse wave data. (3) By inputting the two-dimensional time-frequency matrix into a two-dimensional convolutional neural network fused with CAPM for feature extraction, and then inputting the features extracted by the neural network into the machine learning model, the prediction accuracy and efficiency of the model are further improved, and comprehensive prediction from different perspectives is achieved. Compared with a single one-dimensional time-domain matrix or a deep learning model, this technical solution greatly improves the accuracy and sensitivity of the model in predicting coronary heart disease. At the same time, the lightweight classifier can be adapted to embedded devices to achieve second-level prediction. Based on contact pulse wave photoplethysmography, the risk prediction of coronary heart disease is carried out by converting one-dimensional pulse wave photoplethysmography data into a two-dimensional time-frequency matrix and integrating deep learning and machine learning to improve prediction accuracy. By advancing the prevention time of coronary heart disease, simple, fast, economical and large-scale coronary heart disease screening and early warning can be achieved, greatly improving the popularization rate of coronary heart disease screening and protecting human life and health. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the steps of the coronary heart disease risk prediction method according to Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the pulse wave waveform; Figure 3 This is a schematic diagram of the algorithm for a high-efficiency classifier based on deep feature fusion according to Embodiment 1 of the present invention; Figure 4 This is a schematic diagram of the composition of the coronary heart disease risk prediction system according to Embodiment 2 of the present invention; Figure 5 This is a diagram comparing the waveforms of the healthy group and the coronary heart disease risk group. Figure 6 A schematic diagram comparing the time-frequency difference graphs of wavelet decomposition; Figure 7 This is a schematic diagram of a convolutional neural network structure that incorporates CAPM. Detailed Implementation

[0016] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings, so that the technical solution of the present invention can be more easily understood and mastered.

[0017] Reference Figure 1-4 As shown, a pulse wave-based coronary heart disease risk prediction system based on deep features includes: A contact pulse acquisition device is used to acquire pulse wave photoplethysmography of a target. The preprocessing module is used to process the pulse wave photoplethysmogram to obtain a one-dimensional representation time series matrix; The mapping and transformation module maps the one-dimensional time-series matrix to a two-dimensional time-frequency matrix based on a pre-constructed two-dimensional time-frequency mapping relationship. The feature extraction module extracts coronary heart disease features from the two-dimensional time-frequency matrix based on a preset two-dimensional convolutional neural network that integrates CAPM. The model analysis module is equipped with a disease analysis model, which is used to generate corresponding diagnostic results based on the characteristics of coronary heart disease. First, the pulse wave photoplethysmography (PPG) of the target population is preprocessed by bandpass filtering and baseline correction to obtain a one-dimensional time-series matrix. Contact PPG is obtained through fingertip measurement using a photoelectric sensor and a digitization module. The fingertip pulse wave, as a non-invasive representation of the hemodynamic state of the cardiovascular system, directly reflects the macroscopic functions of cardiac pumping and vascular elasticity in the time domain, while its frequency domain characteristics reveal the intrinsic frequency structure and micro-rhythmic patterns of the hemodynamic signal. The core pathological mechanism of coronary heart disease is coronary artery stenosis leading to myocardial ischemia and thus abnormal cardiac function. When coronary artery stenosis causes myocardial ischemia, myocardial contractility decreases, left ventricular ejection fraction decreases, resulting in a decrease in systolic peak amplitude and a prolonged rise time; aortic elastic reserve is depleted, and vascular sclerosis occurs, leading to a shallower or absent dicrotic wave and a decreased dicrotic ratio. Such waveform changes are pathologically specific, quantifiable, and repeatable, and are highly correlated with coronary heart disease risk levels. Therefore, this invention utilizes time-frequency transformation techniques (including but not limited to wavelet transform, Fourier transform, etc.) to extend a one-dimensional time-series matrix into a two-dimensional time-frequency matrix. This addresses the problem of traditional one-dimensional time-domain analysis severing the physical connection between the time and frequency domains. For example, the disappearance of the dicrotic wave manifests both as amplitude changes in the time domain and energy attenuation in specific frequency bands in the frequency domain. This invention maps a one-dimensional signal to a two-dimensional time-frequency plane through time-frequency transformation, generating a map containing a complete time-frequency structure. This map preserves key time-domain points such as contraction peaks and notches, while revealing the intrinsic micro-rhythms and energy distribution of hemodynamic signals through frequency domain components at different scales. This is a feature enhancement and information augmentation process, transforming hidden, nonlinear pathological patterns into image texture features that are easily captured and visualized by CNNs. Finally, the two-dimensional time-frequency matrix is ​​used as input to a risk prediction model to obtain the corresponding coronary heart disease risk assessment results. (Refer to...) Figure 5 As shown in the figure, this diagram illustrates the specific impact of coronary heart disease on pulse wave morphology through a visual comparison of one-dimensional waveforms between the healthy group and the coronary heart disease group (pathological group): Healthy group waveform: It exhibits a clear systolic peak (A), peak notch (B), and dicrotic wave (C), with a smooth waveform. Key points O, A, C and their projections O', A', C' are marked. Coronary heart disease risk group waveform: It displays non-linear, weak, and unsteady pathological changes, mainly characterized by: reduced systolic peak amplitude, slower rising limb rate, shallower or absent peak notch, shallower or absent dicrotic wave, and an overall blunted waveform. Dashed lines delineate these pathological details, highlighting the non-manual design features that the two-dimensional time-frequency conversion in this invention aims to capture and enhance.

[0018] The core technology utilizes a heterogeneous serial architecture for the deep information extraction module to achieve precise mapping from raw pulse wave data to coronary artery disease risk levels. The deep information extraction module consists of a two-dimensional convolutional neural network (CNN) and a CAPM module (the CNN includes, but is not limited to, VGG, ResNet, and their variants), with the two-dimensional time-frequency matrix as its input. The advantage of 2D CNNs over one-dimensional CNNs or RNNs is that the 2D convolutional kernels can simultaneously convolve the time and frequency axes, thereby capturing joint time-frequency features. A specific convolutional kernel can specifically respond to sudden drops in high-frequency energy during the systolic phase, offering greater pathological specificity compared to analyzing the time or frequency domains alone. This module is not limited to specific networks such as VGG and ResNet; it is specifically optimized and adapted to the characteristics of different networks. The VGG series uses a structure of "small convolutional kernels (3×3) + stacked multiple convolutional layers" to achieve fine-grained feature capture of subtle morphological distortions in pulse waves (such as systolic notches and abnormal dicrotic waves). Its fully connected layers can output high-dimensional dense features, retaining more pathological correlation information. The ResNet series solves the gradient vanishing problem in deep networks through residual connections, which can build deeper network structures and uncover more hidden early coronary heart disease features in pulse waves (such as waveform patterns corresponding to slight hemodynamic abnormalities). CAPM consists of two parallel branches: a time-frequency morphological convolution branch and a time-frequency attention branch for cardiac cycle perception. Finally, it is fused by 1×1 convolution and supplemented by residual connections to obtain a feature matrix that deeply contains details of coronary heart disease features. The common goal of this module is to automatically and end-to-end mine deep features strongly correlated with coronary artery disease (CAD) pathology, capturing subtle pathological features that are difficult to identify manually (such as irregular pulse waveforms caused by early myocardial ischemia and implicit signals of reduced vascular elasticity); automatically learn the association between features and CAD pathological mechanisms (such as feature maps corresponding to abnormal myocardial contractile function and feature patterns related to arteriosclerosis), forming a highly discriminative pathology-feature mapping relationship, avoiding the knowledge completeness problem of traditional methods relying on manually generated features, and providing high-quality feature input for subsequent classification and prediction. (Refer to...) Figure 7As shown in the figure, this diagram details the feature extraction layer structure of a convolutional neural network, focusing on the coronary artery feature perception module (CAPM): Input feature map: After convolution processing in the previous layer. Morphological convolution branch: Based on the time-frequency pathological characteristics of the pulse wave, four convolution kernels of different sizes and functions are preset: main peak rounded convolution kernel (k_main), rising branch rate convolution kernel (k_rate), diphtheria wave microstructure convolution kernel (k_micro), and frequency band energy decay convolution kernel (k_atten). Each convolution kernel captures different morphological details, and after cascading and ReLU activation, morphological enhancement features are obtained. Attention branch: Used for the network to self-perceive the systolic / diastolic phases of the cardiac cycle end-to-end: A time-gated vector G(t) is generated through global pooling and Temporal CNN (temporal convolutional network) to accurately locate key morphological points of the cardiac cycle; a frequency weight W(f) is generated through two fully connected layers and Sigmoid to form an attention mask, focusing the network's attention on key time-frequency regions such as the systolic peak and diphtheria wave, resulting in attention enhancement features. Fusion Output: After feature fusion from the two branches, high-fidelity, pathology-specific deep features are formed, improving the network's extraction and prediction sensitivity. The efficient classifier is composed of traditional machine learning models (including but not limited to support vector machines, random forests, logistic regression, etc.), whose input is the deep feature vector output by the deep information extraction module. This invention adopts a heterogeneous serial architecture of deep feature extraction + traditional machine learning classification. In the medical and health field, especially in early screening for coronary heart disease, obtaining labeled, high-quality samples is costly, and the sample size is usually limited. In small-sample scenarios, using fully connected layers + Softmax in deep classification networks (CNNs) for classification easily leads to overfitting, resulting in poor model generalization ability. In contrast, traditional machine learning models have lower model complexity, stronger small-sample learning ability, and better interpretability; their computational efficiency is far higher than that of deep classification networks, allowing the model to be easily deployed on wearable edge computing devices to achieve real-time prediction within seconds. The architecture proposed in this invention achieves complementary advantages between the powerful representation capabilities of deep learning and the robust generalization capabilities of traditional machine learning.

[0019] The hardware configuration is as follows: it includes a pulse wave data acquisition unit, a user processing unit, an early warning unit, and a cloud and visualization unit. The pulse wave data acquisition module is used to acquire the photoplethysmogram of the fingertips of the target group, and includes at least one photoelectric sensing element, a digitizing element, and a communication interface. The digitized pulse wave signal is transmitted to the user processing unit (including but not limited to a computer, server, or other equipment with data receiving and processing capabilities) through the communication interface. The user processing unit is used to input a two-dimensional time-frequency matrix into a risk prediction model to determine the coronary heart disease risk prediction result for the target group. The two-dimensional time-frequency matrix is ​​obtained through time-frequency transformation technology, and the risk prediction model is constructed by fusing a two-dimensional convolutional neural network with a machine learning classifier. The cloud and visualization unit includes user data uploading, generating corresponding test reports, risk warnings, and personalized health recommendations.

[0020] This embodiment provides a coronary heart disease risk prediction system based on pulse wave photoplethysmography (PPG). Its key feature is that it requires only a single PPG sensor for fingertip measurement, which is superior to traditional multi-sensor collaborative solutions.

[0021] This system is used to screen target groups for coronary heart disease risk using the coronary heart disease risk prediction method described in the above embodiments and obtain prediction results. The risk prediction system consists of four interconnected and collaborative modules, including a pulse wave photoplethysmography acquisition module, a data storage module, a preprocessing and coronary heart disease risk prediction module, and a visualization and result display module.

[0022] The pulse wave photoplethysmography (PPG) module is used to acquire fingertip pulse wave data of the target population. This module includes photoelectric sensing elements, digitizing elements (such as an ADC with a resolution of 12-bit and a sampling rate of 256Hz), and communication interfaces (such as USART, IIC, SPI, etc.). The photoelectric sensing elements are used to convert the light information that passes through the skin tissue and is then reflected back to the sensor into electrical information. The digitizing elements are used to convert the electrical information into digitally transmittable information at a specified sampling rate. The digital information is transmitted to external devices through the communication interface, realizing high-fidelity, low-distortion, and interference-resistant fingertip PPG signal acquisition, providing high-quality raw data for the extraction of coronary heart disease features.

[0023] The data storage module is used to store the raw fingertip pulse wave signal, preprocessed time-frequency characteristics, and final risk report locally and in the cloud. Storage media include PCs, embedded storage, cloud servers, WeChat mini-program cloud databases, etc.

[0024] The preprocessing and coronary heart disease risk prediction module executes the complete algorithm flow described in Example 1: filtering and denoising the original photoplethysmogram, baseline correction, and time-frequency domain transformation are performed to convert the one-dimensional pulse wave signal into a two-dimensional time-frequency feature matrix. The model calculation process is divided into two stages: the forward propagation stage outputs a deep feature vector from the two-dimensional time-frequency matrix through CNN operations, which can be accelerated by a GPU or edge NPU; the classification stage outputs the risk level from the deep feature vector through a random forest model, which involves only simple tree node comparisons and addition operations and can be completed in real time on low-power microcontrollers such as the ARM Cortex-M series. This allows the system to maintain high accuracy while having extremely low hardware cost and power consumption.

[0025] This embodiment provides a method for predicting the risk of coronary artery disease (CAD) by incorporating deep features. The method constructs a CAD risk prediction model through the following steps: acquiring contact pulse wave photoplethysmography (CPPG) of the target population, data preprocessing, converting a one-dimensional matrix to a two-dimensional matrix, extracting deep coronary artery features, and training the CAD risk prediction model. Figure 1 As shown.

[0026] Specifically, the core pathological mechanism of coronary heart disease follows a complete conduction chain: coronary artery stenosis, myocardial ischemia, abnormal cardiac function, hemodynamic disturbances, and peripheral pulse waveform distortion. For example... Figure 2 As shown, myocardial ischemia weakens myocardial contractility and reduces left ventricular ejection fraction (LVEF), directly leading to a decrease in systolic peak amplitude (amplitude at point A) and a prolonged rise time (O→A'). Coronary artery disease is often accompanied by systemic atherosclerosis, resulting in depletion of aortic elastic reserve and a decrease in dicrotic ratio (the ratio of amplitude at point C to point A). Simultaneously, vascular sclerosis and increased peripheral resistance lead to a prolonged fall time (A'→O'). These abnormal combinations of time-domain parameters can serve as important information for early warning of coronary artery disease and can be used to construct a coronary artery disease risk warning model.

[0027] A photoelectric pulse sensor is used to acquire fingertip contact pulse photoplethysmography (PPG) images of a target group sufficient for training purposes, and a one-dimensional time-series matrix is ​​extracted. In one example, specifically, at least 1000 sample objects need to be collected to ensure the comprehensiveness and representativeness of the features of the PPG images used for training the model, and these samples are categorized into three groups based on clinical scales and clinician experience: high risk of coronary heart disease, moderate risk of coronary heart disease, and low risk of coronary heart disease.

[0028] In one embodiment, the preprocessing module is configured with a bandpass filtering unit, which is used to process noise data in the pulse wave photoplethysmography (PPG). The bandpass filtering unit is configured with a Butterworth bandpass filter. Bandpass filtering and baseline correction are performed on the one-dimensional time series matrix to eliminate high-frequency noise and low-frequency artifacts. The original one-dimensional PPG signal is easily overwhelmed by noise, and direct observation cannot identify subtle pathological morphological changes related to coronary heart disease. Filtering and purification are necessary to remove interference in order to preserve and highlight the implicit pathological features of coronary heart disease. Specifically, a Butterworth bandpass filter of 0.5-4Hz can be used. The transfer function of the Butterworth bandpass filter can be expressed as:

[0029] Where sc is the cutoff frequency and n is the filter order; it achieves flat passband filtering, avoids waveform distortion, and protects key morphologies related to coronary heart disease, such as systolic peaks, diphtheria waves, and notches, from being destroyed by filtering.

[0030] The mapping and transformation module is configured with a time-frequency transformation algorithm, which is configured to use a 3-level wavelet decomposition algorithm based on the sym2 wavelet basis.

[0031] The processed one-dimensional time-series matrix is ​​then subjected to time-frequency transformation to expand its dimension to a two-dimensional time-frequency matrix. The original one-dimensional pulse wave can only reflect the amplitude changes in the time domain, failing to present the frequency domain pathological information corresponding to coronary heart disease, and cannot be effectively extracted by deep learning. Through time-frequency transformation, the pathological information of coronary heart disease hidden in the waveform can be expanded from one-dimensional time series to a two-dimensional time-frequency space, realizing the visualization, enhancement, and extractability of pathological features. Specifically, a two-dimensional matrix can be obtained by performing a three-level wavelet decomposition using the sym2 wavelet basis. The mathematical expression for wavelet transform is:

[0032] Where W(a,b) are the continuous wavelet transform coefficients, representing the time-frequency characteristic components of the one-dimensional pulse wave time-series signal x(t) at scale a and translation b; ψ* is the wavelet basis function, where a is the scale parameter, corresponding to different frequency components, used to capture low-frequency arteriosclerosis and high-frequency rhythm abnormalities related to coronary artery disease; b is the translation parameter, locating the time-domain position and accurately matching key morphological points of coronary artery disease such as systolic peaks, notches, and dicrotic waves. Discrete wavelet decomposition achieves signal sub-band decomposition through multi-resolution analysis, decomposing the signal into time-frequency features at different levels, solving the problem that one-dimensional PPG cannot simultaneously display time-domain morphology and frequency-domain pathology. (Refer to...) Figure 6As shown, this figure verifies the effectiveness of wavelet decomposition in enhancing pathological features through a gridded comparison of two sets of time-frequency spectra and their difference maps: Top (healthy group): Shows the one-dimensional waveform and its clear two-dimensional wavelet decomposition spectrum. The healthy group has a clear and coherent diphtheria wave (C), which is manifested as a significant energy distribution in a specific frequency band region of the two-dimensional spectrum. Middle (coronary artery disease risk group): Shows the one-dimensional blunted waveform and the corresponding two-dimensional wavelet spectrum. Affected by pathology, the diphtheria wave disappears, causing its corresponding energy distribution region on the two-dimensional spectrum to shrink or even disappear. Bottom (difference map): Through subtraction, the key time-frequency regions affected by pathology are highlighted on the difference map. Wavelet transform can separate this weak, nonlinear, high-dimensional, and unsteady pathological information from the one-dimensional signal and map it to two-dimensional space, realizing the visualization and extraction of pathological features.

[0033] The feature extraction layer of the two-dimensional convolutional neural network comprises three execution sub-layers and a coronary-aware PPG module (CAPM). Each execution sub-layer includes a convolution module, a batch normalization module, a ReLU activation module, and a max-pooling module. A deep learning-machine learning fusion classification model is constructed, and feature vectors of the two-dimensional time-frequency matrix are extracted using a convolutional neural network. The length N of the feature vector depends on the output size of the fully connected layer of the convolutional neural network. The feature vectors obtained from deep learning training are used as the feature dataset for a machine learning classification model, which is then used to construct a risk prediction model for coronary heart disease.

[0034] In one example, a two-dimensional convolutional neural network model is first constructed to extract features from the original time-frequency matrix. Coronary artery disease (CAD) time-frequency maps are characterized by strong feature locality, dispersed pathological regions, and weak early signals, making traditional manual feature extraction insufficient to cover all pathological patterns. Convolutional neural networks possess powerful feature extraction capabilities, not limited to traditional manual extraction of time-domain, frequency-domain, or nonlinear features. They can deeply mine key CAD features related to myocardial ischemia, arteriosclerosis, and decreased cardiac function in pulse wave photoplethysmography, thus compensating for the shortcomings of traditional manual feature extraction. Specifically, the feature extraction layer of this convolutional neural network consists of three execution sublayers and a CAPM module. Each execution sublayer includes two 3x3 convolutional modules, a batch normalization module, a ReLU activation module, and a 2x2 max-pooling module. After three layers of convolution and pooling, the feature dimension is reduced to a dimension separable by the softmax module through a fully connected layer. Specifically, the convolution operation can be represented as:

[0035] Where I is the input two-dimensional time-frequency feature map; K is the convolution kernel weight matrix; (i, j): the coordinate position of the output feature map; (m, n) is the traversal index inside the convolution kernel; Ii+m,j+n is the pixel value of the input feature map at the corresponding position; Km,n is the weight value of the convolution kernel at the corresponding position; ∗ represents the two-dimensional convolution operation, during which the pathological regions of coronary heart disease such as contraction peaks, notches, and diphtheria waves in the pulse wave time-frequency map can be locally perceived.

[0036] The ReLU activation function is defined as:

[0037] Where x is the output feature value of the convolutional layer or fully connected layer; it is used to introduce nonlinear mapping to enhance the network's ability to fit the complex pathological features of coronary heart disease.

[0038] The max pooling operation can be represented as:

[0039] Where: pi,j is the value of the output feature map at position (i, j) after pooling; w is the range of the pooling window; (m,n) is the index within the pooling window; Ii+m,j+n: the feature value within the area covered by the pooling window; the pooling operation is used for downsampling and dimensionality reduction, preserving key pathological features of coronary heart disease, and suppressing noise and overfitting.

[0040] To preserve sufficient time-frequency resolution for extracting physiological features of coronary artery disease, CAPM is embedded after the second execution sublayer. At this point, the feature map size is approximately (T / 4, F / 4), where T is time and F is frequency. The module can still distinguish between systole and diastole, as well as different frequency components. CAPM consists of two parallel branches: a time-frequency morphological convolution branch and a time-frequency attention branch for cardiac cycle sensing, which are finally fused by 1×1 convolutions and supplemented with residual connections. First, there's the time-frequency morphological convolution branch. This branch directly models waveform pathological features using medical priors. It includes a systolic peak rounding detector using a large kernel convolution (e.g., 7×1) along the time axis, initialized with a Gaussian second derivative kernel to capture the broad and flat peak shape. The rising rate detector uses a differential convolution kernel [-1,0,1] along the time axis to extract the PPG rising slope, reflecting the prolonged rise and fall times characteristic of coronary heart disease mentioned earlier. The diphtheria wave microstructure detector corresponds to the high-frequency sub-band of wavelet decomposition, using a small time-scale convolution (3×1) to enhance the response of early diastolic high-frequency fluctuations. The frequency decay detector uses a 1×5 convolution along the frequency axis to learn the decay trend of the low-frequency-to-high-frequency energy ratio, corresponding to an increased hardening index. The convolution operation in this branch is the same as the convolution operation described above, differing only in the kernel size and direction. Finally, the convolution results are cascaded and activated by ReLU to obtain the morphological enhancement features.

[0041] The time-frequency attention branch for cardiac cycle sensing enables the network to autonomously locate the systolic / diastolic phases of each cardiac cycle and assign importance weights to frequency bands. Temporal gating performs global average pooling along the frequency axis on the input feature map, yielding a time series S(t), which is fed into a small Temporal CNN + Sigmoid layer to generate a time-gated vector G(t). Its peak value represents the systolic peak or the start of systole. Frequency attention performs global average pooling along the time axis on the feature map, yielding a frequency distribution P(f), which is then passed through two fully connected layers + Sigmoid layers to output frequency weights W(f). The outer product A(t,f) = G(t) ⊗ W(f) is calculated to obtain an attention mask with dimensions matching the feature map. This mask is multiplied back into the original feature map to obtain attention-enhanced features, generating the time-frequency attention map. This branch explicitly assigns different levels of attention to different cardiac phases to adapt to pathological localization of "blunting systolic peaks and disappearance of diastolic dicrotic waves."

[0042] Finally, the morphological enhancement features and attention enhancement features are concatenated along the channel direction, and the channel information is integrated through a 1×1 convolution. This integrated information is then added to the input features of CAPM to output the enhanced feature map. This enhanced feature map is then fed into sub-layer 3 and global average pooling to obtain the final feature vector.

[0043] The model training module is used to train the disease analysis model. The model training module is configured with a training strategy, which includes a cross-validation sub-strategy. It is particularly important to note that the deep information extraction module of this application differs fundamentally from the manual feature extraction methods in existing technologies. The latter relies on researchers manually designing statistical features (such as RR interval standard deviation, LF / HF, Poincare scatter plot parameters, etc.). These features are highly dependent on prior clinical knowledge and, once designed, remain fixed, making it difficult to adapt to feature variations in different populations and disease stages. This application uses a two-dimensional convolutional neural network to automatically learn features end-to-end from time-frequency spectra, without any manual feature design. The weights of the convolutional kernels are autonomously optimized during training, enabling the discovery of hidden pathological patterns that human experts have not foreseen. This automated feature mining mechanism makes this application significantly superior to methods relying on fixed manual features in terms of sensitivity and generalization ability in coronary heart disease screening.

[0044] Furthermore, the two-dimensional time-frequency matrix dataset is trained using cross-validation. For example, 1 / 10 and 1 / 5 of the data are used as test sets each time to form 10-fold cross-validation and 5-fold cross-validation, ensuring the randomness and independence of samples during the partitioning process. During training, model parameters are adjusted to optimize the expression of coronary artery disease (CAD) pathological features. These adjustments include, but are not limited to, learning rate, loss function, batch size, number of convolutional kernels, and network depth. The goal is to maximize the extraction of sensitive CAD features and obtain the optimal feature set. CAD data suffers from issues such as a small number of positive samples, class imbalance, and significant individual variability. Conventional training can easily lead to overfitting or missing early pathological features. A dynamic learning rate is employed, with rapid convergence in the early stages to capture the overall time-frequency structure, followed by fine-grained convergence in the later stages to focus on subtle CAD pathological morphologies such as contraction peaks, diphtheria waves, and notches. The number of convolutional kernels is adjusted by progressively increasing the number of kernels layer by layer, allowing the network to learn multi-scale pathological features of CAD from early to high-risk stages from a coarse to a fine scale. The network depth is adjusted to a maximum of three convolutional layers to avoid blurring of pathological features and to ensure that key CAD morphologies are not over-abstracted. Finally, the batch size is adjusted to use mini-batch training, improving the model's ability to learn features from a small number of positive CAD samples and reducing the impact of sample imbalance.

[0045] The loss function commonly used is cross-entropy loss, and the formula is:

[0046] in The overall cross-entropy loss value of the model; N is the total number of samples in a single training batch; yi is the true coronary heart disease risk label of the i-th sample, which adopts multi-class cross-entropy in the multi-classification scenario of this invention, and the label is low, medium and high risk one-hot encoded; The model outputs the predicted probability of coronary heart disease risk for the i-th sample; the loss function is used to measure the error between the predicted probability distribution and the true label, and guides the backpropagation optimization of the network to improve the accuracy of coronary heart disease risk prediction.

[0047] Furthermore, the optimal feature set obtained is used to construct a more accurate coronary heart disease risk prediction model using machine learning algorithms. Although deep learning can extract deep features, it is prone to insufficient generalization on small sample coronary heart disease data. Machine learning can improve robustness, and its accuracy depends on the quality of features and the selection of parameters. As mentioned above, the convolutional neural network addresses the shortcomings of traditional manual feature extraction. Furthermore, the parameters of the machine learning are fine-tuned to further improve the accuracy of coronary heart disease risk prediction.

[0048] The disease analysis model is equipped with a non-linear random forest classifier.

[0049] Specifically, a non-linear random forest classifier can be used. Its decision boundary is smoother than the Softmax layer of a neural network, offering higher tolerance and classification accuracy for edge samples in the feature space. To address the redundancy, noise, and strong non-linearity issues in coronary heart disease features, multiple decision trees are used in parallel voting to reduce the bias of a single model and solve the problem of inaccurate identification of edge and early-stage coronary heart disease samples by a single model. The core logic of the genetic algorithm is to solve the problem that manual parameter tuning cannot adapt to the distribution of coronary heart disease data and is prone to getting trapped in local optima. Through selection, crossover, and mutation, it automatically searches for the optimal parameter combination, using coronary heart disease classification accuracy as the fitness objective, to achieve the best match between features and the classifier.

[0050] Optimal prediction accuracy is achieved by optimizing the number of trees, the depth of each tree, and the number of nodes in a random forest using a genetic algorithm. Random forests predict outcomes through voting among multiple decision trees. For classification problems, the prediction result can be expressed as:

[0051] Where ht(x) represents the prediction of the t-th tree, and T is the number of trees. The genetic algorithm optimizes parameters through selection, crossover, and mutation operations, and the fitness function is the classification accuracy or other metrics.

[0052] It should be noted that although the above example illustrates a method for predicting coronary heart disease using pulse wave photoplethysmography, those skilled in the art will understand that this disclosure is not limited thereto. In fact, users can flexibly customize various implementation methods according to their actual application scenarios.

[0053] In summary, this embodiment provides a coronary heart disease prediction method that differs from existing technologies. Compared with manual statistical features and traditional classification methods, this application employs time-frequency spectrograms and CNN-based automatic feature mining, eliminating the need for manual feature design and avoiding information loss and incomplete knowledge, thus significantly reducing system complexity and signal preprocessing errors. Simultaneously, it uses a nonlinear classification method to output more clinically relevant discrete risk levels (low / medium / high). Furthermore, this application requires only a single PPG sensor, greatly reducing hardware costs and wearing constraints, and achieves efficient nonlinear classification of deep features through a heterogeneous serial architecture, offering significant advantages in prediction accuracy and computational efficiency. Through the synergistic optimization of the above algorithmic pathways, this application achieves non-invasive, rapid, low-cost, and high-precision large-scale screening for coronary heart disease.

[0054] Of course, the above are just typical examples of the present invention. In addition, the present invention may have many other specific embodiments. All technical solutions formed by equivalent substitution or equivalent transformation fall within the scope of protection claimed by the present invention.

Claims

1. A pulse wave-based coronary heart disease risk prediction system based on deep features, characterized in that: include: A contact pulse acquisition device is used to acquire pulse wave photoplethysmography of a target. The preprocessing module is used to process the pulse wave photoplethysmogram to obtain a one-dimensional representation time series matrix; The mapping and transformation module maps the one-dimensional time-series matrix to a two-dimensional time-frequency matrix based on a pre-constructed two-dimensional time-frequency mapping relationship. The feature extraction module extracts coronary heart disease features from the two-dimensional time-frequency matrix representing the disease based on a preset two-dimensional convolutional neural network. The model analysis module is equipped with a disease analysis model, which is used to generate corresponding diagnostic results information based on the characteristics of coronary heart disease.

2. The pulse wave-based coronary heart disease risk prediction system as described in claim 1, characterized in that: The preprocessing module is equipped with a bandpass filter processing unit, which is used to process noise data in the pulse wave photoplethysmography. The bandpass filter processing unit is equipped with a Butterworth bandpass filter.

3. The pulse wave-based coronary heart disease risk prediction system as described in claim 1, characterized in that: The mapping and transformation module is configured with a time-frequency transformation algorithm, which is configured to use a 3-level wavelet decomposition algorithm based on the sym2 wavelet basis.

4. The pulse wave-based coronary heart disease risk prediction system as described in claim 1, characterized in that: The feature extraction layer of the two-dimensional convolutional neural network includes three execution sub-layers and one CAPM module. Each execution sub-layer includes a convolution module, a batch normalization module, a ReLU activation module, and a max pooling module. The CAPM module consists of a time-frequency morphological convolution branch and a time-frequency attention branch for cardiac cycle perception.

5. The pulse wave-based coronary heart disease risk prediction system as described in claim 4, characterized in that: The convolution module is configured with an execution expression, which is: ; in, The input is a two-dimensional time-frequency feature map; Let be the convolution kernel weight matrix; (i, j) be the coordinates of the output feature map; (m, n) be the traversal index within the convolution kernel; I i+m,j+n K represents the pixel value at the corresponding location in the input feature map. m,n The weights of the convolution kernel at the corresponding positions are represented by '*'; '*' indicates a two-dimensional convolution operation.

6. The pulse wave-based coronary heart disease risk prediction system as described in claim 4, characterized in that, The time-frequency morphological convolution branch includes a main peak rounded convolution kernel along the time axis, an ascending branch rate convolution kernel, a diabetic wave microstructure convolution kernel, and a band energy attenuation convolution kernel.

7. The pulse wave-based coronary heart disease risk prediction system as described in claim 4, characterized in that, The cardiac cycle perception time-frequency attention branch includes time gating and frequency attention; time gating generates time weights by global pooling along the frequency axis + temporal CNN + sigmoid; Frequency attention is generated by global pooling along the time axis, two fully connected layers, and a sigmoid function to generate frequency weights; the outer product of the two weights forms an attention mask to enhance key time-frequency features.

8. The pulse wave-based coronary heart disease risk prediction system as described in claim 1, characterized in that: It also includes a model training module, which is used to train the disease analysis model. The model training module is configured with a training strategy, which includes a cross-validation sub-strategy.

9. The pulse wave-based coronary heart disease risk prediction system as described in claim 8, characterized in that: The disease analysis model is equipped with a non-linear random forest classifier.