Deep learning-based elderly heart abnormality detection assistance system

Abstract

The application discloses an old person heart abnormality detection auxiliary system based on deep learning, which comprises a data acquisition module, an edge preprocessing module, a multi-modal space-time representation module and an intelligent detection module. The application relates to the technical field of heart health monitoring, in particular to an old person heart abnormality detection auxiliary system based on deep learning. The one-dimensional convolutional neural network with variable bit width quantization is used to realize efficient denoising. Double attention is introduced in the time domain and the frequency domain, and multi-scale noise suppression is used to enhance the extraction ability of weak pathological waveforms. Hierarchical self-supervised contrast learning is performed to obtain general representation, and then a dynamic atlas convolution network is used to capture the space-time correlation between physiology and environment, thereby improving the accuracy of heart abnormality detection. The edge end preliminary screening and high-precision classification are combined through a double-layer reasoning mechanism, and a risk curve is generated based on continuous time Bayesian variation update, so that real-time detection of heart abnormalities of the old people is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of heart health monitoring, in particular to an old person heart abnormality detection auxiliary system based on deep learning. BACKGROUND

[0002] With the continuous intensification of global population aging, the incidence and mortality of cardiovascular diseases in the elderly continue to rise, and early, continuous and accurate heart abnormality monitoring has become the key to protecting the health of the elderly. However, existing electrocardiogram monitoring mostly relies on large medical equipment, which not only has large equipment size and high power consumption, but also is inconvenient to wear for a long time at home or in community environment, and the model has limited ability in noise suppression and extraction of weak pathological waveform, lacks deep explanation of abnormal detection results, and cannot meet the strict requirements of edge devices on real-time performance, power consumption and bandwidth.

[0003] In addition, the traditional system mostly only uses single electrocardiogram signal, which is difficult to fully utilize unlabeled multi-modal time series data, and lacks deep fusion and spatio-temporal correlation modeling of electrocardiogram, motion and environmental information, resulting in poor model generalization ability when sample is scarce and individual difference is large. At the same time, single-layer diagnosis process cannot balance the rapid early warning of edge devices, nor can it realize high-precision classification and dynamic risk assessment in the cloud, so that there are significant blind spots and delays in the prediction and continuous monitoring of heart abnormalities in the elderly. SUMMARY

[0004] In view of the above, in order to overcome the defects of the prior art, the present application provides an old person heart abnormality detection auxiliary system based on deep learning, aiming at the problems of large model volume, high power consumption, insufficient noise suppression and feature extraction capability, poor interpretability and fixed bit width network leading to edge device inference performance bottleneck in the traditional old person heart abnormality detection auxiliary system, the present scheme automatically balances between error and calculation cost through a variable bit width quantization one-dimensional convolutional neural network, realizes efficient real-time denoising, introduces dual attention multi-scale noise suppression in time domain and frequency domain, enhances the extraction capability of weak pathological waveform, and uploads only the suspected abnormal segment to the cloud based on the fused lightweight features; aiming at the problems that the traditional old person heart abnormality detection system often can only rely on labeled data for training, is difficult to utilize unlabeled time series information, and lacks deep fusion of electrocardiogram, activity and environment multi-modal signals and spatio-temporal correlation modeling, the present scheme first performs hierarchical self-supervised contrast learning on multi-modal unlabeled data to obtain representation, then utilizes a dynamic graph convolutional network to combine electrocardiogram, motion and environmental features to capture the spatio-temporal correlation of physiology and environment, and finally performs few-shot fine-tuning and combines a generative adversarial network to augment samples; aiming at the problems that the traditional old person heart abnormality detection system only relies on a single layer model for diagnosis, it is difficult to balance edge real-time response and cloud high-precision classification, and lacks dynamic risk assessment and early warning capability, the present scheme combines edge primary screening and cloud high-precision classification through a double-layer inference mechanism, and generates short-term, medium-term and long-term risk curves based on continuous time Bayesian variational update, realizing real-time detection of old person heart abnormality.

[0005] The technical scheme adopted by the present application is as follows: The old person heart abnormality detection auxiliary system based on deep learning provided by the present application comprises a data acquisition module, an edge preprocessing module, a multi-modal spatio-temporal representation module and an intelligent detection module.

[0006] The data acquisition module acquires old person historical electrocardiogram signals, activity intensity, environmental data and heart state data.

[0007] The edge preprocessing module utilizes a variable bit width quantization one-dimensional convolutional neural network to perform real-time denoising and lightweight classification on the electrocardiogram signals, and simultaneously generates features through time domain and frequency domain attention to obtain suspected abnormal data.

[0008] The multi-modal spatio-temporal representation module first performs self-supervised contrastive learning on unannotated electrocardiogram, activity and environment multi-modal time series data, maximizes the similarity of same class embedding and minimizes the similarity of different classes by constructing positive and negative sample pairs, thereby obtaining physiological signal representation; then a dynamic graph is constructed based on multi-modal features, a multilayer perceptron is used to fuse electrocardiogram, motion and environment vectors to generate a time-varying adjacency matrix, and then a normalized graph convolution is iteratively updated to capture the spatio-temporal correlation between individual physiology and environment; finally, the global model is fine-tuned based on the annotated samples, and a generative adversarial network is used to expand the samples;

[0009] The intelligent detection module classifies and probabilistically predicts the suspected abnormal electrocardiogram segment through a double-layer inference mechanism; then, based on the output probability vector, the prior and posterior risks of the cardiac abnormality event are dynamically calculated using the continuous time Bayesian update formula, and the integral is accumulated to obtain the short-term, medium-term and long-term risk curves, thereby realizing the detection of cardiac abnormalities.

[0010] Further, the data acquisition module acquires the historical electrocardiogram signals, activity intensity, environment data and cardiac state data of the elderly; the electrocardiogram signal specifically includes the electrocardiogram waveform, sampling rate and timestamp of the elderly; the activity intensity specifically refers to the three-axis acceleration, sampling rate and timestamp of the elderly acquired by the accelerometer; the environment data specifically includes the temperature, relative humidity, air pressure and timestamp of the environment; the cardiac state data includes normal state and abnormal state, and the cardiac state data is set as label data.

[0011] Further, the edge preprocessing module uses a variable bit width quantization one-dimensional convolutional neural network to complete denoising and preliminary classification; at the same time, a double attention is designed, specifically including the following units:

[0012] The denoising unit respectively performs quantization convolution denoising on the candidate bit width and weights the sum of the mean square error and the storage cost, automatically selects the optimal bit width that minimizes the weighted sum, and then quantizes the original signal with the optimal bit width to obtain the denoised signal;

[0013] The double attention unit is introduced, which introduces scaled dot-product attention in the time domain and the frequency domain. The quantized signal is first mapped to query, key and value tensors, then the dot product of the query and the key is divided by the dimension scale factor and normalized to generate the attention weight matrix in the time domain and the frequency domain, and then the value tensor is multiplied respectively 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, and uses a lightweight convolutional network to perform lightweight classification on the fused features, and only uploads the suspected abnormal segments to the cloud.

[0015] Further, the multi-modal spatio-temporal representation module first performs hierarchical self-supervised contrastive learning, performs multi-modal contrastive learning on unannotated electrocardiogram time series data to obtain a representation, then constructs a dynamic graph based on multi-modal features to capture the spatio-temporal correlation between individual physiological signals and the environment, and finally performs gradient fine-tuning on labeled data and expands samples using a generative adversarial network, specifically including the following units:

[0016] The self-supervised contrastive learning unit first encodes electrocardiogram, motion and environmental signals into embedding vectors, constructs positive sample pairs and negative sample pairs under the same time and enhanced view, then introduces cosine similarity and a temperature coefficient to adjust the contrastive loss, maximizes the similarity of positive sample pairs and minimizes the similarity of negative sample pairs through a normalized exponential function, thereby learning multi-modal representations.

[0017] The graph convolution node updating unit constructs a spatio-temporal adjacency matrix based on multi-modal features, dynamically updates the node representation matrix of the graph convolution network, and generates a dynamic graph adjacency matrix by fusing electrocardiogram, activity and environmental features using a multi-layer perceptron, and then iteratively updates the node representation matrix through a normalized graph convolution layer.

[0018] The few-shot fine-tuning and expansion unit performs gradient fine-tuning on labeled data based on the global model and expands samples using a generative adversarial network.

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

[0020] The double-layer reasoning output unit sends data segments suspected to be abnormal by edge preliminary screening to the cloud, and the cloud gives a classification and a probability distribution.

[0021] The Bayesian variational risk scoring unit dynamically calculates short-term, medium-term and long-term cardiac abnormal event risks using a continuous time Bayesian update formula, and generates a risk curve that changes over time.

[0022] The real-time detection unit calculates the cumulative risk scores of all samples in the short-term, medium-term and long-term windows using the Bayesian variational risk scoring unit and pairs them with labels, then generates a threshold set, calculates the sensitivity and specificity at each threshold to draw a 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 integral is accumulated in the window, when the integral of the window exceeds the optimal threshold, the user's cardiac state is determined to be abnormal and an alarm is given. Otherwise, it is determined to be normal.

[0023] The above scheme has the following beneficial effects:

[0024] (1) In view of the problems of large model volume, high power consumption, insufficient noise suppression and feature extraction capability, poor interpretability and fixed bit width network leading to edge device inference performance bottleneck in the traditional auxiliary system for detecting heart abnormalities of the elderly, the scheme realizes efficient and real-time noise reduction by automatically balancing between error and calculation cost through a variable bit width one-dimensional convolutional neural network; the dual attention multi-scale noise suppression is introduced in the time domain and the frequency domain to enhance the extraction capability of weak pathological waveforms; and only the suspected abnormal segment is uploaded to the cloud based on the fused lightweight features, prolonging the endurance and improving the accuracy and interpretability of early detection of heart abnormalities of the elderly.

[0025] (2) In view of the problems that the traditional heart abnormality detection system for the elderly can only rely on labeled data for training, is difficult to utilize unlabeled time series information, and lacks deep fusion of multi-modal signals such as electrocardiogram, activity and environment and spatio-temporal correlation modeling, the scheme first performs hierarchical self-supervised contrastive learning on multi-modal unlabeled data to obtain representation, then utilizes dynamic graph convolutional network to combine electrocardiogram, motion and environmental features to capture the spatio-temporal correlation of physiology and environment, and finally realizes high-accuracy heart abnormality detection for the elderly through few-shot fine-tuning and combining with a generative adversarial network to expand samples.

[0026] (3) In view of the problems that the traditional heart abnormality detection system for the elderly only relies on a single layer model for diagnosis, which is difficult to balance edge real-time response and cloud high-precision classification, and lacks dynamic risk assessment and early warning capability, the scheme combines edge preliminary screening and cloud high-precision classification through a double-layer inference mechanism, and generates short-term, medium-term and long-term risk curves based on continuous time Bayesian variational update, realizing real-time detection of heart abnormalities of the elderly. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 A schematic diagram of the auxiliary system for detecting heart abnormalities of the elderly based on deep learning is provided.

[0028] Figure 2 A schematic diagram of the edge preprocessing module is provided.

[0029] Figure 3 A schematic diagram of the multi-modal spatio-temporal representation module is provided.

[0030] Figure 4 A schematic diagram of the intelligent detection module is provided.

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

[0032] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0033] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

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

[0035] The data acquisition module collects historical electrocardiogram signals, activity intensity, environmental data, and cardiac status data of the elderly; and sends the data to the edge preprocessing module, the multimodal spatiotemporal characterization module, and the intelligent detection module.

[0036] The edge preprocessing module receives data sent by the data acquisition module, uses a variable bit-width quantized one-dimensional convolutional neural network to perform real-time denoising and lightweight classification of ECG signals, generates features through time-domain and frequency-domain attention, acquires 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 from the data acquisition module and the edge preprocessing module. First, it performs self-supervised contrastive 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 dissimilar ones, thereby obtaining physiological signal representations. Then, it constructs a dynamic graph based on multimodal features, using a multilayer perceptron to fuse ECG, motion, and environmental vectors to generate a time-varying adjacency matrix. Next, it iteratively updates node representations through normalized graph convolution to capture the spatiotemporal correlation between individual physiology and the environment. Finally, it fine-tunes the gradient using labeled samples on the global model, combines it 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, the edge preprocessing module and the multi-modal spatio-temporal representation module, classifies and probabilistically predicts the suspected abnormal electrocardio segment through a double-layer inference mechanism; then, based on the output probability vector, the prior and posterior risks of the cardiac abnormal event are dynamically calculated using the continuous time Bayesian update formula, and the integral is accumulated to obtain the short-term, medium-term and long-term risk curves, thereby realizing the detection of the cardiac abnormality.

[0039] Embodiment two, refer to Figure 1 This embodiment is based on the above embodiment, the data acquisition module acquires the historical electrocardio signals, activity intensity, environmental data and cardiac state data of the elderly; the electrocardio signals specifically include the electrocardio waveform, sampling rate and time stamp of the elderly; the activity intensity specifically refers to the three-axis acceleration, sampling rate and time stamp of the elderly collected by the accelerometer; the environmental data specifically includes the temperature, relative humidity, air pressure and time stamp of the environment; the cardiac state data refers to the actual cardiac state of the elderly, including normal state and abnormal state, as label data of the data sample.

[0040] Embodiment three, refer to 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, a double attention is designed, specifically including the following units:

[0041] The denoising unit quantizes and denoises the candidate bit width , respectively, and calculates the weighted sum of the mean square error and the storage cost, automatically selects the optimal bit width that minimizes the weighted sum, and then quantizes the original signal with the optimal bit width to obtain the denoised signal, which is represented as follows:

[0042] ;

[0043] Wherein, represents the original time domain signal of the electrocardiogram, represents the quantized electrocardiogram signal, represents the optimal bit width, represents the signal quantized according to the optimal bit width , and represents one-dimensional convolutional denoising of the signal using the parameter with bit width , and represents the calculation of the mean square error, represents the storage space size corresponding to the bit width , and represents the regularization coefficient for balancing the error and the cost, with a value range of ;

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

[0045] ;

[0046] wherein, , , td_query, td_key, and td_value represent the time domain query, key, and value tensors mapped by a small convolution on the quantized signal, denotes a normalized exponential function, denotes a time domain attention dimension scale factor, and the value range is , denotes a time domain attention weight matrix, denotes a time domain weighted feature representation, , , The frequency domain query, key, and value tensors are mapped by performing a short-time Fourier transform on the quantized signal, denotes a transpose symbol, denotes a frequency domain attention dimension scale factor, and the value range is , denotes a frequency domain attention weight matrix, denotes a frequency domain weighted feature representation;

[0047] The preliminary cardiac abnormality classification unit uses a one-dimensional convolutional neural network to make a lightweight classification on the quantized time domain features. The fusion features are classified by a lightweight convolutional network, and only the suspected abnormal segments are uploaded to the cloud to reduce the bandwidth. It is represented as follows:

[0048] ;

[0049] wherein, denotes the concatenation of the time domain and frequency domain attention features, denotes a one-dimensional convolutional network after quantization, which includes a variable bit width convolution layer and batch normalization, denotes a fully connected layer, denotes an edge prediction category.

[0050] By performing the above operation, in view of the problems of large model volume, high power consumption, insufficient noise suppression and feature extraction capability, poor interpretability and fixed bit width network leading to edge device inference performance bottleneck in traditional elderly heart abnormality detection auxiliary system, the scheme automatically balances between error and calculation cost through variable bit width quantization one-dimensional convolutional neural network, realizes efficient real-time denoising; dual attention multi-scale noise suppression is introduced in time domain and frequency domain to enhance the extraction capability of weak pathological waveform; and only the suspected abnormal segment is uploaded to the cloud based on the fused lightweight features, prolonging the endurance and improving the accuracy and interpretability of early detection of elderly heart abnormalities.

[0051] In an embodiment, the multi-modal spatio-temporal representation module is trained based on a self-supervised contrastive learning method. Figure 1 and Figure 3 The embodiment is based on the above-mentioned embodiment. The multi-modal spatio-temporal representation module first performs hierarchical self-supervised contrastive learning, performs multi-modal contrastive learning on unannotated electrocardiogram time series data, and obtains a representation, wherein the unannotated electrocardiogram time series data refers to electrocardiogram time series data that is not annotated with a heart state; then a dynamic atlas based on multi-modal features is constructed to capture the spatio-temporal correlation between individual physiological signals and the environment, and finally gradient fine-tuning is performed on annotated data, wherein the annotated data refers to electrocardiogram time series data annotated with a heart state; and a generative adversarial network is used to augment samples, specifically including the following units:

[0052] The self-supervised contrastive learning unit first encodes the electrocardiogram, motion and environmental signals into embedding vectors, constructs positive sample pairs and negative sample pairs under the same time and enhanced view, then introduces cosine similarity and temperature coefficient to adjust the contrastive loss, maximizes the similarity of positive sample pairs and minimizes the similarity of negative sample pairs through the normalized exponential function, thereby learning multi-modal representation, which is represented as follows:

[0053] ;

[0054] wherein, represents the self-supervised contrastive learning loss, represents a set of positive sample pairs, i, j, k represent the indices of sample vectors, , and represent multi-modal embedding vectors obtained by encoding with an encoder, , belong to the same positive sample pair, represents all negative sample pairs, represents an exponential function with a natural constant as the base, represents a cosine similarity function, represents a contrastive learning parameter, and the value range is , N represents the total number of all samples;

[0055] The atlas convolution node updating unit constructs a spatio-temporal adjacency matrix based on multi-modal features, dynamically updates the node representation matrix of the graph convolution network, and generates a dynamic atlas adjacency matrix after fusing electrocardiogram, activity and environmental features through a multi-layer perceptron. Then, the node representation matrix is iteratively updated through a normalized graph convolution layer, which is represented as follows:

[0056] ;

[0057] wherein, represents the electrocardiogram feature vector at time t, represents the activity feature vector at time t, represents the environmental feature vector at time t, represents the vector splicing operation, represents the multi-layer perceptron used to generate the adjacency matrix at time t, represents the dynamic atlas adjacency matrix at time t, represents a linear rectification function, and p and q represent the row index and column index of the dynamic atlas adjacency matrix, represents the element of the dynamic atlas adjacency matrix at the pth row and qth column, represents the degree matrix at time t, represents a diagonal matrix with the elements of the main diagonal being , represents the node representation matrix at the l+1th layer at time t, represents the node representation matrix at the lth layer at time t, represents the learnable weight matrix of the lth layer;

[0058] The few-shot fine-tuning augmentation unit performs gradient fine-tuning on the labeled data based on the global model, and augments samples using a generative adversarial network, which is represented as follows:

[0059] ;

[0060] wherein, represents the global model parameters after fine-tuning, represents the global model parameters before fine-tuning, represents the fine-tuning learning rate, and the value range is , represents the global model loss, represents the gradient of the global model loss with respect to the parameters, represents the labeled sample set, represents the cross-entropy loss, represents the global model composed of two layers of multi-layer perceptron, and the first layer of the multi-layer perceptron contains 64 neurons with The activated neurons in the second layer include 32 neurons with linear rectification function activation, and a fully connected layer is used before output, and finally the normalized exponential function is activated to generate the final classification probability. represents the current global model parameters annotated data the model output obtained by forward inference, represents the generative adversarial network loss, which is generated by the discriminator D to distinguish the samples generated by the generator G, represents the hyperparameter for balancing the fine-tuning loss and the adversarial loss, and the value range is .

[0061] By performing the above operations, in view of the problems that in the traditional elderly heart abnormality detection system, it is often only possible to rely on labeled data for training, it is difficult to utilize unlabeled time series information, and there is a lack of deep fusion of electrocardiogram, activity and environment multi-modal signals and spatiotemporal correlation modeling, the present scheme first performs hierarchical self-supervised contrastive learning on multi-modal unlabeled data to obtain representation, then utilizes dynamic graph convolutional network to combine electrocardiogram, motion and environmental features to capture the spatiotemporal correlation of physiology and environment, and finally through few-shot fine-tuning and combining generative adversarial network to expand samples, high-accuracy heart abnormality detection for the elderly is realized.

[0062] Embodiment five, refer to Figure 1 and Figure 4 This embodiment is based on the above-mentioned embodiments, and the intelligent detection module specifically includes the following units:

[0063] The double-layer reasoning output unit judges the data segments suspected to be abnormal in the edge preliminary screening Send to the cloud, and the cloud gives the classification and probability distribution, represented as follows:

[0064] ;

[0065] wherein, represents the set of suspected abnormal data segments, represents the label of the normal state of the heart state, represents the prediction probability vector of the cloud for the heart abnormality event, and represent the weight parameters and bias parameters of the global model after fine-tuning;

[0066] The Bayesian variational risk scoring unit dynamically calculates the short-term, medium-term and long-term heart abnormality event risks by using the continuous time Bayesian update formula, and generates a risk curve that changes with time t, represented as follows:

[0067] ;

[0068] wherein, prior risk probability at time t, posterior risk probability at time t, multimodal input segment at time t, conditional probability, likelihood function of observing in the case of a cardiac abnormal event occurred, likelihood function of observing in the case of no cardiac abnormal event occurred, denotes the estimation window length, taking 1 hour, 4 hours, 24 hours; cumulative risk integral from t to ;

[0069] The real-time detection unit obtains the multimodal time series data of the collected ECG signals, activity intensity, environmental data and cardiac state data of all the elderly, and calculates the cumulative risk integral of each time point on the 1-hour, 4-hour and 24-hour window using the Bayesian variational risk scoring unit, respectively. After pairing with the label, a threshold set is created, which is to generate a threshold set starting from 0 and increasing by 0.05 to 1. The candidate threshold set is traversed, and the sensitivity and specificity are calculated under each threshold to draw the ROC curve, and the threshold that makes the Youden index maximum is selected as the optimal threshold of each window. In real-time detection, the unit receives the preprocessed ECG, motion and environmental data in real time, calls the continuous-time Bayesian update formula to iteratively calculate the posterior risk probability and integrate and accumulate within each time window. The higher the integral, the higher the risk of cardiac abnormalities. When the risk integral of the window exceeds the optimal threshold, a cardiac abnormality alarm is triggered, and the user's cardiac state is abnormal. Otherwise, the user's cardiac state is normal.

[0070] By performing the above operation, the traditional elderly cardiac abnormality detection system only relies on a single-layer model for diagnosis, which is difficult to balance edge real-time response and cloud high-precision classification, and lacks dynamic risk assessment and early warning capability. The scheme combines edge preliminary screening and cloud high-precision classification through a double-layer reasoning mechanism, and generates short-term, medium-term and long-term risk curves based on continuous-time Bayesian variational update, realizing real-time detection of cardiac abnormalities in the elderly.

[0071] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting; it is not intended to exclude myriad other embodiments of the present application that other presenters can develop. It is also possible, however, that only a single element can be present. Terms such as "a", "an", and "the" are not intended to refer to only a singular entity but include the general class of which a single element is only one species, unless otherwise indicated. Furthermore, the use of the terms "primary" and "secondary", "first" and "second", etc., designate different Stages in the process, and are not intended to otherwise limit the number of stages which can be employed. The terminology includes the words specifically noted above, derivatives thereof, and words of similar import. The designation of a component as "optional" indicates that the component is "optional" and can or can not be present or used in the practice of the application, but that when it is present or used, it can be used in varying amounts, configurations and / or types. The use of "including", "containing", "having" or "with" are not intended to be interdictive or to mean that the various embodiments of the present application solely comprise, have or are only accompanied by the enumerated

[0072] While the embodiments of the application have been shown and described, it is to be understood that various modifications, substitutions, combinations, staging, and variations can be made in the application without parting from the spirit and principles of the application.

[0073] The above description of the application and its embodiments is not restrictive, and the embodiments shown in the drawings are only one of the embodiments of the application, and the actual structure is not limited thereto. In general, if a person skilled in the art is inspired by it, without departing from the purpose of the application, without creative design, similar structure and embodiments of the technical solution can be developed, which should belong to the protection scope of the application.

Claims

1. A deep learning-based elderly heart abnormality detection assistance system, characterized by: The system comprises a data acquisition module, an edge preprocessing module, a multi-modal spatio-temporal representation module and an intelligent detection module. The data acquisition module acquires historical electrocardio signals, activity intensity, environmental data and heart state data of the elderly; The edge preprocessing module uses a variable-bit-width quantization one-dimensional convolutional neural network to perform real-time denoising and lightweight classification on the electrocardio signals, and generates features through time domain and frequency domain attention to obtain suspected abnormal data, and specifically comprises the following units: The denoising unit performs quantization convolution denoising on the candidate bit width respectively, and weights the mean square error and the storage cost to obtain the optimal bit width that minimizes the weighted sum, and then quantizes the original signal with the optimal bit width to obtain the denoised signal. The double attention unit introduces scaled dot-product attention in the time domain and the frequency domain, maps the quantized signal to query, key and value tensors, generates time domain and frequency domain attention weight matrices by taking the dot product of the query and the key, normalizing and dividing by the dimension scale factor, and then multiplies the value tensor to obtain enhanced time domain and frequency domain feature representations. The preliminary heart abnormality classification unit uses a one-dimensional convolutional neural network to perform lightweight classification on the quantized time domain features, and uses a lightweight convolutional network to perform lightweight classification on the fused features, and only uploads suspected abnormal segments to the cloud. The multi-modal spatio-temporal representation module first performs self-supervised contrastive learning on unannotated electrocardio, activity and environmental multi-modal time series data, maximizes the similarity of the same class embedding and minimizes the similarity of different classes by constructing positive and negative sample pairs, thereby obtaining physiological signal representation; then constructs a dynamic graph based on multi-modal features, generates a time-varying adjacency matrix using a multi-layer perceptron to fuse electrocardio, motion and environmental vectors, and iteratively updates node representation through normalized graph convolution to capture the spatio-temporal association between individual physiology and environment; finally, the global model is fine-tuned based on annotated samples, and the generated samples are expanded using a generative adversarial network. The intelligent detection module classifies and probabilistically predicts suspected abnormal electrocardio segments through a double-layer inference mechanism; subsequently, based on the output probability vector, the prior and posterior risks of the cardiac abnormality event are dynamically calculated using the continuous time Bayesian update formula, and the integral is accumulated to obtain short-term, medium-term and long-term risk curves, thereby realizing detection of cardiac abnormalities. 2.The deep learning-based elderly heart abnormality detection assistance system of claim 1, wherein: The multi-modal spatio-temporal representation module first performs hierarchical self-supervised contrastive learning, performs multi-modal contrastive learning on unannotated electrocardio time series data to obtain representation; then constructs a dynamic graph based on multi-modal features to capture the spatio-temporal association between individual physiological signals and environment, and finally fine-tunes the gradient based on labeled data and expands the scarce cardiac abnormality samples using a generative adversarial network, specifically comprising the following units: The self-supervised contrastive learning unit first encodes electrocardio, motion and environmental signals into embedding vectors, constructs positive sample pairs and negative sample pairs under the same time and enhanced view, then introduces cosine similarity and temperature coefficient to adjust the contrast loss, and maximizes the similarity of positive sample pairs and minimizes the similarity of negative sample pairs through a normalized exponential function, thereby learning multi-modal representation. The atlas convolution node updating unit constructs a space-time adjacency matrix based on multi-modal features, dynamically updates a node representation matrix of a graph convolution network, and generates a dynamic atlas adjacency matrix by fusing electrocardiogram, activity, and environmental features using a multi-layer perceptron, and then iteratively updates the node representation matrix through a normalized graph convolution layer. The few-shot fine-tuning and expansion unit performs gradient fine-tuning on labeled data based on a global model, and expands scarce abnormal samples using a generative adversarial network. 3.The deep learning-based elderly heart abnormality detection assistance system of claim 1, wherein: The intelligent detection module specifically includes the following units: The double-layer reasoning output unit sends data segments suspected to be abnormal to the cloud for edge preliminary screening, and the cloud gives a classification and a probability distribution. The Bayesian variational risk scoring unit dynamically calculates short-term, medium-term, and long-term cardiac abnormal event risks using a continuous-time Bayesian update formula, and generates a risk curve that changes over time. The real-time detection unit calculates the cumulative risk scores of all samples in the short-term, medium-term, and long-term windows using the Bayesian variational risk scoring unit, and pairs them with labels, then generates a threshold set, iteratively calculates the sensitivity and specificity of each threshold to draw a 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 integral is accumulated in the window. When the integral of the window exceeds the optimal threshold, the user's cardiac status is determined to be abnormal and an alarm is given. Otherwise, it is determined to be normal. 4.The deep learning-based elderly heart abnormality detection assistance system of claim 1, wherein: The data acquisition module acquires the historical electrocardiogram signals, activity intensity, environmental data, and cardiac status data of the elderly. The electrocardiogram signals specifically include the electrocardiogram waveform, sampling rate, and timestamp of the elderly. The activity intensity specifically refers to the three-axis acceleration, sampling rate, and timestamp of the elderly collected by an accelerometer. The environmental data specifically includes the temperature, relative humidity, air pressure, and timestamp of the environment. The cardiac status data includes normal and abnormal states, and the cardiac status data is set as label data.

Images