Electrocardiogram monitoring intelligent diagnosis system based on deep learning and multi-modal fusion technology

Abstract

The invention relates to the technical field of health monitoring, in particular to an electrocardiograph monitoring intelligent diagnosis system based on deep learning and multi-modal fusion technology, which is characterized in that a multi-modal synchronous sensing unit synchronously acquires ECG, PCG and PPG data streams with aligned timestamps, and an edge intelligent diagnosis unit operates a lightweight model with the parameter smaller than 1MB, so that the ECG monitoring intelligent diagnosis system is obtained. The cloud fusion analysis unit is used for dynamically fusing multi-source data by adopting a cross-modal attention mechanism to generate a ventricular fibrillation probability PVF, and the dynamic collaborative decision-making unit is used when R is equal to III or Cilt, and the cloud fusion analysis unit is used for dynamically fusing the multi-source data by adopting a cross-modal attention mechanism to generate a ventricular fibrillation probability PVF; deep diagnosis, PVFgt, is triggered within 5 seconds at 0.9; and when 0.95, a third-level audible and visual alarm is triggered, and the closed-loop optimization unit continuously optimizes the model by distilling and compressing the misdiagnosis sample. According to the intelligent diagnosis system for electrocardiograph monitoring, rapid edge diagnosis is achieved through a lightweight model, critical cases can respond rapidly through cloud cooperation, and the accuracy and timeliness of heart disease diagnosis are improved.

Description

Technical Field

[0001] This invention relates to the field of health monitoring technology, specifically to an intelligent diagnostic system for electrocardiogram monitoring based on deep learning and multimodal fusion technology. Background Technology

[0002] In the field of health monitoring technology, electrocardiogram (ECG) monitoring systems provide early warning and diagnostic support for cardiovascular diseases by real-time monitoring and analysis of ECG signals. For example, dynamic ECG monitoring devices collect data through wearable sensors and use threshold algorithms or shallow machine learning models for initial screening of arrhythmias, thereby achieving real-time monitoring and intelligent analysis of cardiac function and aiming to improve the early detection rate and diagnostic accuracy of cardiac emergencies.

[0003] Existing ECG monitoring and diagnosis methods have several problems. ECG interpretation relies entirely on physician experience, and subjective judgment can easily lead to misdiagnosis, especially when subtle changes in ST segment deviation cannot be accurately quantified and identified. Current AI systems only analyze single ECG signals, ignoring cross-validation between heart sounds and pulse waves, resulting in insufficient confidence in marginal diagnoses and frequent need for cloud intervention. There is a lack of second-level early warning mechanisms for high-risk cases such as ventricular fibrillation, making it impossible to achieve in-depth cloud-based diagnostic feedback and three-level audible and visual alarms within 5 seconds. The system lacks closed-loop optimization capabilities, and the lightweight model cannot be continuously updated based on misdiagnosis samples, causing diagnostic accuracy to decrease over time.

[0004] Therefore, in order to improve the real-time diagnostic accuracy and response speed of ECG monitoring, an intelligent diagnostic system for ECG monitoring based on deep learning and multimodal fusion technology is proposed. This system achieves millisecond-level edge analysis and second-level cloud-based deep diagnosis of critical cases through multimodal signal fusion and edge-cloud collaborative architecture. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent diagnostic system for electrocardiogram (ECG) monitoring based on deep learning and multimodal fusion technology, which solves the core defects of traditional ECG monitoring, such as high misdiagnosis rate due to reliance on human experience, insufficient reliability of single-modal analysis, and the inability of existing AI systems to dynamically optimize models.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an intelligent diagnostic system for electrocardiogram monitoring based on deep learning and multimodal fusion technology, comprising the following units:

[0007] The multimodal synchronous sensing unit synchronously acquires ECG, PCG, and PPG data streams and achieves timestamp alignment through a hardware clock.

[0008] The edge intelligent diagnostic unit connects to the multimodal synchronous sensing unit, runs a lightweight arrhythmia detection model with fewer than 1MB of parameters, and outputs the confidence level C in real time. eand risk level R∈{I, II, III};

[0009] The cloud-based fusion analysis unit receives the C output from the edge intelligent diagnostic unit. e R, through a cross-modal attention mechanism, dynamically weights and fuses ECG features, PCG and PPG data to generate a deep cardiac pathology report;

[0010] Dynamic collaborative decision-making unit, connected to cloud-based fusion analysis unit, when R=III or C e When the value is less than 0.9, a deep cloud-based diagnostic is triggered within 5 seconds and the results are fed back to the closed-loop optimization unit.

[0011] The closed-loop optimization unit receives misdiagnosed samples from cloud-based diagnostics, updates model parameters, and generates a lightweight version through distillation and compression. The lightweight model output by the closed-loop optimization unit is then deployed to the edge intelligent diagnostic unit to complete the optimization cycle.

[0012] Furthermore, the edge intelligent diagnostic unit performs:

[0013] The QRS localization submodule employs a one-dimensional trainable convolutional kernel W. q calculate

[0014] The risk marking submodule forces R=III when the ST segment offset ΔST>0.1mV;

[0015] in, denoted as QRS complex location; X represents the input ECG signal segment; ReLU is the corrected linear unit function; * represents a one-dimensional convolution operation.

[0016] Furthermore, the risk marking submodule calculates the risk level:

[0017]

[0018] Where ΔST is the voltage offset of the ST segment; p is the probability distribution vector of arrhythmia; max(p) is the highest probability confidence level; if is the conditional judgment keyword; mV is millivolt; otherwise is other cases.

[0019] Furthermore, the cloud-based fusion analysis unit includes a cross-modal alignment module and an attention fusion module;

[0020] The cross-modal alignment module aligns time series using a dynamic time warping algorithm;

[0021] The weight calculation of the attention fusion module satisfies w i =softmax(v T tanh(W s f i ));

[0022] Output the probability P of ventricular fibrillation VF ;

[0023] Among them, w i Here are the modal weights; softmax is the normalized exponential function; v T is the vector transpose; v is the attention weight vector; tanh is the hyperbolic tangent activation function; W s For shared weight matrix; f i It is a single-modal feature vector.

[0024] Furthermore, the dynamic collaborative decision-making unit enables a low-latency channel for cases with R=III, and when outputting the ventricular fibrillation probability P... VF A level 3 audible and visual alarm is triggered when the value is >0.95.

[0025] Furthermore, the closed-loop optimization unit generates a lightweight model through distillation and compression, satisfying a parameter size of less than 1MB.

[0026] This invention discloses an edge computing node for an intelligent diagnostic system for electrocardiogram monitoring, wherein the edge computing node performs:

[0027] The hardware accelerator performs QRS positioning operations;

[0028] The dynamic cache stores the original data of case R=III.

[0029] This invention also discloses a cloud-based analysis platform for an intelligent diagnostic system for electrocardiogram monitoring, wherein the cloud-based analysis platform performs:

[0030] Implement attention-weighted fusion in the attention fusion module of the cloud-based fusion analysis unit;

[0031] Perform model distillation compression in the closed-loop optimization unit.

[0032] Furthermore, the lightweight model has a latency of ≤50ms and a cloud response time of ≤5 seconds for critical cases.

[0033] Furthermore, the risk levels {I, II, III} correspond to:

[0034] Level I: Edge diagnostics are covered;

[0035] Level II: Requires cloud-based analysis;

[0036] Level III: Real-time cloud-based in-depth diagnostics and alerts.

[0037] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0038] 1. The multimodal synchronous sensing unit enhances diagnostic confidence by simultaneously acquiring ECG, PCG, and PPG data streams and using a hardware clock for timestamp alignment. This ensures temporal consistency of the multimodal signals. The aligned data stream output by this multimodal synchronous sensing unit enhances the pathological feature extraction capabilities of subsequent analysis units, overcomes the limitations of single-modal signal analysis, and improves the output confidence level C of the edge intelligent diagnostic unit. e Reliability.

[0039] 2. The edge intelligent diagnostic unit achieves millisecond-level real-time diagnosis, runs a lightweight arrhythmia detection model with less than 1MB of parameters, performs one-dimensional trainable convolution operations through the QRS localization submodule, and combines the risk labeling submodule to quantify the ST segment offset. The edge intelligent diagnostic unit completes real-time analysis on the local device, avoiding the offline processing delay of traditional ECG monitoring and ensuring immediate classification output of arrhythmia risk.

[0040] 3. The cloud-based fusion analysis unit enhances the accuracy of pathological identification, receiving the confidence level C output from the edge intelligent diagnostic unit. e The risk level R is determined, and the time series is aligned using a dynamic time warping algorithm executed through the cross-modal alignment module. The modal weight coefficients w are then calculated using the attention fusion module. i This cloud-based fusion analysis unit dynamically weights and fuses ECG features, PCG, and PPG data to generate a cardiomyopathy depth report, improving the probability of ventricular fibrillation. VF The accuracy of key indicators such as...

[0041] 4. Dynamic collaborative decision-making units ensure zero-delay response to high-risk cases, based on risk level R and confidence level C. e Automatically triggers a graded response mechanism when R=III or C e When the probability of ventricular fibrillation is <0.9, a cloud-based deep diagnostic process is forcibly initiated within 5 seconds; when the probability P0.9 is << 0.9, a cloud-based deep diagnostic process is forcibly initiated within 5 seconds; VF When the value is >0.95, a level 3 audible and visual alarm is immediately activated, eliminating the sluggish response of traditional systems to critical cases.

[0042] 5. The closed-loop optimization unit enables the continuous evolution of the diagnostic model. It receives misdiagnosed samples confirmed by cloud diagnosis, updates the lightweight model parameters through distillation and compression technology, and deploys the optimized model with less than 1MB of parameters to the edge intelligent diagnostic unit. The closed-loop mechanism enables the system to continuously adapt to new pathological features and maintain the accuracy of risk level R determination in the long term. Attached Figure Description

[0043] Figure 1 This is a diagram showing the overall architecture of the intelligent ECG monitoring and diagnostic system of the present invention.

[0044] Figure 2This is a flowchart of the edge intelligent diagnostic process of the present invention;

[0045] Figure 3 This is a schematic diagram of the closed-loop optimization of the present invention;

[0046] Figure 4 This is a cloud-based fusion analysis diagram of the present invention. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] Please see Figures 1 to 4 The present invention provides a technical solution: an intelligent diagnostic system for electrocardiogram monitoring based on deep learning and multimodal fusion technology, comprising the following units:

[0049] The multimodal synchronous sensing unit synchronously acquires ECG, PCG, and PPG data streams and achieves timestamp alignment through a hardware clock.

[0050] The edge intelligent diagnostic unit connects to the multimodal synchronous sensing unit, runs a lightweight arrhythmia detection model with fewer than 1MB of parameters, and outputs the confidence level C in real time. e and risk level R∈{I, II, III};

[0051] The cloud-based fusion analysis unit receives the C output from the edge intelligent diagnostic unit. e R, through a cross-modal attention mechanism, dynamically weights and fuses ECG features, PCG and PPG data to generate a deep cardiac pathology report;

[0052] Dynamic collaborative decision-making unit, connected to cloud-based fusion analysis unit, when R=III or C e When the value is less than 0.9, a deep cloud-based diagnostic is triggered within 5 seconds and the results are fed back to the closed-loop optimization unit.

[0053] The closed-loop optimization unit receives misdiagnosed samples from the cloud, updates the model parameters, and generates a lightweight version through distillation and compression. The lightweight model is then deployed to the edge intelligent diagnostic unit to complete the optimization cycle.

[0054] The multimodal synchronous sensing unit synchronously acquires ECG, PCG, and PPG data streams via a hardware clock, and outputs them to the edge intelligent diagnostic unit after timestamp alignment. The edge intelligent diagnostic unit runs a lightweight arrhythmia detection model with fewer than 1MB of parameters and outputs a confidence level C. e and risk levels R∈{I, II, III} are transmitted to the cloud-based fusion analysis unit; the cloud-based fusion analysis unit, based on the received C... e Using R data, a cross-modal attention mechanism is employed to dynamically weight and fuse ECG features, PCG, and PPG data to generate a cardiomyology depth report; the dynamic collaborative decision-making unit operates at R=III or C e When the value is less than 0.9, a cloud-based deep diagnostic process is initiated within 5 seconds, and the diagnostic results are fed back to the closed-loop optimization unit. The closed-loop optimization unit receives the misdiagnosed samples confirmed by the cloud, generates a lightweight model with less than 1MB of parameters through distillation and compression, and deploys it to the edge intelligent diagnostic unit.

[0055] The edge intelligent diagnostic unit executes:

[0056] The QRS localization submodule employs a one-dimensional trainable convolutional kernel W. q calculate

[0057] The risk marking submodule forces R=III when the ST segment offset ΔST>0.1mV;

[0058] in, The location of the QRS complex; X represents the segment of the input ECG signal; ReLU is the corrected linear unit function; * represents a one-dimensional convolution operation;

[0059] The QRS localization submodule calculates the QRS complex position of an ECG signal segment using a one-dimensional trainable convolutional neural network layer. The risk labeling submodule outputs a risk level R=III when it detects an ST segment offset exceeding 0.1 mV.

[0060] The risk labeling submodule calculates the risk level:

[0061]

[0062] Where ΔST is the voltage offset of the ST segment; p is the probability distribution vector of arrhythmia; max(p) is the highest probability confidence level; if is the conditional judgment keyword; mV is millivolt; otherwise is other cases;

[0063] The risk labeling submodule outputs the risk level according to preset conditions: when the ST segment offset exceeds 0.1 mV, it outputs R=III; when the highest confidence level of the arrhythmia probability distribution vector is in the range of 0.7 to 0.9, it outputs R=II; otherwise, it outputs R=I.

[0064] The cloud-based fusion analytics unit includes a cross-modal alignment module and an attention fusion module;

[0065] The cross-modal alignment module aligns time series using a dynamic time warping algorithm;

[0066] The weight calculation of the attention fusion module satisfies w i =softmax(v T tanh(W s f i ));

[0067] Output the probability P of ventricular fibrillation VF ;

[0068] Among them, w i Here are the modal weights; softmax is the normalized exponential function; v T is the vector transpose; v is the attention weight vector; tanh is the hyperbolic tangent activation function; W s For shared weight matrix; f i It is a single-modal feature vector;

[0069] The cross-modal alignment module aligns ECG, PCG, and PPG time series using a dynamic time warping algorithm, while the attention fusion module calculates modal weight coefficients using the hyperbolic tangent function and the normalized exponential function, outputting the ventricular fibrillation probability P. VF .

[0070] The dynamic collaborative decision-making unit activates a low-latency channel for cases with R=III, and outputs the ventricular fibrillation probability P. VF A level 3 audible and visual alarm is triggered when the value is >0.95.

[0071] The dynamic collaborative decision-making unit activates a low-latency transmission channel for cases with R=III disease level, when the probability of ventricular fibrillation P VF If the value exceeds 0.95, a level 3 audible and visual alarm command will be output.

[0072] The closed-loop optimization unit generates a lightweight model through distillation and compression, ensuring that the number of parameters is less than 1MB.

[0073] The closed-loop optimization unit compresses the model size using knowledge distillation technology, generating a lightweight model with fewer than 1MB of parameters.

[0074] This invention discloses an edge computing node for an intelligent diagnostic system for electrocardiogram monitoring. The edge computing node performs the following:

[0075] The hardware accelerator performs QRS positioning operations;

[0076] The dynamic cache stores the original data for cases R=III;

[0077] The hardware accelerator of the edge computing node performs convolutional neural network operations on the QRS localization submodule; the dynamic buffer stores the raw multimodal data stream of cases with risk level R=III.

[0078] This invention also discloses a cloud-based analysis platform for an intelligent diagnostic system for electrocardiogram monitoring. The cloud-based analysis platform executes:

[0079] Implement attention-weighted fusion in the attention fusion module of the cloud-based fusion analysis unit;

[0080] Perform model distillation and compression in the closed-loop optimization unit;

[0081] The cloud-based analytics platform performs modal weighting calculations for the attention fusion module and runs the model distillation and compression process for the closed-loop optimization unit.

[0082] Lightweight model latency ≤50ms, cloud response time for critical cases ≤5 seconds;

[0083] The edge intelligent diagnostic unit takes no more than 50 milliseconds to complete a single analysis; the dynamic collaborative decision-making unit responds to cloud-based in-depth diagnosis of R=III level cases in no more than 5 seconds.

[0084] Risk levels {I, II, III} correspond to:

[0085] Level I: Edge diagnostics are covered;

[0086] Level II: Requires cloud-based analysis;

[0087] Level III: Real-time cloud-based in-depth diagnostics and alerts;

[0088] Level I cases are handled independently by the edge intelligent diagnostic unit, Level II cases request cloud-based auxiliary analysis, and Level III cases trigger real-time cloud-based in-depth diagnosis and audible and visual alarms.

[0089] Example 1: Complete workflow for Level I risk cases

[0090] First, the multimodal synchronous sensing unit executes;

[0091] The system synchronously acquires patient ECG, PCG, and PPG data streams using a hardware clock. It timestamps the ECG, PCG, and PPG signals separately using a hardware clock to achieve timestamp alignment of the three signals and outputs time-aligned ECG, PCG, and PPG data streams to the edge intelligent diagnostic unit.

[0092] Next, the edge intelligent diagnostic unit executes;

[0093] A lightweight arrhythmia detection model with a runtime of 0.92 MB;

[0094] The QRS localization submodule performs a one-dimensional trainable convolution operation on the input ECG signal segment X to calculate... Output the QRS group position of the current segment.

[0095] The risk labeling submodule detects the ST segment offset ΔST = 0.05 mV and obtains the highest probability value of the arrhythmia probability distribution vector p with a confidence level of max(p) = 0.95.

[0096] According to the risk level determination formula:

[0097]

[0098] When ΔST ≤ 0.1 mV and max(p) ≥ 0.9, the output R = I;

[0099] Output risk level R = I and confidence level C e =0.96;

[0100] Then, the dynamic collaborative decision-making unit executes;

[0101] Receive the C output from the edge intelligent diagnostic unit e =0.96 and R=I;

[0102] The judgment does not satisfy R=III or C e Triggering condition <0.9;

[0103] Do not send diagnostic requests to the cloud-based fusion analytics unit;

[0104] Secondly, the cloud-based integrated analysis unit executes;

[0105] Since no trigger command was received, the cross-modal alignment module and attention fusion module remain in standby mode and do not generate a cardiac pathology depth report.

[0106] Finally, the closed-loop optimization unit is executed;

[0107] If no misdiagnosed sample is received from the cloud for confirmation, the model distillation, compression, and deployment operations will not be performed.

[0108] Example 2: Management of Level III Risk Critical Cases

[0109] First, the multimodal synchronous sensing unit executes;

[0110] The system synchronously acquires patient ECG, PCG, and PPG data streams using a hardware clock. It timestamps the ECG, PCG, and PPG signals separately using a hardware clock to achieve timestamp alignment of the three signals and outputs time-aligned ECG, PCG, and PPG data streams to the edge intelligent diagnostic unit.

[0111] Next, the edge intelligent diagnostic unit executes;

[0112] A lightweight arrhythmia detection model with a runtime of 0.98 MB;

[0113] The QRS localization submodule performs a one-dimensional trainable convolution operation on the input ECG signal segment X to calculate... Output the QRS group position of the current segment.

[0114] The risk marking submodule detects that the ST segment offset ΔST = 0.22 mV is greater than the 0.1 mV threshold.

[0115] According to the mandatory setting rule, when ΔST > 0.1 mV, the output R = III;

[0116] Output risk level R=III and confidence level C e =0.83;

[0117] Then, the dynamic collaborative decision-making unit executes;

[0118] Receive the C output from the edge intelligent diagnostic unit e =0.83 and R=III:

[0119] Activate the low-latency transmission channel immediately;

[0120] Send a deep diagnostic request to the cloud-based fusion analysis unit;

[0121] Start a 5-second response countdown, starting at time T0;

[0122] Secondly, the cloud-based integrated analysis unit executes;

[0123] After receiving the request:

[0124] The cross-modal alignment module aligns ECG, PCG, and PPG time series using a dynamic time warping algorithm.

[0125] Attention Fusion Module:

[0126] Calculate the modal weighting coefficient w i =softmax(v T tanh(W s fi ));

[0127] Dynamically weighted fusion of three-modal features outputs the probability P of ventricular fibrillation. VF =0.98;

[0128] Secondly, dynamic collaborative decision-making units execute;

[0129] At T0+4.1 seconds:

[0130] Receive P VF =0.98, determine P VF => 0.95;

[0131] Activate the level three audible and visual alarm device;

[0132] Finally, the closed-loop optimization unit is executed;

[0133] No misdiagnosed samples were received, so the model update operation was not performed.

[0134] Example 3: Optimization of the Level II Risk Case Model

[0135] First, the multimodal synchronous sensing unit executes;

[0136] The system synchronously acquires patient ECG, PCG, and PPG data streams using a hardware clock. It timestamps the ECG, PCG, and PPG signals separately using a hardware clock to achieve timestamp alignment of the three signals and outputs time-aligned ECG, PCG, and PPG data streams to the edge intelligent diagnostic unit.

[0137] Next, the edge intelligent diagnostic unit executes;

[0138] A lightweight arrhythmia detection model with a runtime of 0.95MB;

[0139] The QRS localization submodule performs a one-dimensional trainable convolution operation on the input ECG signal segment X to calculate... Output the QRS group position of the current segment.

[0140] The risk labeling submodule detects the ST segment offset ΔST = 0.08 mV, ≤ 0.1 mV, and obtains the highest probability value of the arrhythmia probability distribution vector p with confidence max(p) = 0.82.

[0141] According to the risk level determination formula:

[0142]

[0143] When 0.7 ≤ max(p) ≥ 0.9, the output is R = II;

[0144] Output risk level R=II and confidence level C e =0.75;

[0145] Then, the dynamic collaborative decision-making unit executes;

[0146] Receive the C output from the edge intelligent diagnostic unit e =0.75 and R=II after;

[0147] Judgment C e If the value is less than 0.9, a cloud-based auxiliary analysis process is triggered.

[0148] Secondly, the cloud-based integrated analysis unit executes;

[0149] The cross-modal alignment module aligns three-modal time series using a dynamic time warping algorithm.

[0150] Attention fusion module, calculates w i Coefficients and fused features:

[0151] Corrected diagnosis: Ventricular premature beats; original marginal diagnosis result was atrial fibrillation.

[0152] This sample was marked as a misdiagnosed case;

[0153] Secondly, dynamic collaborative decision-making units execute;

[0154] Receive raw ECG, PCG, and PPG data of misdiagnosed samples, along with cloud-based correction tags;

[0155] A new lightweight model was generated using distillation compression technology: the new model has parameters of 0.89MB < 1MB;

[0156] Deploy and update the model to the edge intelligent diagnostic unit;

[0157] Finally, the model is updated and validated;

[0158] The updated edge intelligent diagnostic unit handles similar cases as follows:

[0159] QRS positioning submodule execution Operations;

[0160] Risk Marking Submodule Output: C e =0.96>0.75, the correct diagnosis is premature ventricular contractions, risk level R=II.

[0161] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0162] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An intelligent diagnostic system for electrocardiogram monitoring based on deep learning and multimodal fusion technology, characterized in that: Includes the following units: The multimodal synchronous sensing unit synchronously acquires ECG, PCG, and PPG data streams and achieves timestamp alignment through a hardware clock. The edge intelligent diagnostic unit connects to the multimodal synchronous sensing unit, runs a lightweight arrhythmia detection model with fewer than 1MB of parameters, and outputs the confidence level C in real time. e and risk level R∈{I, II, III}; The cloud-based fusion analysis unit receives the C output from the edge intelligent diagnostic unit. e R, through a cross-modal attention mechanism, dynamically weights and fuses ECG features, PCG and PPG data to generate a deep cardiac pathology report; Dynamic collaborative decision-making unit, connected to cloud-based fusion analysis unit, when R=III or C e When the value is less than 0.9, a deep cloud-based diagnostic is triggered within 5 seconds and the results are fed back to the closed-loop optimization unit. The closed-loop optimization unit receives misdiagnosed samples from cloud-based diagnostics, updates model parameters, and generates a lightweight version through distillation and compression. The lightweight model output by the closed-loop optimization unit is then deployed to the edge intelligent diagnostic unit to complete the optimization cycle.

2. The intelligent ECG monitoring and diagnostic system based on deep learning and multimodal fusion technology according to claim 1, characterized in that, The edge intelligent diagnostic unit performs: The QRS localization submodule employs a one-dimensional trainable convolutional kernel W. q calculate The risk marking submodule forces R=III when the ST segment offset ΔST>0.1mV; in, denoted as QRS complex location; X represents the input ECG signal segment; ReLU is the corrected linear unit function; * represents a one-dimensional convolution operation.

3. The intelligent ECG monitoring and diagnostic system based on deep learning and multimodal fusion technology according to claim 2, characterized in that, The risk labeling submodule calculates the risk level: Where ΔST is the voltage offset of the ST segment; p is the probability distribution vector of arrhythmia; max(p) is the highest probability confidence level; if is the conditional judgment keyword; mV is millivolt; otherwise is other cases.

4. The intelligent diagnostic system for electrocardiogram monitoring based on deep learning and multimodal fusion technology according to claim 1, characterized in that, The cloud-based fusion analysis unit includes a cross-modal alignment module and an attention fusion module; The cross-modal alignment module aligns time series using a dynamic time warping algorithm; The weight calculation of the attention fusion module satisfies w i =softmax(v T tanh(W s f i )); Output the probability P of ventricular fibrillation VF ; Among them, w i Here are the modal weights; softmax is the normalized exponential function; v T is the vector transpose; v is the attention weight vector; tanh is the hyperbolic tangent activation function; W s For shared weight matrix; f i It is a single-modal feature vector.

5. The intelligent diagnostic system for electrocardiogram monitoring based on deep learning and multimodal fusion technology according to claim 4, characterized in that, The dynamic collaborative decision-making unit activates a low-latency channel for cases with R=III, and outputs the ventricular fibrillation probability P. VF A level 3 audible and visual alarm is triggered when the value is >0.

95.

6. The intelligent diagnostic system for electrocardiogram monitoring based on deep learning and multimodal fusion technology according to claim 1, characterized in that, The closed-loop optimization unit generates a lightweight model through distillation and compression, satisfying the requirement that the number of parameters is less than 1MB.

7. The edge computing node of the intelligent diagnostic system for electrocardiogram monitoring according to claims 1-6, characterized in that, The edge computing node performs: The hardware accelerator performs QRS positioning operations; The dynamic cache stores the original data of case R=III.

8. The cloud-based analysis platform of the intelligent diagnostic system for electrocardiogram monitoring according to claims 1-6, characterized in that, The cloud analytics platform performs the following: Implement attention-weighted fusion in the attention fusion module of the cloud-based fusion analysis unit; Perform model distillation compression in the closed-loop optimization unit.

9. The intelligent diagnostic system for electrocardiogram monitoring based on deep learning and multimodal fusion technology according to claim 6, characterized in that, The lightweight model has a latency of ≤50ms, and the cloud response time for critical cases is ≤5 seconds.

10. The intelligent diagnostic system for electrocardiogram monitoring based on deep learning and multimodal fusion technology according to claim 1, characterized in that, The risk levels {I, II, III} correspond to: Level I: Edge diagnostics are covered; Level II: Requires cloud-based analysis; Level III: Real-time cloud-based in-depth diagnostics and alerts.

Images