Single-lead ecg-based atrial fibrillation risk prediction and closed-loop management method and system

Abstract

The application discloses a method and system for atrial fibrillation risk prediction and closed-loop management based on single-lead ECG, the method comprising: fusing the quality parameters of ECG signals and the motion quality parameters of three-axis acceleration signals to calculate signal quality index SQI; inputting the ECG signal segments evaluated as high quality into an integrated learning detection model to output the classification results of the current ECG signal segments; when the classification result of the current ECG signal segment is non-atrial fibrillation, extracting the ECG features in the continuous historical monitoring data and inputting the pre-trained time series deep learning model to output the atrial fibrillation occurrence probabilities of multiple time windows; fusing the classification result of the current ECG signal segment, the atrial fibrillation occurrence probabilities of multiple time windows and the like to generate a comprehensive risk score CRS through weighted calculation; and dynamically adjusting the collection frequency and duration of subsequent ECG signals according to the risk level of the CRS. The application forms a complete closed-loop management system and realizes intelligent linkage of risk assessment and monitoring strategies.

Description

Technical Field

[0001] This invention relates to the field of digital healthcare and intelligent health management technology, and in particular to a method and system for atrial fibrillation risk prediction and closed-loop management based on single-lead ECG. Background Technology

[0002] Currently, atrial fibrillation detection devices based on single-lead electrocardiograms (such as AliveCor Kardia Mobile, Apple Watch Series 4 and later models) represent the most representative existing technology on the market. Their technical solutions are as follows: Figure 1 As shown, the existing technology has the following drawbacks: 1) Limited functionality, confined to current detection: Current technology can only determine "whether atrial fibrillation exists in the current ECG recording." For patients with paroxysmal atrial fibrillation, sinus rhythm ECGs recorded when not experiencing an episode are considered normal, thus missing a large number of potential patients. It cannot use ECG signals that do not contain atrial fibrillation to predict the risk of developing atrial fibrillation in the future (e.g., within the next 14 days), missing the optimal opportunity for early intervention.

[0003] 2) It cannot utilize long-term continuous monitoring data. Currently, the algorithm only analyzes single records and cannot form long-term risk information such as risk trends, risk change rates, and disease window predictions.

[0004] 3) The warning threshold is fixed. The existing system uses a uniform threshold: risk > 0.5, which triggers an alarm and cannot be dynamically adjusted according to individual circumstances.

[0005] 4) Simple signal quality control, prone to false alarms: Existing devices are prone to motion artifacts when the user moves because the dry electrodes have unstable contact with the skin. Their signal quality assessment methods are relatively simple (e.g., based solely on lead impedance), resulting in a large number of noise segments being fed into the algorithm for analysis, leading to unclassifiable or misclassified "room problems," causing unnecessary anxiety or medical consultations for the user.

[0006] 5) Lack of a closed-loop monitoring system, making it impossible to dynamically adjust monitoring strategies based on risk changes: The existing equipment's workflow is detection-result output-end, which is a one-time, open-loop detection mode. The system cannot automatically adjust the monitoring frequency and data collection strategy according to the user's risk change trends, resulting in insufficient data collection during high-risk phases and wasted resources during low-risk phases, thus failing to form an effective closed-loop risk management system. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for atrial fibrillation risk prediction and closed-loop management based on single-lead ECG, forming a complete closed-loop management system to achieve intelligent linkage between risk assessment and monitoring strategies.

[0008] To achieve the above objectives, the present invention provides the following solution: A method for predicting and managing atrial fibrillation risk based on single-lead ECG includes the following steps: S1. Collect the user's single-lead ECG signal and synchronous triaxial acceleration signal through a wearable device; S2. Combine the quality parameters of the ECG signal and the motion quality parameters of the triaxial acceleration signal to calculate the signal quality index (SQI) and classify the quality level according to the SQI. S3. Construct and train an ensemble learning detection model. Input the ECG signal segment evaluated as high quality into the trained ensemble learning detection model and output the classification result of the current ECG signal segment. S4. When the current ECG signal segment classification result is non-atrial fibrillation, extract the ECG features from the continuous historical monitoring data, input them into the pre-trained time series deep learning model, and output the probability of atrial fibrillation occurrence in multiple future time windows. S5. Continuously record the probability of atrial fibrillation occurrence at multiple time points to construct a risk trajectory curve, calculate the risk growth rate, risk growth acceleration, risk volatility index and duration of sustained high risk, and identify risk evolution patterns. S6. By integrating the classification results of the current ECG signal segment, the probability of atrial fibrillation occurrence in multiple time windows, risk trajectory status, signal quality level, and the number of historical abnormal events, a comprehensive risk score (CRS) is generated through weighted calculation. S7. Based on the risk level of CRS, dynamically adjust the frequency and duration of subsequent ECG signal acquisition to form a closed-loop management of risk assessment, strategy adjustment, and reassessment.

[0009] Preferably, in step S2, the signal quality index (SQI) is calculated, and the quality level is classified according to the SQI, specifically including: Quality parameters of the ECG signal are extracted, including signal-to-noise ratio, kurtosis, skewness, R-wave recognition rate, and RR interval continuity; motion quality parameters are extracted, including acceleration mean square error and attitude change rate; the SQI is calculated based on the above parameters, and the quality levels are classified as follows: Grade A: SQI ≥ 80; Grade B: 60 ≤ SQI < 80; Grade C: 40 ≤ SQI < 60; Grade D: SQI < 40. Grades A and B are high-quality data, while Grades C and D are low-quality data. For Grades C and D, their participation in high-risk decision-making should be restricted, or the user should be prompted to recollect the data.

[0010] Preferably, in S3, the ensemble learning detection model is constructed and trained, specifically including: The analysis results of ECG signal segments from multiple open-source atrial fibrillation detection algorithms with different structures are used as feature vectors and input into a random forest meta-classifier for final decision-making, resulting in the final ensemble learning detection model. The analysis results include at least the regularity of the RR interval, the presence of the P wave, heart rate variability, and rhythm complexity.

[0011] Preferably, in S5, the identified risk evolution patterns include: low-risk stable type, continuously rising type, volatile rising type and continuously high-risk type.

[0012] Preferably, in S6, the formula for calculating and generating the Comprehensive Risk Score (CRS) is as follows: CRS = f(a1*classification result score + a2*atrial fibrillation probability score + a3*trajectory score + a4*quality factor + a5*historical frequency score); Among them, a1-a4 are preset weighting factors based on different clinical significance; when the signal quality is low, the confidence of CRS is reduced and a re-acquisition is prompted by adjusting the weight of quality factor a4.

[0013] Preferably, in S7, the acquisition frequency and duration of subsequent ECG signals are dynamically adjusted according to the risk level of CRS, specifically including: When CRS < 0.3, it is a Level I risk, and the normal mode is executed, with ECG collected every 4 hours; when 0.3 <= CRS < 0.5, it is a Level II risk, and the attention mode is executed, with ECG collected every 1 hour; when 0.5 <= CRS < 0.75, it is a Level III risk, and the enhanced mode is executed, with ECG collected every 15 minutes; when CRS >= 0.75, it is a Level IV risk, and the continuous monitoring mode is executed, starting continuous ECG monitoring, and the accelerometer is activated for fall detection.

[0014] Preferably, step S7 further includes a dynamic threshold adjustment step: The CRS threshold for triggering warnings is dynamically adjusted based on the user's individual characteristics. These characteristics include at least age, baseline heart rate, historical average risk, number of atrial fibrillation episodes, and comorbidities. For elderly individuals, those with abnormal baseline heart rate, or those with a history of atrial fibrillation, the CRS warning threshold is automatically lowered.

[0015] This invention also provides a single-lead ECG-based atrial fibrillation risk prediction and closed-loop management system, applied to any of the above-mentioned single-lead ECG-based atrial fibrillation risk prediction and closed-loop management methods, comprising: The data acquisition module is used to acquire the user's single-lead ECG signal and synchronous triaxial acceleration signal through the wearable device; The quality level classification module is used to fuse the quality parameters of ECG signals and the motion quality parameters of triaxial acceleration signals, calculate the signal quality index (SQI), and classify the quality level based on the SQI. The real-time detection module is used to build and train an ensemble learning detection model. It takes ECG signal segments evaluated as high quality as input to the trained ensemble learning detection model and outputs the classification result of the current ECG signal segment. The risk analysis module is used to extract ECG features from continuous historical monitoring data when the current signal segment classification result is not atrial fibrillation, input them into a pre-trained time series deep learning model, and output the probability of atrial fibrillation occurrence in multiple future time windows. The trajectory analysis module is used to continuously record the probability of atrial fibrillation occurrence at multiple time points to construct a risk trajectory curve, calculate the risk growth rate, risk growth acceleration, risk volatility index, and duration of sustained high risk, and identify risk evolution patterns. The comprehensive risk score generation module is used to integrate the classification results of the current ECG signal segment, the probability of atrial fibrillation occurrence in multiple time windows, risk trajectory status, signal quality level, and the number of historical abnormal events, and generate a comprehensive risk score (CRS) through weighted calculation. The strategy adjustment module is used to dynamically adjust the frequency and duration of subsequent ECG signal acquisition based on the risk level of CRS, forming a closed-loop management of risk assessment, strategy adjustment, and reassessment.

[0016] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements a method for atrial fibrillation risk prediction and closed-loop management based on single-lead ECG as described above.

[0017] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: (1) This invention is the first to establish a three-tiered risk management system of detection, prediction, and trajectory analysis, achieving a fundamental shift from detecting atrial fibrillation to providing early warning of atrial fibrillation: This invention can not only detect atrial fibrillation in the current ECG segment (Level 1), but also predict the risk of future atrial fibrillation in multiple time windows based on ECGs without atrial fibrillation (Level 2), and further capture the changing trend of risk through risk trajectory analysis (Level 3). This three-tiered progressive risk management system makes early warning and proactive intervention for patients with paroxysmal atrial fibrillation possible, rather than passively waiting for the next attack.

[0018] (2) This invention utilizes the dynamic characteristics of risk trajectories to identify onset trends in advance, significantly improving the advance warning time: By calculating the risk growth rate (DR / DT) and risk growth acceleration (D2R / DT2), the system can distinguish between two distinct risk evolution modes: stable low risk and continuous deterioration. When the risk trajectory exhibits a continuously increasing characteristic, even if the current CRS has not yet triggered the alarm threshold, the system can issue an early warning, gaining a valuable time window for clinical intervention.

[0019] (3) The present invention adopts a joint decision-making mechanism (CRS) based on multi-source information fusion, which significantly reduces the risk of misjudgment by a single model: The present invention does not rely on the output of any single model, but integrates five-dimensional information such as detection results, prediction results, trajectory status, signal quality and historical events to form a comprehensive risk score. This collaborative decision-making architecture effectively makes up for the shortcomings of a single model and greatly reduces the false alarm rate and the false negative rate. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic flowchart of the atrial fibrillation detection method based on single-lead electrocardiogram in the prior art provided by the present invention; Figure 2 This is a flowchart illustrating the atrial fibrillation risk prediction and closed-loop management method based on single-lead ECG provided by the present invention. Detailed Implementation

[0022] 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.

[0023] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0024] like Figure 2 As shown, the present invention provides a method for atrial fibrillation risk prediction and closed-loop management based on single-lead ECG, comprising the following steps: S1. Collect the user's single-lead ECG signal and synchronous triaxial acceleration signal through a wearable device; S2. Combine the quality parameters of the ECG signal and the motion quality parameters of the triaxial acceleration signal to calculate the signal quality index (SQI) and classify the quality level according to the SQI. S3. Construct and train an ensemble learning detection model. Input the ECG signal segment evaluated as high quality into the trained ensemble learning detection model and output the classification result of the current ECG signal segment. S4. When the current ECG signal segment classification result is non-atrial fibrillation, extract the ECG features from the continuous historical monitoring data, input them into the pre-trained time series deep learning model, and output the probability of atrial fibrillation occurrence in multiple future time windows. S5. Continuously record the probability of atrial fibrillation occurrence at multiple time points to construct a risk trajectory curve, calculate the risk growth rate, risk growth acceleration, risk volatility index and duration of sustained high risk, and identify risk evolution patterns. S6. By integrating the classification results of the current ECG signal segment, the probability of atrial fibrillation occurrence in multiple time windows, risk trajectory status, signal quality level, and the number of historical abnormal events, a comprehensive risk score (CRS) is generated through weighted calculation. S7. Based on the risk level of CRS, dynamically adjust the frequency and duration of subsequent ECG signal acquisition to form a closed-loop management of risk assessment, strategy adjustment, and reassessment.

[0025] The core of this invention lies in constructing a five-layer closed-loop risk management system, achieving a fundamental shift from passively detecting atrial fibrillation to proactively warning of it. Its overall architecture is as follows: Figure 2 As shown, it specifically includes: First layer - Real-time atrial fibrillation detection: Determines whether an atrial fibrillation rhythm exists in the current ECG segment; Second layer - Multi-time window atrial fibrillation risk prediction: When atrial fibrillation is not detected by the current ECG, it automatically predicts the risk of atrial fibrillation in the next 24 hours, 72 hours, 7 days, and 14 days; The third layer - risk trajectory analysis: continuously records the risk prediction values ​​at multiple time points, constructs the risk trajectory curve, calculates the risk growth rate, acceleration and volatility index, and identifies risk evolution patterns; The fourth layer - joint decision-making and dynamic risk stratification: integrates current detection results, risk prediction results, risk trajectory status, signal quality level and number of historical abnormal events to form a comprehensive risk score (Cardiac Risk Score, CRS). Fifth layer - Adaptive adjustment of monitoring strategy: Automatically adjust the monitoring frequency and data collection strategy according to the CRS risk level, forming a closed-loop management mechanism of risk increase - increase monitoring frequency - obtain more data - reassess risk.

[0026] The method for closed-loop management of atrial fibrillation risk prediction and active monitoring based on single-lead ECG as described in this embodiment of the invention specifically includes: Step S1: Signal Acquisition The wearable device of the present invention can be a wristband (such as a smartwatch), a chest patch, or a handheld device.

[0027] The device uses a single-lead dry electrode to continuously or on-demand acquire the user's ECG signals at a sampling rate of ≥250Hz. A triaxial accelerometer can also be integrated internally for synchronous acquisition of motion signals.

[0028] Step S2: Signal quality assessment Preprocessing: The acquired raw ECG signals were bandpass filtered from 0.5 to 40 Hz to remove baseline drift and high-frequency electromyographic noise.

[0029] Quality assessment: A signal quality assessment algorithm based on multi-parameter fusion is employed. Specifically, two types of indicators are calculated: (1) ECG quality parameters: including signal-to-noise ratio (SNR), kurtosis, skewness, R-wave recognition rate, and RR interval continuity; (2) Motion quality parameters: including the root mean square error of acceleration and the rate of change of posture.

[0030] The above parameters are combined to form the Signal Quality Index (SQI, with a value ranging from 0 to 100), and the quality is divided into four levels: Grade A (SQI>=80): High quality, can be directly used for AI analysis; Grade B (60<=SQI<80): Acceptable, to be used after enhancement; Grade C (40 <= SQI < 60): Poor quality, only suitable for auxiliary analysis; Grade D (SQI < 40): Low quality. Pause the analysis and prompt the user to re-collect data.

[0031] Grade C signals can be used for heart rate variability (HRV) trend statistics and long-term baseline characteristic calculation, but are not used as independent inputs for the real-time atrial fibrillation detection model (S3) and risk prediction model (S4); the analysis results of Grade C signals do not independently trigger risk warnings of Grade III (CRS≥0.5) and above, and need to be combined with the assessment results of Grade A / B signals in the same period for joint judgment. When the quality level is below the threshold, the analysis is automatically paused and the user is prompted that the signal quality is insufficient and to remain still before re-acquiring data, thereby reducing the false alarm rate caused by motion artifacts from the source.

[0032] Step S3: Real-time atrial fibrillation detection Input: High-quality ECG clips (e.g., 30-60 seconds) that have undergone quality assessment.

[0033] Model: An ensemble learning model is employed. Specifically, the outputs of multiple open-source atrial fibrillation detection algorithms with varying structures are used as features and input into a random forest meta-classifier for final decision-making. This meta-classifier is pre-trained on a large-scale labeled dataset containing multi-source data.

[0034] Analysis includes: RR interval regularity, P wave presence, heart rate variability (HRV), and rhythm complexity.

[0035] Function: Outputs the classification result of the current ECG segment - atrial fibrillation, non-atrial fibrillation, or cannot be determined.

[0036] When atrial fibrillation is detected, the method directly enters the early warning process; when no atrial fibrillation is detected, the risk prediction in step S4 is automatically initiated.

[0037] Step S4: Predicting the Risk of Atrial Fibrillation in Multiple Time Windows Input: The system extracts ECG segments without atrial fibrillation from continuous historical monitoring data.

[0038] Feature extraction: The model automatically extracts the following three types of features from ECG segments: (1) ECG morphological characteristics: P wave duration, P wave terminal potential (PTFV1), QT interval, ST segment characteristics, T wave morphology; (2) Heart rate variability (HRV) characteristics: SDNN (standard deviation of RR intervals of all normal sinus beats), RMSSD (root mean square of the difference between adjacent RR intervals), pNN50 (percentage of adjacent RR intervals with a difference >50ms); (3) Rhythm stability characteristics: RR dispersion and periodic fluctuation index.

[0039] Model: A CNN-LSTM cascaded deep learning model is used, which can predict the probability of atrial fibrillation in multiple time windows after training.

[0040] Output: Risk probability in the next 24 hours, risk probability in the next 72 hours, risk probability in the next 7 days, and risk probability in the next 14 days.

[0041] For example: If a user wears the device for 24 hours and all ECG segments are non-atrial fibrillation, the system outputs: 24h risk: 22%, 72h risk: 45%, 7-day risk: 67%, 14-day risk: 84%, indicating that the risk is showing a continuous upward trend.

[0042] Step S5: Risk Trajectory Analysis The system continuously records the risk prediction values ​​R1, R2, R3, ..., Rn at multiple time points, forming a risk trajectory curve.

[0043] Based on this, the following dynamic indicators are further calculated: (1) Risk growth rate (DR / DT): measures the magnitude of change in risk value per unit time, reflecting the speed of risk change; (2) Risk growth acceleration (D2R / DT2): measures the changing trend of the risk growth rate itself, reflecting whether the risk is accelerating its deterioration; (3) Risk volatility index: reflects the degree of short-term volatility of risk value and is used to distinguish between stable changes and violent fluctuations; (4) Duration of sustained high risk: The duration during which the statistical risk value is continuously higher than the threshold.

[0044] Based on the above indicators, the system automatically classifies risk evolution patterns into the following four types: Type 1 - Low-risk stable type: The risk value remains low for a long time with very little fluctuation, indicating that the patient's condition is stable; Type 2 - Continuously Rising Type: The risk value shows a continuous upward trend, DR / DT>0 and continues to increase, indicating that the condition is gradually deteriorating and early intervention is required; Type 3 - Increased volatility: The overall risk value is elevated but accompanied by large fluctuations, indicating that the patient is in an unstable state with paroxysmal characteristics; Type 4 - Persistently High Risk: The risk value remains high for a long period, indicating that the patient is in a high-risk state and requires close monitoring.

[0045] Compared to a single risk prediction which only provides a snapshot of the risk at the current moment, risk trajectory analysis can capture the changing trends of risk, thereby identifying the trend of impending atrial fibrillation and providing an earlier warning window for clinical intervention.

[0046] Step S6: Joint Decision-Making Mechanism This system does not directly use the output of any single model for early warning. Instead, it constructs a joint decision-making mechanism that integrates multi-source information to form a comprehensive risk score (Cardiac Risk Score, CRS).

[0047] CRS calculation incorporates the following five types of information: (1) Current atrial fibrillation detection results (from S3): Weighting factor a1, significantly improves CRS when atrial fibrillation is detected; (2) Risk prediction results of multiple time windows (from S4): weight factor a2, weighted and fused according to the predicted values ​​of each time window; (3) Risk trajectory status (from S5): Weighting factor a3, a continuously rising type or a continuously high-risk type will significantly increase CRS; (4) Signal quality level (from S2): Weighting factor a4, when the signal quality is low, the confidence of CRS is reduced and a reacquisition is prompted; (5) Number of historical abnormal events: weighting factor a5, which counts the frequency of atrial fibrillation events within a certain period of time.

[0048] Final CRS = f(a1 * detection score + a2 * prediction score + a3 * trajectory score + a4 * quality factor + a5 * historical frequency score) Based on the CRS value, the system outputs a four-level risk level: Level I risk (CRS<0.3): Low risk, maintain routine monitoring; Level II risk (0.3 <= CRS < 0.5): Low to medium risk, monitor frequency should be increased appropriately; Level III Risk (0.5 <= CRS < 0.75): Medium to high risk, require significantly enhanced monitoring and user alerts. Level IV risk (CRS>=0.75): High risk, triggering the highest level of warning, immediate medical evaluation is recommended.

[0049] The core value of the joint decision-making mechanism lies in: by integrating multi-dimensional information for collaborative judgment, it can significantly reduce the risk of misjudgment caused by a single model, and avoid excessive alarms caused by a single abnormal signal or delays caused by a single model's missed detection.

[0050] Step S7: Dynamic threshold adjustment This invention abandons the fixed, uniform alarm threshold in existing technologies and adopts an individualized, dynamic threshold adjustment mechanism.

[0051] The system automatically adjusts the CRS alarm threshold based on the following individual characteristics: (1) Age: The older the age, the lower the threshold should be to improve the sensitivity of the early warning; (2) Baseline heart rate: The threshold should be appropriately reduced for those with abnormal resting heart rate; (3) Long-term risk mean: The threshold for individuals with a high historical risk mean is appropriately lowered; (4) Number of atrial fibrillation episodes in history: The more episodes, the lower the threshold should be. (5) Comorbid factors: such as hypertension, diabetes, heart failure and other atrial fibrillation risk factors.

[0052] For example, the CRS alarm threshold for the general healthy population is set at 0.75, while for high-risk patients over 65 years of age with a history of hypertension and who have experienced atrial fibrillation within the past 6 months, the system automatically lowers the threshold to 0.55, achieving individualized risk management and striking the optimal balance between sensitivity and specificity.

[0053] Step S8: Adaptive adjustment of monitoring strategy This invention achieves a closed-loop linkage between monitoring strategies and risk assessment results. The monitoring mode is automatically adjusted based on the current CRS risk level. Level I risk (normal mode): ECG is collected every 4 hours, each collection lasts 30-60 seconds; Level II Risk (Watch Mode): ECG is collected every 1 hour, with each collection lasting 60 seconds; Level III Risk (Enhanced Mode): ECG is collected every 15 minutes, with each collection lasting 60 seconds; Level IV Risk (Continuous Monitoring Mode): Continuous ECG monitoring is initiated, and the accelerometer is activated for fall detection.

[0054] The adjustment of monitoring strategies follows the following closed-loop logic: Risk level increases - monitoring frequency is automatically increased - more high-density data is obtained - risk is reassessed based on new data - if the risk decreases, regular monitoring is resumed; if the risk continues to increase, monitoring is maintained or further strengthened.

[0055] This closed-loop management mechanism ensures that sufficient data density can be captured for accurate assessment during high-risk phases, while avoiding unnecessary resource consumption during low-risk phases, thus achieving a dynamic balance between risk assessment and monitoring strategies.

[0056] Step S9: Interpretability Output For the output results of risk prediction and joint decision-making, the system automatically generates a feature contribution map, highlighting the waveform regions that contribute the most to the decision, including: P-wave abnormal regions, RR interval abnormal regions, and QT interval variation regions.

[0057] Possible implementation methods include: Grad-CAM (gradient-weighted class activation mapping), SHAP (Shapley Additive Explanations), Integrated Gradients, etc.

[0058] Example scenario: When the model predicts that a certain segment of sinus rhythm ECG is at high risk of future atrial fibrillation, the heat map highlights the terminal part of the P wave, suggesting to clinicians that there may be structural or electrical conduction abnormalities in the patient's left atrium, providing interpretable clinical evidence for the AI ​​prediction results.

[0059] Step S10: Federated Learning Model Update This method supports federated learning frameworks and employs a hierarchical update strategy to continuously optimize the model in the system: (1) Federated update of S4 risk prediction model (main update object) The CNN-LSTM deep learning model in S4 is the core objective of federated updates. The specific process is as follows: Local training: Each local node (user device or hospital server) trains the CNN-LSTM model using newly added labeled data locally, without exporting the original ECG data, and calculates the model gradient or weight update. Local training labels include: subsequently clinically confirmed atrial fibrillation events (positive samples) and long-term follow-up confirmed no-atrial fibrillation records (negative samples).

[0060] Parameter upload: Each node only uploads the updated model parameters (such as gradients or differential weights) that have been anonymized to the central aggregation server, and does not upload any original ECG signals, user identity information or other privacy data.

[0061] Secure aggregation: The central server uses a federated averaging (FedAvg) algorithm to weight and aggregate the parameter updates uploaded by each node to generate a new global model. Optional aggregation strategies include: weighted averaging based on the amount of data from each node, or using differential privacy technology to add noise to the gradient before aggregation to further protect privacy.

[0062] Model distribution: The aggregated global model is distributed back to each local node to replace the original local CNN+LSTM model, completing one round of federated update.

[0063] Iteration frequency: Federated updates can be triggered at fixed intervals (such as once a week) or after the cumulative number of newly labeled samples reaches a preset number.

[0064] (2) Updating the S3 ensemble learning detection model (auxiliary update object) The update of the random forest meta-classifier in S3 is achieved in the following way: The central server uses the high-quality feature representation generated by the aggregated global CNN-LSTM model, combined with the desensitized statistical information reported by each node (such as the confusion matrix statistics of atrial fibrillation detection), to periodically retrain the random forest meta-classifier. The retrained meta-classifier is distributed to each local node along with the model distribution steps of the federated learning framework.

[0065] (3) The core value of layered updates S4's predictive model, as the main target of federated updates, can continuously learn the evolution patterns of atrial fibrillation risk in different populations from multi-source data, so that the predictive performance can be continuously improved as the user population expands. S3's detection model's meta-classifier maintains its compatibility with the output of the base algorithm through periodic retraining; data remains local, and the model continues to evolve, achieving continuous model evolution while protecting user data privacy.

[0066] This invention also achieves the following technical effects: (1) Achieve individualized dynamic threshold adjustment to achieve the optimal balance between sensitivity and specificity: Abandon fixed uniform thresholds and dynamically adjust alarm thresholds based on individual characteristics such as age, baseline heart rate, history of atrial fibrillation, and comorbidities. Lower alarm thresholds are used for high-risk patients to improve sensitivity, while higher thresholds are used for low-risk individuals to reduce false alarms, thus achieving truly individualized risk management.

[0067] (2) A complete closed-loop management system is formed to realize the intelligent linkage between risk assessment and monitoring strategies: When the risk level increases, the system automatically increases the monitoring frequency, obtains high-density data, and reassesses the risk, forming a dynamic closed loop of assessment-adjustment-reassessment. This ensures that there is enough data to support accurate decision-making in the high-risk stage, while avoiding unnecessary resource consumption in the low-risk stage.

[0068] (3) By integrating a four-level signal quality assessment system with motion sensors, the false alarm rate is reduced from the source: The signal quality assessment module of this invention not only considers the characteristics of ECG itself, but also integrates acceleration sensor data to judge the motion state, and finely divides the signal quality into four levels: A / B / C / D, effectively filtering low-quality signals caused by motion artifacts, and significantly reducing the situation of being unable to classify and false alarms.

[0069] (4) Significantly improves the robustness and cross-platform generalization ability of the model: The present invention adopts an ensemble learning model (multi-foundation algorithm-random forest meta-classifier). In the cross-validation of multiple public datasets, its average F1 score variation index (V_P) is much lower than any single foundation algorithm. It can maintain stable and reliable performance on data collected from different devices such as AliveCor and Apple Watch.

[0070] This invention also provides a single-lead ECG-based atrial fibrillation risk prediction and closed-loop management system, applied to any of the above-mentioned single-lead ECG-based atrial fibrillation risk prediction and closed-loop management methods, comprising: The data acquisition module is used to acquire the user's single-lead ECG signal and synchronous triaxial acceleration signal through the wearable device; The quality level classification module is used to fuse the quality parameters of ECG signals and the motion quality parameters of triaxial acceleration signals, calculate the signal quality index (SQI), and classify the quality level based on the SQI. The real-time detection module is used to build and train an ensemble learning detection model. It takes ECG signal segments evaluated as high quality as input to the trained ensemble learning detection model and outputs the classification result of the current ECG signal segment. The risk analysis module is used to extract ECG features from continuous historical monitoring data when the current signal segment classification result is not atrial fibrillation, input them into a pre-trained time series deep learning model, and output the probability of atrial fibrillation occurrence in multiple future time windows. The trajectory analysis module is used to continuously record the probability of atrial fibrillation occurrence at multiple time points to construct a risk trajectory curve, calculate the risk growth rate, risk growth acceleration, risk volatility index, and duration of sustained high risk, and identify risk evolution patterns. The comprehensive risk score generation module is used to integrate the classification results of the current ECG signal segment, the probability of atrial fibrillation occurrence in multiple time windows, risk trajectory status, signal quality level, and the number of historical abnormal events, and generate a comprehensive risk score (CRS) through weighted calculation. The strategy adjustment module is used to dynamically adjust the frequency and duration of subsequent ECG signal acquisition based on the risk level of CRS, forming a closed-loop management of risk assessment, strategy adjustment, and reassessment.

[0071] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements a method for atrial fibrillation risk prediction and closed-loop management based on single-lead ECG as described above.

[0072] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0073] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for atrial fibrillation risk prediction and closed-loop management based on single-lead ECG, characterized in that, Includes the following steps: S1. Collect the user's single-lead ECG signal and synchronous triaxial acceleration signal through a wearable device; S2. Combine the quality parameters of the ECG signal and the motion quality parameters of the triaxial acceleration signal to calculate the signal quality index (SQI), and classify the quality level according to the signal quality index (SQI). S3. Construct and train an ensemble learning detection model. Input the ECG signal segment evaluated as high quality into the trained ensemble learning detection model and output the classification result of the current ECG signal segment. S4. When the current ECG signal segment classification result is non-atrial fibrillation, extract the ECG features from the continuous historical monitoring data, input them into the pre-trained time series deep learning model, and output the probability of atrial fibrillation occurrence in multiple future time windows. S5. Continuously record the probability of atrial fibrillation occurrence at multiple time points to construct a risk trajectory curve, calculate the risk growth rate, risk growth acceleration, risk volatility index and duration of sustained high risk, and identify risk evolution patterns. S6. By integrating the classification results of the current ECG signal segment, the probability of atrial fibrillation occurrence in multiple time windows, risk trajectory status, signal quality level, and the number of historical abnormal events, a comprehensive risk score (CRS) is generated through weighted calculation. S7. Based on the risk level of the CRS, dynamically adjust the acquisition frequency and duration of subsequent ECG signals to form a closed-loop management of risk assessment, strategy adjustment, and reassessment.

2. The method for atrial fibrillation risk prediction and closed-loop management based on single-lead ECG according to claim 1, characterized in that, In step S2, the signal quality index (SQI) is calculated, and the quality level is classified according to the SQI, specifically including: Quality parameters of the ECG signal are extracted, including signal-to-noise ratio, kurtosis, skewness, R-wave recognition rate, and RR interval continuity; motion quality parameters are extracted, including acceleration mean square error and attitude change rate; the SQI is calculated based on the above parameters, and the quality levels are classified as follows: Grade A: SQI ≥ 80; Grade B: 60 ≤ SQI < 80; Grade C: 40 ≤ SQI < 60; Grade D: SQI < 40. Grades A and B are high-quality data, while Grades C and D are low-quality data. For Grades C and D, their participation in high-risk decision-making should be restricted, or the user should be prompted to recollect the data.

3. The method for atrial fibrillation risk prediction and closed-loop management based on single-lead ECG according to claim 1, characterized in that, In step S3, constructing and training the ensemble learning detection model specifically includes: The analysis results of ECG signal segments from multiple open-source atrial fibrillation detection algorithms with different structures are used as feature vectors and input into a random forest meta-classifier for final decision-making to obtain the final ensemble learning detection model; the analysis results include at least the regularity of RR intervals, the presence of P waves, heart rate variability, and rhythm complexity.

4. The method for atrial fibrillation risk prediction and closed-loop management based on single-lead ECG according to claim 1, characterized in that, In S5, the identified risk evolution patterns include: low-risk stable type, continuously rising type, fluctuating type, and continuously high-risk type.

5. The method for atrial fibrillation risk prediction and closed-loop management based on single-lead ECG according to claim 1, characterized in that, In step S6, the formula for calculating and generating the Comprehensive Risk Score (CRS) is as follows: CRS = f(a1*classification result score + a2*atrial fibrillation probability score + a3*trajectory score + a4*quality factor + a5*historical frequency score); Among them, a1-a4 are preset weighting factors based on different clinical significance; when the signal quality is low, the confidence of CRS is reduced and a re-acquisition is prompted by adjusting the weight of quality factor a4.

6. The method for atrial fibrillation risk prediction and closed-loop management based on single-lead ECG according to claim 1, characterized in that, In step S7, the frequency and duration of subsequent ECG signal acquisition are dynamically adjusted according to the risk level of the CRS, specifically including: When CRS < 0.3, it is a Level I risk, and the normal mode is executed, with ECG collected every 4 hours; when 0.3 <= CRS < 0.5, it is a Level II risk, and the attention mode is executed, with ECG collected every 1 hour; when 0.5 <= CRS < 0.75, it is a Level III risk, and the enhanced mode is executed, with ECG collected every 15 minutes; when CRS >= 0.75, it is a Level IV risk, and the continuous monitoring mode is executed, starting continuous ECG monitoring, and the accelerometer is activated for fall detection.

7. The method for atrial fibrillation risk prediction and closed-loop management based on single-lead ECG according to claim 1, characterized in that, S7 further includes a dynamic threshold adjustment step: The CRS threshold for triggering warnings is dynamically adjusted based on the user's individual characteristics; these individual characteristics include at least age, baseline heart rate, historical average risk, number of historical atrial fibrillation episodes, and comorbidities; for elderly individuals, individuals with abnormal baseline heart rate, or individuals with a history of atrial fibrillation, the CRS threshold is automatically lowered.

8. A risk prediction and closed-loop management system for atrial fibrillation based on single-lead ECG, applied to the execution of the risk prediction and closed-loop management method for atrial fibrillation based on single-lead ECG as described in any one of claims 1-7, characterized in that, include: The data acquisition module is used to acquire the user's single-lead ECG signal and synchronous triaxial acceleration signal through the wearable device; The quality level classification module is used to fuse the quality parameters of the ECG signal and the motion quality parameters of the triaxial acceleration signal, calculate the signal quality index (SQI), and classify the quality level according to the signal quality index (SQI). The real-time detection module is used to build and train an ensemble learning detection model. It takes ECG signal segments evaluated as high quality as input to the trained ensemble learning detection model and outputs the classification result of the current ECG signal segment. The risk analysis module is used to extract ECG features from continuous historical monitoring data when the current signal segment classification result is not atrial fibrillation, input them into a pre-trained time series deep learning model, and output the probability of atrial fibrillation occurrence in multiple future time windows. The trajectory analysis module is used to continuously record the probability of atrial fibrillation occurrence at multiple time points to construct a risk trajectory curve, calculate the risk growth rate, risk growth acceleration, risk volatility index, and duration of sustained high risk, and identify risk evolution patterns. The comprehensive risk score generation module is used to integrate the classification results of the current ECG signal segment, the probability of atrial fibrillation occurrence in multiple time windows, risk trajectory status, signal quality level, and the number of historical abnormal events, and generate a comprehensive risk score (CRS) through weighted calculation. The strategy adjustment module is used to dynamically adjust the acquisition frequency and duration of subsequent ECG signals according to the risk level of the CRS, forming a closed-loop management of risk assessment-strategy adjustment-reassessment.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements a method for atrial fibrillation risk prediction and closed-loop management based on single-lead ECG as described in any one of claims 1-7.

Images