A Multimodal Screening System for Cardiovascular and Cerebrovascular Diseases Guided by Manual Features

Abstract

This invention discloses a multimodal cardiovascular and cerebrovascular disease (CCD) screening system guided by handcrafted features, comprising: a data acquisition and processing module for acquiring and preprocessing multi-channel physiological signal data; a multi-granularity feature extraction module for extracting coarse-grained features from the physiological signal data, forming a tabular data set as a coarse-grained handcrafted feature matrix; simultaneously, passing the physiological signal data through a deep convolutional network to form a fine-grained signal feature matrix; a feature fusion module for weighting the coarse-grained handcrafted feature matrix into the fine-grained feature matrix using attention, filtering out redundant features with low importance to obtain fused features; and a preliminary screening and prediction module for inputting the fused features into a deep learning classification network to output the risk of cardiovascular and cerebrovascular diseases, thus performing preliminary screening for the risk of these diseases. Using this invention, the accuracy of predicting the incidence of cardiovascular and cerebrovascular diseases can be improved, achieving preliminary screening for the risk of these diseases.

Description

Technical Field

[0001] This invention relates to the field of initial screening and prediction technology for cardiovascular and cerebrovascular diseases, and in particular to a multimodal initial screening system for cardiovascular and cerebrovascular diseases based on manual feature guidance. Background Technology

[0002] Many factors influence cardiovascular and cerebrovascular diseases (CVD), and early and accurate diagnosis is crucial for promoting effective treatment and reducing the risk of related mortality. The diagnostic process generates a wide variety of data, including tabular data and physiological signal data. Multiple modalities of data can assist in preliminary screening and treatment, improving the accuracy and stability issues that arise with single-modal data diagnosis.

[0003] However, traditional methods for detecting cardiovascular and cerebrovascular diseases rely on manual operation by doctors and only utilize medical diagnostic data from a single modality, resulting in insufficient intelligent information fusion.

[0004] Chinese patent document CN114903445A discloses an intelligent monitoring and early warning system for cardiovascular and cerebrovascular diseases, comprising: a wearable device, a data acquisition module, a transmission module, a smart terminal module, and a display module; the wearable device is worn on the wrist of the subject, and the data acquisition module, transmission module, and display module are all mounted on the wearable device; the data acquisition module is used to acquire physiological signals from the subject; the transmission module is used to transmit the physiological signals to the smart terminal module, and also to transmit data from the smart terminal module to the display module; the smart terminal module is used to process the physiological signals; and the display module is used to provide early warning of cardiovascular and cerebrovascular disease risks and display the predicted cardiovascular and cerebrovascular disease outcomes.

[0005] Chinese patent document CN116269391A discloses a method and system for cardiac-brain coupling analysis and assessment. The method involves real-time synchronous acquisition of brain oxygen and electrocardiogram (ECG) signals from multiple test users, followed by data preprocessing to obtain processed brain oxygen signals, processed ECG signals, and brain blood oxygen variability. Feature extraction is performed on the processed ECG signals to calculate ECG feature sets, heart rate signals, and heart rate variability feature sets. Cardiac-brain coupling analysis and assessment are then performed based on the processed brain oxygen and heart rate signals to obtain a cardiac-brain coupling feature set. Finally, an integrated learning method is used to establish a mental stress assessment model, which corrects the output mental stress values, thereby improving the accuracy and applicability of mental stress assessment.

[0006] The aforementioned patents only employ traditional time-domain, frequency-domain, and heart rate variability feature extraction methods. These methods often rely on manual design and may fail to fully extract deep information from the original signal, neglecting the extraction of local detailed features. Deep convolution can automatically learn more robust feature representations and incorporates knowledge from the cardiovascular and neurological domains, which is particularly important for complex cardiovascular and neurological signals.

[0007] Therefore, the clinical application of an intelligent screening method for cardiovascular and cerebrovascular diseases is of great significance, as it can help patients and doctors prevent diseases in their early stages. Summary of the Invention

[0008] This invention provides a multimodal screening system for cardiovascular and cerebrovascular diseases based on manual feature guidance, which can improve the accuracy of predicting the incidence of cardiovascular and cerebrovascular diseases and achieve preliminary screening of the risk of cardiovascular and cerebrovascular diseases.

[0009] A multimodal primary screening system for cardiovascular and cerebrovascular diseases based on handcrafted feature guidance includes:

[0010] The data acquisition and processing module is used to acquire multi-channel physiological signal data and perform preprocessing.

[0011] The multi-granularity feature extraction module is used to extract coarse-grained features from the preprocessed physiological signal data and organize the extracted features into tabular data as a coarse-grained manual feature matrix. At the same time, the preprocessed physiological signal data is processed by a deep convolutional network to form a fine-grained signal feature matrix.

[0012] The feature fusion module is used to weight the coarse-grained manual feature matrix into the fine-grained feature matrix in the form of attention, filter out redundant features with low importance, and obtain the fused features.

[0013] The initial screening and prediction module inputs fused features into a deep learning classification network and outputs results for cardiovascular and cerebrovascular diseases, thus providing an initial screening of the risk of developing these diseases.

[0014] Furthermore, physiological signal data include electroencephalogram (EEG), electrocardiogram (ECG), blood oxygen saturation (SpO2), blood pressure (HP), and respiratory rate (RPF).

[0015] The data acquisition and processing module acquires multi-channel physiological signal data, specifically:

[0016] At a specific sampling frequency f c Simultaneous acquisition of signals from multiple physiological signal channels, and their arrangement into a multidimensional data structure in chronological order. Where c is the number of signal channels; to facilitate the representation of the downsampled signal, n·t is used as the subscript in the timestamp sequence, n·t = 1, 2, 3...T, where T represents the total number of samples; the sampled value of each physiological signal channel corresponds to a specific timestamp, and there is a correlation between the time series of each signal channel of the same subject.

[0017] Preprocessing includes denoising and downsampling. Denoising involves using high-pass filtering, low-pass filtering, and wavelet processing to remove high-frequency noise, low-frequency interference, and baseline drift from the signal. Downsampling is then applied to the denoised signal to reduce the sampling rate by retaining some sampling points from the original signal. The downsampling rate is n.

[0018] In the multi-granularity feature extraction module, coarse-grained feature extraction is performed on the preprocessed physiological signal data, specifically as follows:

[0019] EEG signals electrocardiogram signal blood oxygen saturation blood pressure respiratory rate Extracting general frequency domain features F fre General time-domain features F time For electrocardiogram signals Extracted heart rate variability feature F hrv General time-domain features include maximum, minimum, mean, standard deviation, variance, root mean square (RMS), skewness, and kurtosis; general frequency-domain features include mean frequency, RMS frequency, and standard deviation frequency; heart rate variability features include the standard deviation F of the RR interval extracted based on the time interval sequence between adjacent heartbeats. sdnn RR interval root mean square F rmssd Number of times the RR interval difference exceeds 50ms (F) nn50 and percentage F pnn50 ;

[0020] The general frequency domain feature F fre General time-domain features F time Heart rate variability characteristics F hrv By piecing them together, coarse-grained handcrafted features are formed.

[0021] In the multi-granularity feature extraction module, the preprocessed physiological signal data is processed through a deep convolutional network to form a fine-grained signal feature matrix, specifically:

[0022] Preprocessed multichannel physiological signals A multi-layer deep convolutional network is used for layer-by-layer convolution, with each layer containing a pair of convolutional layers C1(·) and C2(·) and a residual connection layer R(·). The following formula is used to process multi-channel physiological signals. Fine-grained feature extraction is performed to obtain fine-grained feature F. sig

[0023]

[0024] In the formula, l represents the number of convolutional layers in the deep network.

[0025] In the feature fusion module, the coarse-grained handcrafted feature matrix is ​​weighted into the fine-grained feature matrix using attention, specifically as follows:

[0026] Handcrafted feature F hand via fully connected function f fc (·) Feature projection, the projected handcrafted features are f fc (F hand Then calculate the weighted feature F after coarse-grained manual feature guidance. weight ;Weighted feature F weight Compared with the projected handcrafted features f fc (F hand The features are then spliced ​​together to obtain the fusion characteristics.

[0027] The weighted features guided by coarse-grained manual features are calculated using the following formula:

[0028] F weigh =softmax(f fc (F hand )·F sig T )·F sig

[0029] In the formula, F sig It is a fine-grained feature.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] This invention utilizes multiple data sources and an innovative fusion-guided method to extract multi-granularity feature information from physiological signal monitoring data, including coarse-grained tabular handcrafted features and fine-grained physiological signal features. The coarse-grained tabular features are fused into the signal features using a weighted attention approach, removing redundant information from long-term series data while selectively considering deep convolutional features whose potential distribution aligns with the coarse-grained features, thereby improving prediction accuracy. Attached Figure Description

[0032] Figure 1 This is a structural diagram of a multimodal primary screening system for cardiovascular and cerebrovascular diseases based on manual feature guidance according to the present invention;

[0033] Figure 2 This is a structural diagram of the multi-granularity feature extraction module in this invention;

[0034] Figure 3 This is a structural diagram of the feature fusion module in this invention;

[0035] Figure 4 This is a comparison chart of prediction results in an embodiment of the present invention. Detailed Implementation

[0036] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be noted that the embodiments described below are intended to facilitate the understanding of the present invention and do not constitute any limitation thereof.

[0037] like Figure 1 As shown, a multimodal screening system for cardiovascular and cerebrovascular diseases based on manual feature guidance includes: a data acquisition and processing module, a multi-granularity feature extraction module, a feature fusion module, and a screening prediction module.

[0038] The data acquisition and processing module is used to acquire raw multi-channel physiological signal data. In home or hospital environments, multi-channel physiological monitoring signal data of specific groups are acquired using wearable medical devices. These multi-channel physiological monitoring signals include electroencephalogram (EEG), electrocardiogram (ECG), blood oxygen saturation (SpO2), blood pressure (HP), and respiratory rate (RPF). The multi-channel physiological signals acquired at a high sampling rate are preprocessed using methods including noise reduction and downsampling.

[0039] A multi-granularity feature extraction module is used to extract coarse-grained features from physiological signal data. This feature extraction is performed under the guidance of knowledge pairs in the field of cardiovascular and cerebrovascular diseases, and includes general frequency domain features, general time domain features, and heart rate variability features extracted from electrocardiogram (EEG) signals. The extracted features are compiled into tabular data, serving as a coarse-grained handcrafted feature matrix. The physiological signal data is then processed through a deep convolutional network to form a fine-grained signal feature matrix.

[0040] The feature fusion module is used to fuse multimodal and multi-granularity feature matrices. This invention employs an advanced hand-crafted feature-guided multimodal feature fusion technique, which weights the coarse-grained hand-crafted feature matrix into the fine-grained feature matrix using attention, filters out redundant features with low importance, and obtains the fused features.

[0041] The initial screening and prediction module is used to predict and assess an individual's cardiovascular and cerebrovascular risk and compare it with the actual category. Through the fusion of multimodal and multigranular features, it obtains the prevalence of cardiovascular and cerebrovascular diseases, and performs preliminary screening for the risk of developing these diseases.

[0042] In this embodiment of the invention, multi-channel physiological signal acquisition is performed at a specific sampling frequency f. c Simultaneous acquisition of signals from multiple physiological signal channels, and their arrangement into a multidimensional data structure in chronological order. Where c is the number of signal channels. To facilitate the representation of the downsampled signal, n·t is used as the subscript in the timestamp sequence, n·t = 1, 2, 3...T, where T represents the total number of samples. The sampled value of each physiological signal channel corresponds to a specific timestamp, and there is a correlation between the time series of different signal channels of the same subject.

[0043] Data preprocessing methods are a type of method based on multi-channel physiological signals. Based on the inherent properties of the signal, high-pass filtering, low-pass filtering, and wavelet denoising are used to remove high-frequency noise, low-frequency interference, and baseline drift. A comprehensive downsampling method is then applied to the denoised signal. A portion of the original signal's sampling points are retained to reduce the sampling rate, which is n. The resulting c-dimensional physiological signal is represented as follows:

[0044] Due to multiple factors such as respiration, body movement, poor electrode contact, and skin electrode impedance, multi-channel physiological monitoring signals often contain noise or unwanted artifacts, which can affect signal analysis. Therefore, in the data acquisition and processing module, a multi-signal denoising method is used for preprocessing, including high-pass filtering, low-pass filtering, and wavelet denoising. This method not only removes baseline drift, power frequency interference, and high-frequency noise from the signal but also preserves key signal features, avoiding the loss of useful information. For the denoised signal, downsampling is performed to reduce unnecessary redundant information generated at high sampling frequencies.

[0045] Multi-granularity feature extraction module, such as Figure 2 The image illustrates a feature extraction method for multi-channel physiological signal data, guided by knowledge in the field of cardiovascular and cerebrovascular diseases. Based on knowledge in this field, the heart rate variability of ECG signals contains rich features related to the heart and brain. Specifically, the time interval between two adjacent R-wave peaks represents a complete cardiac cycle, and its standard deviation, root mean square value, and heart rate variability can all reflect cardiac health. Coarse-grained feature extraction includes analysis of the electroencephalogram (EEG) signal... electrocardiogram signal blood oxygen saturation blood pressure respiratory rate Extracting general frequency domain features F fre General time-domain features F time and electrocardiogram signals Heart rate variability features extracted from F hrv The time-domain characteristics of the signal include the maximum value F. max Minimum value F min Mean F mean Standard deviation F std Variance F var Root mean square value Frms skewness F ske , kurtosis F kur The frequency domain characteristics of the signal include the frequency mean F. avg , frequency root mean square F square Frequency standard deviation F fre_std The heart rate variability features include the standard deviation F of the RR interval extracted based on the time interval sequence between adjacent heartbeats. sdnn RR interval root mean square F rmssd Number of times the RR interval difference exceeds 50ms (F) nn50 and percentage F pnn50 The calculations are performed using the following formulas:

[0046]

[0047]

[0048]

[0049] in, It is a signal The spectrum of the signal, k = 1, 2, ..., K, where K is the number of spectral lines, f(k) is the frequency value of the k-th spectral line, N is the total number of RR intervals in the ECG signal, and I(·) is an indicator function that takes a value of 1 when the condition in parentheses is true, and 0 otherwise. Time-domain characteristics Frequency domain characteristics Heart rate variability characteristics

[0050] Time-domain features, frequency-domain features, and heart rate variability features are concatenated to form coarse-grained handcrafted features. This constitutes a handcrafted feature matrix based on domain knowledge.

[0051] like Figure 2 As shown in the fine-grained signal feature extraction, this is a method for preprocessing multi-channel physiological signals. The method employed is a layer-by-layer convolutional network with a depth of d layers, consisting of stacked l basic network layers. Each basic layer contains a pair of convolutional layers C1(·) and C2(·) and a residual connection layer R(·). For signal time series, the following formula is used to calculate the signal... Fine-grained cross-time step feature extraction is performed to obtain fine-grained features. The fine-grained feature F sig The dimension is n×d, where n is the number of samples and d is the feature length after processing by the deep convolutional network.

[0052] Multidimensional physiological monitoring signals are complex and variable data structures containing multi-layered information. In such cases, relying solely on single-granularity feature information may be insufficient to comprehensively capture the effective information within the signal, especially in the early stages of disease manifestation. Therefore, our proposed multi-granularity feature extraction module combines coarse-grained and fine-grained features extracted from multidimensional physiological monitoring signals to achieve a comprehensive representation of different information levels within the signal. Manually extracted coarse-grained features represent information highly relevant to cardiovascular and cerebrovascular health research, while fine-grained features, extracted through deep convolutional networks, capture complex features that are difficult to design manually, achieving higher feature discrimination. The necessity of this multi-granularity feature extraction method lies in its ability to strike a balance between refinement and comprehensiveness, providing richer feature information for accurate disease diagnosis.

[0053] The weighted fusion method in the feature fusion module is as follows: Figure 3 As shown, a classification model for cardiovascular and cerebrovascular diseases guided by handcrafted features from multi-granular and multimodal data is proposed, which can balance feature information at different levels between coarse-grained and fine-grained features. The core idea is to utilize an attention mechanism to influence fine-grained, multi-channel physiological signal features with coarse-grained handcrafted tabular features, and then fuse the weighted signal features guided by the handcrafted tabular features with the handcrafted tabular features. Disease classification is then performed based on the fused multimodal features. The handcrafted features F... hand via fully connected function f fc (·) Feature projection, which lies in the same feature space R as the signal features. n×d Where n is the number of samples in the group, and d is the number of feature dimensions. Then, the weighted feature F, guided by coarse-grained manual features, is calculated. weight =softmax(f fc (F han )·F sig T )·F sig The weighted features are then concatenated with the projected hand-crafted features to obtain the fused features. First, the handcrafted feature matrix, representing domain knowledge, enables the weighted selection of fine-grained features, ensuring the model focuses on key features of cardiovascular and cerebrovascular diseases and improving feature discriminative power. Second, through an attention mechanism, the model can automatically learn the weight distribution of different features, effectively enhancing its ability to capture potential correlations between multimodal data. This multimodal feature fusion model not only improves the accuracy of classification and recognition but also enhances its robustness, making it suitable for analyzing complex physiological signals and detecting disease states.

[0054] The initial screening and prediction module integrates features and processes them through a deep learning classification network to screen individuals for cardiovascular and cerebrovascular diseases. Based on the individual classification results, it determines whether there is a risk of developing cardiovascular and cerebrovascular diseases, thus achieving early screening.

[0055] like Figure 4 As shown, the method proposed in this invention is compared with two other single-modal methods, and their accuracy metrics are evaluated on the same dataset. The method proposed in this invention achieves superior performance, demonstrating the beneficial effects of this invention.

[0056] The embodiments described above provide a detailed explanation of the technical solutions and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multimodal primary screening system for cardiovascular and cerebrovascular diseases based on manual feature guidance, characterized in that, include: The data acquisition and processing module is used to acquire multi-channel physiological signal data and perform preprocessing. The multi-granularity feature extraction module is used to perform coarse-grained feature extraction on the preprocessed physiological signal data, and to organize the extracted features into tabular data as a coarse-grained manual feature matrix. Simultaneously, the preprocessed physiological signal data is processed through a deep convolutional network to form a fine-grained signal feature matrix. Specifically, the coarse-grained feature extraction on the preprocessed physiological signal data involves: EEG signals electrocardiogram signal blood oxygen saturation blood pressure respiratory rate Extracting general frequency domain features F fre and general time-domain features F time For electrocardiogram signals Extracted heart rate variability feature F hrv General time-domain features include maximum, minimum, mean, standard deviation, variance, root mean square (RMS), skewness, and kurtosis; general frequency-domain features include mean frequency, RMS frequency, and standard deviation frequency; heart rate variability features include the standard deviation F of the RR interval extracted from the time interval sequence between adjacent heartbeats. sdnn RR interval root mean square F rmssd F times the RR interval difference exceeds 50ms nn50 and percentage F pnn50 ; the general frequency domain feature F fre General time-domain features F time Heart rate variability characteristics F hrv By piecing them together, coarse-grained handcrafted features are formed. The preprocessed physiological signal data is then processed through a deep convolutional network to form a fine-grained signal feature matrix, specifically: Preprocessed multi-channel physiological signals A multi-layer deep convolutional network is used for layer-by-layer convolution, with each layer containing a pair of convolutional layers C1(·) and C2(·) and a residual connection layer R(·). The following formula is used to process multi-channel physiological signals. Fine-grained feature extraction is performed to obtain fine-grained feature F. sig In the formula, l represents the number of convolutional layers in the deep network; The feature fusion module is used to weight the coarse-grained manual feature matrix into the fine-grained feature matrix in the form of attention, filter out redundant features with low importance, and obtain the fused features. The coarse-grained handcrafted feature matrix is ​​weighted into the fine-grained feature matrix using attention techniques, specifically as follows: Handcrafted feature F hand via fully connected function f fc (·) Feature projection, the projected handcrafted features are f fc (F hand Then calculate the weighted feature F after coarse-grained manual feature guidance. weight ;Weighted feature F weight Compared with the projected handcrafted features f fc (F hand The features are then spliced ​​together to obtain the fusion characteristics. The weighted features guided by coarse-grained manual features are calculated using the following formula: F weight =softmax(f fc (F hand )·F sig T )·F sig In the formula, F sig It is a fine-grained feature; The initial screening and prediction module inputs fused features into a deep learning classification network and outputs results for cardiovascular and cerebrovascular diseases, thus providing an initial screening of the risk of developing these diseases.

2. The multimodal primary screening system for cardiovascular and cerebrovascular diseases based on manual feature guidance according to claim 1, characterized in that, Physiological signal data include electroencephalogram (EEG) signals, electrocardiogram (ECG) signals, blood oxygen saturation, blood pressure, and respiratory rate.

3. The multimodal primary screening system for cardiovascular and cerebrovascular diseases based on manual feature guidance according to claim 1, characterized in that, The data acquisition and processing module acquires multi-channel physiological signal data, specifically: At a specific sampling frequency f c Simultaneous acquisition of signals from multiple physiological signal channels, and their arrangement into a multidimensional data structure in chronological order. Where c is the number of signal channels; to facilitate the representation of the downsampled signal, n·t is used as the subscript in the timestamp sequence, n·t = 1, 2, 3...T, where T represents the total number of samples; the sampled value of each physiological signal channel corresponds to a specific timestamp, and there is a correlation between the time series of each signal channel of the same subject.

4. The multimodal primary screening system for cardiovascular and cerebrovascular diseases based on manual feature guidance according to claim 1, characterized in that, In the data acquisition and processing module, preprocessing includes noise reduction and downsampling; In the denoising process, high-pass filtering, low-pass filtering, and wavelet processing are used to remove high-frequency noise, low-frequency interference, and baseline drift from the signal. The denoised signal is then downsampled to retain some sampling points from the original signal to reduce the sampling rate, which is n.

Images