Old people heart abnormality detection auxiliary system based on deep learning

Abstract

The invention discloses an old people heart anomaly detection auxiliary system based on deep learning. The old people heart anomaly detection auxiliary system comprises a data acquisition module, an edge preprocessing module, a multi-modal space-time representation module and an intelligent detection module. The invention relates to the technical field of heart health monitoring, in particular to an old people heart anomaly detection auxiliary system based on deep learning, which realizes efficient denoising through a variable bit width quantized one-dimensional convolutional neural network; double attention is introduced into the time domain and the frequency domain, noise is suppressed in a multi-scale mode, and the extraction capacity of weak pathological waveforms is enhanced; layered self-supervised comparative learning is carried out to obtain general representation, and a dynamic map convolutional network is used to capture time-space association of physiology and environment, so that the accuracy of heart anomaly detection is improved; edge end preliminary screening and high-precision classification are combined through a double-layer inference mechanism, and a risk curve is generated based on continuous time Bayesian variation updating, so that real-time detection of the heart abnormality of the old people is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of cardiac health monitoring, and specifically to a deep learning-based auxiliary system for detecting cardiac abnormalities in the elderly. Background Art

[0002] As the global population continues to age, the incidence and mortality rates of cardiovascular diseases in the elderly continue to rise. Early, continuous, and accurate monitoring of cardiac abnormalities has become key to protecting the health of the elderly. However, existing ECG monitoring systems mostly rely on large medical equipment. Not only are these devices large and power-hungry, making them inconvenient to wear for long periods of time at home or in community settings, but the models also have limited capabilities in noise suppression and weak pathological waveform extraction, lack in-depth interpretation of abnormality detection results, and cannot meet the strict real-time, power consumption, and bandwidth requirements of edge devices.

[0003] In addition, most traditional systems only use a single ECG signal, which makes it difficult to fully utilize unlabeled multimodal time series data. They also lack deep fusion and spatiotemporal correlation modeling of ECG, motion and environmental information, resulting in poor model generalization when samples are scarce and individual differences are large. At the same time, the single-layer diagnostic process can neither take into account rapid early warning at the edge nor achieve high-precision classification and dynamic risk assessment in the cloud, resulting in significant blind spots and delays in the prediction, early warning and continuous monitoring of cardiac abnormalities in the elderly. Summary of the Invention

[0004] In view of the above situation, in order to overcome the defects of the prior art, the present invention provides an auxiliary system for detecting cardiac abnormalities in the elderly based on deep learning. In view of the problems of large model size, high power consumption, insufficient noise suppression and feature extraction capabilities, poor interpretability, and fixed bit width network leading to edge device inference performance bottlenecks in traditional auxiliary systems for detecting cardiac abnormalities in the elderly, this solution automatically balances the error and computational cost through a one-dimensional convolutional neural network with variable bit width quantization to achieve efficient real-time denoising; introduces dual attention multi-scale noise suppression in the time domain and frequency domain to enhance the extraction capability of weak pathological waveforms; and uploads only suspected abnormal fragments to the cloud based on the fused lightweight features; and in view of the fact that traditional systems for detecting cardiac abnormalities in the elderly can often only rely on labeled data for training and find it difficult to use unlabeled time series information, To address the problem of lack of deep fusion and spatiotemporal correlation modeling of multimodal signals such as ECG, activity, and environment, this solution first performs hierarchical self-supervised comparative learning on multimodal unlabeled data to obtain representation, then uses a dynamic graph convolutional network to combine ECG, movement, and environmental features to capture the spatiotemporal correlation between physiology and the environment, and finally amplifies the samples through small-sample fine-tuning combined with generative adversarial networks. Traditional cardiac abnormality detection systems for the elderly rely only on single-layer models for diagnosis, which makes it difficult to balance real-time response at the edge and high-precision classification on the cloud, and also lacks dynamic risk assessment and early warning capabilities. This solution combines initial screening at the edge with high-precision classification on the cloud through a two-layer reasoning mechanism, and generates short-term, medium-term, and long-term risk curves based on continuous-time Bayesian variational updates, thereby realizing real-time detection of cardiac abnormalities in the elderly.

[0005] The technical solution adopted by the present invention is as follows: The deep learning-based elderly heart abnormality detection auxiliary system provided by the present invention includes a data acquisition module, an edge preprocessing module, a multimodal spatiotemporal representation module and an intelligent detection module;

[0006] The data acquisition module collects the elderly person's historical electrocardiogram signals, activity intensity, environmental data and heart status data;

[0007] The edge preprocessing module uses a one-dimensional convolutional neural network with variable bit width quantization to perform real-time denoising and lightweight classification on the ECG signal, while generating features through time domain and frequency domain attention to obtain suspected abnormal data;

[0008] The multimodal spatiotemporal representation module first performs self-supervised comparative learning on unlabeled ECG, activity, and environmental multimodal time series data. By constructing positive and negative sample pairs, it maximizes the similarity of similar embeddings and minimizes the similarity of different embeddings, thereby obtaining physiological signal representations. It then constructs a dynamic graph based on multimodal features, uses a multi-layer perceptron to fuse ECG, motion, and environmental vectors to generate a time-varying adjacency matrix, and then iteratively updates node representations through normalized graph convolution to capture the spatiotemporal correlation between individual physiology and the environment. Finally, based on the global model, it performs gradient fine-tuning using labeled samples and combines them with a generative adversarial network to amplify the samples.

[0009] The intelligent detection module uses a two-layer reasoning mechanism to classify and probabilistically predict suspected abnormal ECG segments. Subsequently, based on the output probability vector, it uses the continuous-time Bayesian update formula to dynamically calculate the prior and posterior risks of abnormal cardiac events, and accumulates integrals to obtain short-term, medium-term, and long-term risk curves, thereby realizing the detection of cardiac abnormalities.

[0010] Furthermore, the data acquisition module collects the elderly's historical electrocardiogram signals, activity intensity, environmental data and heart state data; the electrocardiogram signals specifically include the elderly's electrocardiogram waveform, sampling rate and timestamp; the activity intensity specifically refers to the elderly's three-axis acceleration, sampling rate and timestamp collected by the accelerometer; the environmental data specifically includes the environment's temperature, relative humidity, air pressure and timestamp; the heart state data includes normal state and abnormal state, and the heart state data is set as label data.

[0011] Furthermore, the edge preprocessing module uses a one-dimensional convolutional neural network with variable bit width quantization to perform denoising and preliminary classification. At the same time, dual attention is designed, specifically including the following units:

[0012] The denoising unit performs quantized convolution denoising on candidate bit widths and takes the weighted sum of the mean square error and storage cost. It automatically selects the optimal bit width that minimizes the weighted sum and then quantizes the original signal using the optimal bit width to obtain the denoised signal.

[0013] A dual attention unit is introduced, which uses scaled dot product attention in the time domain and frequency domain respectively. The quantized signal is first mapped into query, key, and value tensors. Then, the dot product of the query and key is divided by the dimension scaling factor and normalized to generate the attention weight matrices in the time domain and frequency domain. These are then multiplied with the value tensor to obtain enhanced time domain and frequency domain feature representations.

[0014] The preliminary cardiac abnormality classification unit uses a one-dimensional convolutional neural network to perform lightweight classification on the quantized time domain features. It also uses the lightweight convolutional network to perform lightweight classification on the fused features, and only uploads suspected abnormal fragments to the cloud.

[0015] Furthermore, the multimodal spatiotemporal representation module first performs hierarchical self-supervised contrastive learning on unlabeled ECG time series data to obtain a representation; then constructs a dynamic map based on multimodal features to capture the spatiotemporal correlation between individual physiological signals and the environment; finally, gradient fine-tuning is performed using labeled data, and samples are amplified using a generative adversarial network. Specifically, it includes the following units:

[0016] The self-supervised contrastive learning unit first encodes ECG, motion, and environmental signals into embedding vectors, constructs positive pairs of samples from the same enhanced view and negative pairs of samples from different views at the same moment, then introduces cosine similarity and temperature coefficient to adjust the contrastive loss. Using a normalized exponential function, it maximizes the similarity of positive pairs and minimizes the similarity of negative pairs, thereby learning multimodal representations.

[0017] The graph convolution node update unit constructs a spatiotemporal adjacency matrix based on multimodal features, dynamically updates the graph convolution network node representation matrix, uses a multi-layer perceptron to fuse ECG, activity, and environmental features to generate a dynamic graph adjacency matrix, and then iteratively updates the node representation matrix through a normalized graph convolution layer;

[0018] The few-sample fine-tuning and amplification unit performs gradient fine-tuning on the global model through labeled data and amplifies samples using a generative adversarial network.

[0019] Furthermore, the intelligent detection module specifically includes the following units:

[0020] The two-layer reasoning output unit sends data segments that are initially identified as abnormal by the edge to the cloud, which then provides classification and probability distribution.

[0021] The Bayesian variational risk scoring unit uses a continuous-time Bayesian update formula to dynamically calculate the short-term, medium-term, and long-term risks of cardiac abnormalities and generate a risk curve that changes over time;

[0022] The real-time detection unit uses the Bayesian variational risk scoring unit to calculate the cumulative risk scores of all samples in the short-term, medium-term and long-term windows and pair them with labels. It then generates a threshold set, traverses and calculates the sensitivity and specificity under each threshold to draw the ROC curve, and selects the threshold that maximizes the Youden index as the optimal threshold for each window; finally, the user data to be detected is collected and preprocessed, the preprocessed user data is received in real time, the posterior risk is iteratively updated and the points are accumulated in the window. When the integral of the window exceeds the optimal threshold, the user's heart state is determined to be abnormal and an alarm is issued; otherwise, it is determined to be normal.

[0023] The beneficial effects achieved by the present invention using the above scheme are as follows:

[0024] (1) In order to address the problems of large model size, high power consumption, insufficient noise suppression and feature extraction capabilities, poor interpretability, and fixed bit-width networks in traditional elderly heart anomaly detection auxiliary systems, this solution uses a one-dimensional convolutional neural network with variable bit-width quantization to automatically balance error and computational cost, achieving efficient real-time denoising; introducing dual attention multi-scale noise suppression in the time domain and frequency domain to enhance the ability to extract weak pathological waveforms; and based on the fused lightweight features, only suspected abnormal fragments are uploaded to the cloud, extending battery life and improving the accuracy and interpretability of early detection of elderly heart anomalies.

[0025] (2) In view of the problem that traditional cardiac anomaly detection systems for the elderly can only rely on labeled data for training and have difficulty in utilizing unlabeled time series information, and lack deep fusion and spatiotemporal correlation modeling of multimodal signals of ECG, activity and environment, this solution first performs hierarchical self-supervised comparative learning on multimodal unlabeled data to obtain representations, then uses a dynamic graph convolutional network to combine ECG, movement and environmental features to capture the spatiotemporal correlation between physiology and the environment, and finally achieves high-accuracy cardiac anomaly detection for the elderly through small-sample fine-tuning combined with generative adversarial networks to amplify samples.

[0026] (3) In view of the problem that the traditional elderly heart abnormality detection system only relies on a single-layer model for diagnosis, it is difficult to take into account both real-time response at the edge and high-precision classification at the cloud, and it also lacks dynamic risk assessment and early warning capabilities. This solution combines the initial screening at the edge with high-precision classification at the cloud through a two-layer reasoning mechanism, and generates short-term, medium-term, and long-term risk curves based on continuous-time Bayesian variational updates, thereby realizing real-time detection of heart abnormalities in the elderly. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 A schematic diagram of the deep learning-based elderly cardiac abnormality detection auxiliary system provided by the present invention;

[0028] Figure 2 Schematic diagram of the edge preprocessing module;

[0029] Figure 3 Schematic diagram of the multimodal spatiotemporal representation module;

[0030] Figure 4 Schematic diagram of the intelligent detection module.

[0031] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION

[0032] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0033] In the description of the present invention, it should be understood that terms such as "upper", "lower", "front", "back", "left", "right", "top", "bottom", "inside" and "outside" indicating directions or positional relationships are based on the directions or positional relationships shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operated in a specific direction. Therefore, they should not be understood as limiting the present invention.

[0034] Example 1, see Figure 1 The present invention provides an auxiliary system for detecting cardiac abnormalities in the elderly based on deep learning, which includes a data acquisition module, an edge preprocessing module, a multimodal spatiotemporal representation module and an intelligent detection module;

[0035] The data acquisition module collects the elderly person's historical electrocardiogram signals, activity intensity, environmental data, and heart status data; and sends the data to the edge preprocessing module, the multimodal spatiotemporal representation module, and the intelligent detection module;

[0036] The edge preprocessing module receives the data sent by the data acquisition module, performs real-time denoising and lightweight classification on the ECG signal using a one-dimensional convolutional neural network with variable bit width quantization, and simultaneously generates features through time domain and frequency domain attention to obtain suspected abnormal data, and sends the data to the multimodal spatiotemporal representation module and the intelligent detection module;

[0037] The multimodal spatiotemporal representation module receives data sent by the data acquisition module and the edge preprocessing module, first performs self-supervised comparative learning on unlabeled ECG, activity, and environmental multimodal time series data, and obtains physiological signal representation by constructing positive and negative sample pairs to maximize the embedding similarity of the same type and minimize the similarity of different types. It then constructs a dynamic graph based on multimodal features, uses a multi-layer perceptron to fuse ECG, motion, and environmental vectors to generate a time-varying adjacency matrix, and then iteratively updates the node representation through normalized graph convolution to capture the spatiotemporal correlation between individual physiology and the environment. Finally, based on the global model, it uses labeled samples to perform gradient fine-tuning, combines with a generative adversarial network to amplify the samples, and sends the data to the intelligent detection module.

[0038] The intelligent detection module receives data sent by the data acquisition module, edge preprocessing module and multimodal spatiotemporal representation module, and classifies and probabilistically predicts suspected abnormal ECG segments through a two-layer reasoning mechanism. Subsequently, based on the output probability vector, the continuous-time Bayesian update formula is used to dynamically calculate the prior and posterior risks of abnormal cardiac events, and the integrals are accumulated to obtain short-term, medium-term and long-term risk curves, thereby realizing the detection of cardiac abnormalities.

[0039] Example 2, see Figure 1 This embodiment is based on the above embodiment. The data acquisition module collects the elderly person's historical electrocardiogram signals, activity intensity, environmental data and heart state data; the electrocardiogram signals specifically include the elderly person's electrocardiogram waveform, sampling rate and timestamp; the activity intensity specifically refers to the elderly person's three-axis acceleration, sampling rate and timestamp collected by the accelerometer; the environmental data specifically includes the environment's temperature, relative humidity, air pressure and timestamp; the heart state data refers to the elderly person's actual heart state, including normal state and abnormal state, as label data of the data sample.

[0040] Example 3, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. The edge preprocessing module uses a one-dimensional convolutional neural network with variable bit width quantization to complete denoising and preliminary classification. At the same time, dual attention is designed, specifically including the following units:

[0041] Denoising unit, in candidate bit width The above quantized convolution denoising is performed separately and the mean square error and storage cost are weighted summed. The optimal bit width that minimizes the weighted sum is automatically selected. The original signal is then quantized with the optimal bit width to obtain the denoised signal, which is expressed as follows:

[0042] ;

[0043] in, represents the original time domain signal of the electrocardiogram, represents the quantized electrocardiogram signal, Table optimal bit width, Indicates that the signal According to the optimal bit width To quantify, Indicates that the bit width used is The parameters of the signal Perform one-dimensional convolution denoising, represents the calculation of mean square error, Indicates bit width The corresponding storage space size, Represents the regularization coefficient that balances error and cost, and its value range is ;

[0044] The dual attention unit is introduced, and scaled dot product attention is introduced in the time domain and frequency domain respectively. The quantized signal is first mapped into query, key, and value tensors. Then, the dot product of the query and key is divided by the dimension scaling factor and normalized to generate the attention weight matrix in the time domain and frequency domain. The attention weight matrix is then multiplied with the value tensor to obtain the enhanced time domain and frequency domain feature representation, which are expressed as follows:

[0045] ;

[0046] in, 、 、 They represent the time domain query, key, and value tensors obtained by mapping the quantized signal through a small convolution. represents the normalized exponential function, Represents the temporal attention dimension scaling factor, with a value range of , represents the time domain attention weight matrix, represents the time domain weighted feature representation, 、 、 By performing short-time Fourier transform on the quantized signal, we can get the frequency domain query, key, and value tensors. represents the transpose symbol, Represents the frequency domain attention dimension scaling factor, with a value range of , represents the frequency domain attention weight matrix, represents the frequency domain weighted feature representation;

[0047] The preliminary cardiac anomaly classification unit uses a one-dimensional convolutional neural network to perform lightweight classification on quantized time-domain features. It then uses the lightweight convolutional network to perform lightweight classification on the fused features, uploading only suspected abnormal segments to the cloud to reduce bandwidth. This is shown below:

[0048] ;

[0049] in, Indicates the splicing of time domain and frequency domain attention features, Represents a quantized one-dimensional convolutional network, including variable bit width convolutional layers and batch normalization, represents the fully connected layer, Represents the edge prediction category.

[0050] By performing the above operations, this solution addresses the problems of large model size, high power consumption, insufficient noise suppression and feature extraction capabilities, poor interpretability, and edge device inference performance bottlenecks caused by fixed bit-width networks in traditional elderly heart abnormality detection auxiliary systems. This solution automatically balances error and computational cost through a one-dimensional convolutional neural network with variable bit-width quantization to achieve efficient real-time denoising; dual attention multi-scale noise suppression is introduced in the time and frequency domains to enhance the extraction capability of weak pathological waveforms; and based on the fused lightweight features, only suspected abnormal fragments are uploaded to the cloud, extending battery life and improving the accuracy and interpretability of early detection of heart abnormalities in the elderly.

[0051] Example 4, see Figure 1 and Figure 3 This embodiment is based on the above embodiment. The multimodal spatiotemporal representation module first performs hierarchical self-supervised contrastive learning on unlabeled ECG time series data to obtain a representation, wherein the unlabeled ECG time series data refers to ECG time series data that is not labeled with heart status; then a dynamic graph based on multimodal features is constructed to capture the spatiotemporal correlation between individual physiological signals and the environment; finally, gradient fine-tuning is performed using labeled data, wherein the labeled data refers to ECG time series data labeled with heart status; and a generative adversarial network is used to amplify samples, specifically including the following units:

[0052] The self-supervised contrastive learning unit first encodes the ECG, motion, and environmental signals into embedding vectors, constructs positive sample pairs and negative sample pairs of different samples under the same enhanced view at the same moment, then introduces cosine similarity and temperature coefficient to adjust the contrast loss. Through the normalized exponential function, the similarity of positive sample pairs is maximized and the similarity of negative sample pairs is minimized, thereby learning multimodal representations, which can be expressed as follows:

[0053] ;

[0054] in, represents the self-supervised contrastive learning loss, Represents a set of positive sample pairs, i, j, k represent the index of the sample vector, 、 and represents the multimodal embedding vector obtained by encoding, 、 belong to the same positive sample pair, represents all negative sample pairs, represents an exponential function with a natural constant as the base, represents the cosine similarity function, Represents the contrastive learning parameter, with a value range of , N represents the total number of samples;

[0055] The graph convolution node update unit constructs a spatiotemporal adjacency matrix based on multimodal features, dynamically updates the graph convolution network node representation matrix, and uses a multi-layer perceptron to fuse ECG, activity, and environmental features to generate a dynamic graph adjacency matrix. The node representation matrix is then iteratively updated through a normalized graph convolution layer, as shown below:

[0056] ;

[0057] in, represents the ECG feature vector at time t, represents the activity feature vector at time t, represents the environmental feature vector at time t, Represents vector concatenation operation, represents the multilayer perceptron used to generate the adjacency matrix at time t, represents the dynamic graph adjacency matrix at time t, represents the linear rectification function, p and q represent the row index and column index of the dynamic graph adjacency matrix, Represents the element in the pth row and qth column of the dynamic graph adjacency matrix, represents the degree matrix at time t, The elements representing the main diagonal are The diagonal matrix of represents the node representation matrix of the l+1th layer at time t, represents the node representation matrix of the lth layer at time t, represents the learnable weight matrix of layer l;

[0058] The few-shot fine-tuning and amplification unit performs gradient fine-tuning on the global model using labeled data and amplifies samples using a generative adversarial network, as shown below:

[0059] ;

[0060] in, represents the global model parameters after fine-tuning, represents the global model parameters before fine-tuning, Indicates fine-tuning learning rate, the value range is , represents the global model loss, represents the gradient of the global model loss with respect to the parameters, represents a set of labeled samples, represents the cross entropy loss, Represents a global model composed of two layers of multilayer perceptrons. The first layer of the multilayer perceptron contains 64 The second layer of activated neurons contains 32 neurons with linear rectification function activation, and passes through a fully connected layer before output, and finally activated by a normalized exponential function to generate the final classification probability; Indicates using the current global model parameters For labeled data The model output obtained by forward reasoning, Represents the loss of the generative adversarial network, which discriminates the samples generated by the generator G through the discriminator D. Represents the hyperparameter that balances fine-tuning loss and adversarial loss, and its value range is .

[0061] By performing the above operations, we can address the problems in traditional elderly heart anomaly detection systems that often rely only on labeled data for training, have difficulty utilizing unlabeled time series information, and lack deep fusion and spatiotemporal correlation modeling of multimodal signals of ECG, activity, and environment. This solution first performs hierarchical self-supervised comparative learning on multimodal unlabeled data to obtain representations, then uses a dynamic graph convolutional network to combine ECG, motion, and environmental features to capture the spatiotemporal correlation between physiology and the environment, and finally achieves high-accuracy heart anomaly detection in the elderly through few-sample fine-tuning combined with generative adversarial networks to amplify samples.

[0062] Example 5, see Figure 1 and Figure 4 This embodiment is based on the above embodiment, and the intelligent detection module specifically includes the following units:

[0063] The two-layer reasoning output unit initially screens the edge to determine the suspected abnormal data segments The data is sent to the cloud, which then provides the classification and probability distribution, as shown below:

[0064] ;

[0065] in, Represents a collection of suspected abnormal data fragments, A label indicating that the heart state is normal, represents the cloud's predicted probability vector for abnormal cardiac events, and Represents the weight parameters and bias parameters of the global model after fine-tuning;

[0066] The Bayesian variational risk scoring unit uses the continuous-time Bayesian update formula to dynamically calculate the short-term, medium-term, and long-term risks of cardiac abnormalities and generate a risk curve that changes with time t, as shown below:

[0067] ;

[0068] in, represents the prior risk probability at time t, represents the posterior risk probability at time t, represents the multimodal input segment at time t, represents the conditional probability, Indicates that an abnormal cardiac event has occurred In the case of The likelihood function of Indicates that no cardiac abnormalities occur In the case of The likelihood function of Indicates the estimated window length, with values of 1 hour, 4 hours, and 24 hours; From t to The cumulative risk score of

[0069] The real-time detection unit obtains the multimodal time series data of the collected ECG signals, activity intensity, environmental data and heart status data of all elderly people, and uses the Bayesian variational risk scoring unit to calculate the cumulative risk score at each time point on the 1-hour, 4-hour and 24-hour windows respectively. After pairing with the label, a threshold set is created, specifically a threshold set starting from 0 and increasing to 1 in steps of 0.05; the candidate threshold set is traversed, the sensitivity and specificity are calculated at each threshold to draw the ROC curve, and the threshold that maximizes the Youden index is selected as the optimal threshold for each window; during real-time detection, the unit receives the pre-processed ECG, motion and environmental data in real time, calls the continuous-time Bayesian update formula to iteratively calculate the posterior risk probability and accumulates the integral in each time window. The higher the integral, the higher the risk of heart abnormality; when the risk integral of the window exceeds the optimal threshold, a heart abnormality alarm is triggered, and the user's heart state is abnormal; otherwise, the user's heart state is normal.

[0070] By performing the above operations, the traditional elderly heart abnormality detection system only relies on a single-layer model for diagnosis, which makes it difficult to balance real-time response on the edge and high-precision classification on the cloud, and lacks dynamic risk assessment and early warning capabilities. This solution combines edge initial screening with high-precision classification on the cloud through a two-layer inference mechanism, and generates short-term, medium-term, and long-term risk curves based on continuous-time Bayesian variational updates, thereby realizing real-time detection of heart abnormalities in the elderly.

[0071] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0072] While the embodiments of the present invention have been shown and described, it will be apparent to those skilled in the art that various changes, modifications, substitutions, and alterations can be made to the embodiments without departing from the principles and spirit of the invention.

[0073] The present invention and its embodiments are described above. This description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this and, without departing from the purpose of the present invention, designs structures and embodiments similar to this technical solution without inventiveness, they shall fall within the scope of protection of the present invention.

Claims

1. A deep learning-based auxiliary system for detecting cardiac abnormalities in the elderly, characterized by: It includes data acquisition module, edge pre-processing module, multimodal spatiotemporal representation module and intelligent detection module; The data acquisition module collects the elderly person's historical electrocardiogram signals, activity intensity, environmental data and heart status data; The edge preprocessing module uses a one-dimensional convolutional neural network with variable bit width quantization to perform real-time denoising and lightweight classification on the ECG signal, while generating features through time domain and frequency domain attention to obtain suspected abnormal data; The multimodal spatiotemporal representation module first performs self-supervised comparative learning on unlabeled ECG, activity, and environmental multimodal time series data. By constructing positive and negative sample pairs, it maximizes the similarity of similar embeddings and minimizes the similarity of different embeddings, thereby obtaining physiological signal representations. It then constructs a dynamic graph based on multimodal features, uses a multi-layer perceptron to fuse ECG, motion, and environmental vectors to generate a time-varying adjacency matrix, and then iteratively updates node representations through normalized graph convolution to capture the spatiotemporal correlation between individual physiology and the environment. Finally, based on the global model, it performs gradient fine-tuning using labeled samples and combines them with a generative adversarial network to amplify the samples. The intelligent detection module classifies and predicts the probability of suspected abnormal ECG segments through a two-layer reasoning mechanism; Subsequently, based on the output probability vector, the continuous-time Bayesian update formula is used to dynamically calculate the prior and posterior risks of cardiac abnormality events, and the integrals are accumulated to obtain short-term, medium-term and long-term risk curves, thereby realizing the detection of cardiac abnormalities.

2. The deep learning-based elderly cardiac anomaly detection auxiliary system according to claim 1 is characterized by: The edge preprocessing module uses a one-dimensional convolutional neural network with variable bit width quantization to perform denoising and preliminary classification. It also designs dual attention, specifically including the following units: The denoising unit performs quantized convolution denoising on candidate bit widths and takes the weighted sum of the mean square error and storage cost. It automatically selects the optimal bit width that minimizes the weighted sum and then quantizes the original signal using the optimal bit width to obtain the denoised signal. A dual attention unit is introduced, which uses scaled dot product attention in the time domain and frequency domain respectively. The quantized signal is first mapped into query, key, and value tensors. Then, the dot product of the query and key is divided by the dimension scaling factor and normalized to generate the attention weight matrices in the time domain and frequency domain. These are then multiplied with the value tensor to obtain enhanced time domain and frequency domain feature representations. The preliminary cardiac abnormality classification unit uses a one-dimensional convolutional neural network to perform lightweight classification on the quantized time domain features. It also uses the lightweight convolutional network to perform lightweight classification on the fused features, and only uploads suspected abnormal fragments to the cloud.

3. The deep learning-based elderly cardiac anomaly detection auxiliary system according to claim 1 is characterized by: The multimodal spatiotemporal representation module first performs hierarchical self-supervised contrastive learning on unlabeled ECG time series data to obtain a representation. It then constructs a dynamic graph based on multimodal features to capture the spatiotemporal correlation between individual physiological signals and the environment. Finally, it performs gradient fine-tuning using labeled data and uses a generative adversarial network to amplify scarce cardiac abnormality samples. Specifically, it includes the following units: The self-supervised contrastive learning unit first encodes ECG, motion, and environmental signals into embedding vectors, constructs positive pairs of samples from the same enhanced view and negative pairs of samples from different views at the same moment, then introduces cosine similarity and temperature coefficient to adjust the contrastive loss. Using a normalized exponential function, it maximizes the similarity of positive pairs and minimizes the similarity of negative pairs, thereby learning multimodal representations. The graph convolution node update unit constructs a spatiotemporal adjacency matrix based on multimodal features, dynamically updates the graph convolution network node representation matrix, uses a multi-layer perceptron to fuse ECG, activity, and environmental features to generate a dynamic graph adjacency matrix, and then iteratively updates the node representation matrix through a normalized graph convolution layer; The few-sample fine-tuning and amplification unit performs gradient fine-tuning on the global model through labeled data, and uses a generative adversarial network to amplify scarce abnormal samples.

4. The deep learning-based elderly cardiac anomaly detection auxiliary system according to claim 1 is characterized by: The intelligent detection module specifically includes the following units: The two-layer reasoning output unit sends data segments that are initially identified as abnormal by the edge to the cloud, which then provides classification and probability distribution. The Bayesian variational risk scoring unit uses a continuous-time Bayesian update formula to dynamically calculate the short-term, medium-term, and long-term risks of cardiac abnormalities and generate a risk curve that changes over time; The real-time detection unit uses the Bayesian variational risk scoring unit to calculate the cumulative risk score of all samples in the short-term, medium-term, and long-term windows and pair them with the labels. It then generates a threshold set, traverses and calculates the sensitivity and specificity under each threshold to draw the ROC curve, and selects the threshold that maximizes the Youden index as the optimal threshold for each window; Finally, the user data to be tested is collected and preprocessed. The preprocessed user data is received in real time, the posterior risk is iteratively updated, and the integral is accumulated within the window. When the integral of the window exceeds the optimal threshold, the user's heart condition is determined to be abnormal and an alarm is issued. Otherwise, it is judged to be in normal state.

5. The deep learning-based elderly cardiac anomaly detection auxiliary system according to claim 1 is characterized by: The data acquisition module collects the elderly's historical electrocardiogram (ECG) signals, activity intensity, environmental data, and heart status data; the ECG signals specifically include the elderly's ECG waveform, sampling rate, and timestamp; the activity intensity specifically refers to the elderly's three-axis acceleration, sampling rate, and timestamp collected by the accelerometer; the environmental data specifically includes the environment's temperature, relative humidity, air pressure, and timestamp; the heart status data includes normal and abnormal states, and the heart status data is set as label data.

Images