Atrial fibrillation ablation postoperative recurrence continuous learning prediction system based on multi-source electrocardio data

Abstract

The invention relates to an atrial fibrillation ablation postoperative recurrence continuous learning prediction system based on multi-source electrocardiogram data. The system comprises a data acquisition module, a data processing module, a feature recognition module and a risk prediction module. According to the invention, the electrocardiogram representation network is constructed to process different pieces of follow-up visit electrocardiogram information during follow-up visit, so that corresponding risk prediction results can be automatically output at different stages; meanwhile, the model can still obtain stable and transferable electrocardio characteristic representation under the multi-source and multi-device conditions. The cross-domain and cross-device capability enables the model to stably work in a real multi-center clinical environment, retraining for each medical center is not needed, the deployment difficulty and cost are greatly reduced, and controlled updating can be realized by using a small amount of follow-up visit data without depending on historical original electrocardiogram data and retraining the complete model.

Description

Technical Field

[0001] This invention relates to medical data processing technology, and more specifically, to a continuous learning prediction system for recurrence after atrial fibrillation ablation based on multi-source electrocardiogram data. Background Technology

[0002] Predicting recurrence after atrial fibrillation catheter ablation usually relies on a comprehensive analysis of electrocardiogram (ECG) signals, clinical information, and structural indicators. To extract stable features from ECG signals, current technologies often employ deep feature extraction networks to automatically encode waveform structures, replacing manually constructed feature methods (Lamba S, Kumar S, Diwakar M. Feature extraction and arrhythmia classification using deep learning from electrocardiograph signals [J]. Discover Artificial Intelligence, 2025, 5:290.); (Guerra RT, et al. Deep learning approach for automatic heartbeat classification and feature extraction using LSTM autoencoders [J]. Sensors, 2025, 25(5): 1400.); (Prakash AJ, et al. Advances in machine and deep learning for ECG beat classification [J]. Frontiers in Digital Health, 2025,3: 1649923.). Deep networks extract waveform information related to rhythmic changes through local convolution, temporal modeling, or transform domain structures, in order to capture morphological differences associated with recurrence risk.To improve feature consistency and transferability under cross-device and cross-individual conditions, a pre-trained ECG model is introduced into the technical system. This model learns universal representations on large-scale signals, forming feature codes that can be reused in different environments (Zhang S, Du Y, Wang W, et al. ECGFM: A foundation model for ECG analysis trained on a multi-center million-ECG dataset [J]. Information Fusion, 2025: 103363.); (McKeen K, Oliva L, Masood S, et al. ECG-FM: An open electrocardiogram foundation model [J]. JAMIA Open, 2025.); (Cai M, Jiang J, Huang W, et al. SuPreME: A supervised pre-training framework for multimodal ECG representation learning [J / Preprint]. 2025.). Furthermore, in multi-center, multi-device ECG environments, a feature alignment mechanism is used to mitigate signal shifts caused by differences in acquisition conditions, ensuring relatively consistent input representations and thus providing a stable feature foundation for postoperative recurrence prediction.

[0003] In practical implementation, existing postoperative recurrence prediction methods mainly employ static training models, pre-trained representation transfer, multi-center collaborative training, and feature alignment-based adaptation frameworks (Zhang S, Du Y, Wang W, et al. ECGFM: A foundation model for ECG analysis trained on a multi-center million-ECG dataset [J]. Information Fusion, 2025: 103363.); (Alreshidi FS, Alsaffar M, Chengoden R, et al. Fed-CL: an atrial fibrillation prediction system using ECG signals employing federated learning mechanism [J]. Scientific Reports, 2024, 14: 21038.). Static training models establish a mapping relationship between signal features and recurrence status within a fixed time range, but the model parameters remain unchanged after deployment, lacking dynamic adaptation capabilities during follow-up. Pre-trained feature transfer methods can provide cross-domain representations, but their feature space is fixed and cannot automatically adjust to changes in ECG signals during long-term follow-up. Collaborative training mechanisms expand model coverage through parameter aggregation, but rely on batch data construction, making them unsuitable for the continuous generation of small batches of non-stationary data during follow-up. Feature alignment methods can mitigate inconsistencies caused by equipment differences, but their alignment structure is typically fixed during training and cannot be updated as postoperative ECG characteristics gradually change. These techniques constitute the main implementation path for current recurrence prediction.

[0004] However, the performance degradation of the aforementioned existing technologies in practical applications is extremely prominent. Long-term follow-up data of atrial fibrillation patients confirms that the performance of existing technologies declines significantly over time: Performance degradation of static models: In long-term postoperative follow-up scenarios, the performance of statically trained models will gradually decline when faced with continuously changing input signals.

[0005] Decreased generalization ability of pre-trained representations: Pre-trained representations cannot maintain stable consistency on long-term non-stationary data. In cross-domain validation of multi-center clinical data, patient data from different medical centers show significant differences. This severe decline in cross-domain generalization ability indicates that pre-trained representations lack the ability to adapt to changes in equipment and acquisition conditions, becoming a fundamental bottleneck for deployment in multiple medical centers.

[0006] In long-term postoperative follow-up settings, atrial structural remodeling and changes in electrophysiological activity lead to a continuous shift in ECG signal characteristics over time; different monitoring devices, sampling rates, and noise conditions further exacerbate signal inconsistencies. Static models gradually degrade in performance when faced with continuously changing input signals, pre-trained representations cannot maintain stable consistency on long-term non-stationary data, co-training methods cannot provide structured updates to progressively arriving follow-up data, and fixed feature alignment frameworks are ill-suited to adapting to the dynamic changes in postoperative features (Zhang S, Du Y, Wang W, et al. ECGFM: Afoundation model for ECG analysis trained on a multi-center million-ECG dataset [J]. Information Fusion, 2025: 103363.); (Alreshidi FS, Alsaffar M, Chengoden R, et al. Fed-CL: an atrial fibrillation prediction system using ECG signals employing federated learning mechanism [J]. Scientific Reports, 2024, 14: 21038.). Due to the lack of a mechanism to continuously adapt the model under long-term, multi-source heterogeneous conditions, existing technologies struggle to maintain the ability to recognize early feature patterns in postoperative dynamic follow-up scenarios, and also cannot maintain stable output when signal distribution continues to change. This is precisely the core problem that this invention aims to solve. Summary of the Invention

[0007] In view of this, the purpose of this invention is to provide a continuous learning prediction system for recurrence after atrial fibrillation ablation based on multi-source electrocardiogram data.

[0008] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: A continuous learning prediction system for recurrence after atrial fibrillation ablation based on multi-source electrocardiogram data, characterized in that it includes a data acquisition module, a data processing module, a feature recognition module, and a risk prediction module; The data acquisition module is used to collect follow-up electrocardiogram data; The data processing module includes a noise processing sub-strategy and a normalization sub-strategy. The noise processing sub-strategy is used to process the noise in the follow-up ECG data, and the normalization sub-strategy is used to normalize the follow-up ECG data to obtain a standard input tensor. The feature recognition module is configured with an electrocardiogram (ECG) representation network, which is used to identify the representation vectors in the standard input tensor to generate follow-up feature representations. The risk prediction module is used to generate relapse prediction information based on follow-up characteristics.

[0009] Furthermore, it also includes an update trigger module and an update calculation module; The update triggering module is configured with a dynamic threshold and a drift calculation sub-strategy. The drift calculation sub-strategy is used to calculate the current characterization drift based on the patient's follow-up characteristics. If the characterization drift is greater than the dynamic threshold, the update calculation module will work. The update calculation module includes a parameter learning strategy used to update the learned parameters of the ECG representation network. Compared to the passive approach of manually specifying update time points in traditional static models, this invention can automatically identify moments when significant changes occur in the ECG distribution. Instead of updating the model at each fixed follow-up point, this invention precisely triggers updates based on dynamic thresholds. This reduces unnecessary updates, significantly lowering computational costs and storage requirements, while improving the system's practicality while maintaining accuracy.

[0010] Furthermore, it also includes a network building module, which includes a temporal convolutional encoder, a rhythm-sensitive filter, and an adaptive aggregation strategy.

[0011] Furthermore, the network building module also includes consistency constraints.

[0012] Furthermore, the drift calculation sub-strategy includes averaging the representation vectors of all patients at the current follow-up node to obtain the follow-up mean, and calculating the difference between the current follow-up mean and the follow-up mean at the previous time point to generate the representation drift. This invention achieves automatic detection of changes in electrocardiogram distribution during the follow-up period by introducing a feature drift based on changes in the mean of electrocardiogram representation.

[0013] Furthermore: the parameter learning strategy includes a parameter update representation, which is as follows: Where η is the preset learning rate and λ is the preset regularization weight. For the first The orthogonal components corresponding to the term parameters, For the first The significance value of the parameter in the previous time step. For the first The parameter value of the item. For the first The parameter values ​​of each term at the previous time step. This invention employs a joint update mechanism of parameter importance constraints and gradient orthogonalization, enabling the model to retain early rhythmic structures while adapting to new follow-up data.

[0014] Furthermore, the parameter learning strategy also includes a regularization term protection sub-strategy and an orthogonal component calculation sub-strategy. The regularization term protection sub-strategy includes calculating the importance values ​​of the parameters and constructing parameter protection regularization terms. The orthogonal component calculation sub-strategy includes calculating the subspace matrix and calculating the corresponding orthogonal components based on the subspace matrix. Through selective protection at the parameter level and orthogonal constraints on the gradient space, the catastrophic forgetting problem commonly encountered in incremental training is significantly reduced. High-performance knowledge retention ensures the stability of the model's output throughout the follow-up period, reduces the risk of false alarms and missed diagnoses, and improves clinical reliability.

[0015] Furthermore, the follow-up ECG data includes device type sub-data, acquisition parameter sub-data, acquisition content sub-data, and environmental status sub-data. The device type sub-data reflects the type of acquisition device at the time of acquisition, the acquisition parameter sub-data reflects the acquisition parameters at the time of acquisition, the acquisition content sub-data reflects the waveform information at the time of acquisition, and the environmental status sub-data reflects the information on environmental influencing factors at the time of acquisition.

[0016] Furthermore, the noise processing sub-strategy includes filtering the ECG signal in the acquired content sub-data using a high-pass filter to obtain a first filtered signal, filtering the first filtered signal using a notch filter to obtain a second filtered signal, and removing outliers in the second filtered signal using an outlier removal algorithm to obtain denoised follow-up ECG data.

[0017] Further: The normalization sub-strategy includes calculating the mean and standard deviation of the ECG signal to make the amplitude of the ECG data lie in a uniform data range to obtain a standard ECG signal; obtaining each sub-data in the follow-up ECG data to obtain a mapping sub-vector from a preset corresponding embedding table, concatenating the mapping sub-vectors to generate a domain embedding vector, and concatenating the domain embedding vector and the standard ECG signal to generate the standard input tensor.

[0018] The main technical advantages of this invention are reflected in the following aspects: This invention processes different follow-up ECG information by constructing an ECG representation network to automatically output corresponding risk prediction results at different stages; simultaneously, it enables the model to obtain stable and transferable ECG feature representations under multi-source and multi-device conditions. Cross-domain and cross-device capabilities allow the model to work stably in real multi-center clinical environments without needing to retrain for each medical center, significantly reducing deployment difficulty and cost. This invention does not rely on historical raw ECG data or require retraining of the complete model; it can achieve controlled updates using a small amount of follow-up data. Attached Figure Description

[0019] Figure 1 This is a flowchart of the dynamic electrocardiogram feature modeling and recurrence prediction method for follow-up after catheter ablation of atrial fibrillation proposed in this invention. Figure 2 A schematic diagram of a unified input coding structure consisting of multi-source ECG signal input and domain embedding; Figure 3 This is a schematic diagram illustrating feature extraction and cross-center consistency constraints based on a pre-trained ECG representation network. Figure 4 This is a schematic diagram of the feature drift detection mechanism during the follow-up period (showing the changes in the mean values ​​of electrocardiogram characteristics at different follow-up stages). Figure 5 This is a schematic diagram of a controlled update mechanism for a model based on parameter importance constraints and gradient orthogonality. Figure 6 This is a schematic diagram illustrating the on-demand update effect of the recurrence prediction model of this invention on the follow-up time axis. Detailed Implementation

[0020] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings, so that the technical solution of the present invention can be more easily understood and mastered.

[0021] A recurrence prediction system for long-term follow-up after atrial fibrillation catheter ablation is proposed. This system constructs a unified electrocardiogram (ECG) representation space, introduces a feature drift detection mechanism within the follow-up period, and combines it with a constrained parameter update strategy to achieve continuous model adaptation under multi-source ECG signals and long-term non-stationary conditions. Unlike existing static training models that rely on one-time data and struggle to keep up with post-operative ECG feature changes after deployment, this invention integrates the entire process from input encoding, feature learning, distribution change monitoring to parameter updates. This allows the recurrence prediction model to perform structured updates for new follow-up data without retaining historical raw data, while maintaining the ability to recognize early rhythm patterns. Specifically: A continuous learning prediction system for recurrence after atrial fibrillation ablation based on multi-source electrocardiogram data, characterized in that it includes a data acquisition module, a data processing module, a feature recognition module, and a risk prediction module; The data acquisition module is used to collect follow-up electrocardiogram (ECG) data. The follow-up ECG data includes device type sub-data, acquisition parameter sub-data, acquisition content sub-data, and environmental status sub-data. The device type sub-data reflects the type of acquisition device at the time of acquisition, the acquisition parameter sub-data reflects the acquisition parameters at the time of acquisition, the acquisition content sub-data reflects the waveform information at the time of acquisition, and the environmental status sub-data reflects the information of environmental influencing factors at the time of acquisition.

[0022] This implementation first establishes a follow-up data acquisition process. For patients undergoing atrial fibrillation catheter ablation, fixed follow-up time points are set according to routine clinical follow-up plans, such as the first week, first month, third month, and sixth month post-procedure. This follow-up time point design not only aligns with clinical trials—frequent monitoring in the early stages to detect early recurrence and gradually extending observation intervals in later stages—but also provides an ideal data distribution for the time-aware design of this invention. During each follow-up, a fixed-length electrocardiogram (ECG) signal is acquired: typically 10 seconds, corresponding to 500-5000 sampling points at different sampling rates. Simultaneously, the model of the acquisition device, the number of leads, the sampling frequency, the amplification factor, and the recording environment information are recorded. These signals and metadata are stored in the database along with the follow-up time. Due to the long time intervals between different follow-up points, these ECG records often exhibit gradual changes in statistical structure; this time-series characteristic is precisely the applicable scenario for the dynamic modeling method of this invention. Metadata entries include: Sampling frequency: 125 Hz, 250 Hz, 500 Hz, etc., affecting the signal's temporal resolution; Device model: Philips PageWriter, GE MAC 5500, etc., determining hardware characteristics and noise patterns; Lead configuration: 12 leads, 3 leads, single lead, etc., affecting the integrity of ECG characteristics; Gain coefficient: standard gain versus other gain settings, directly affecting signal amplitude; Filter information: high-pass and low-pass frequencies, parameterization of preprocessing; Acquisition environment: supine, standing, active, etc., affecting electrode contact quality and noise levels.

[0023] The data processing module includes a noise processing sub-strategy and a normalization sub-strategy. The noise processing sub-strategy is used to process the noise in the follow-up ECG data, and the normalization sub-strategy is used to normalize the follow-up ECG data to obtain a standard input tensor. The noise processing sub-strategy includes filtering the ECG signal in the acquired content sub-data using a high-pass filter to obtain a first filtered signal, filtering the first filtered signal using a notch filter to obtain a second filtered signal, and removing outliers from the second filtered signal using an outlier removal algorithm to obtain denoised follow-up ECG data. After acquiring the original signal, the system preprocesses each segment of the ECG signal and completes unified encoding. First, baseline drift correction and power frequency interference suppression are performed on the ECG signal to ensure waveform stability. Preprocessing includes: using a high-pass filter to remove slow changes in the ECG baseline, which is crucial for accurate detection of low-frequency components. Power frequency interference suppression: using a 50 Hz notch filter to remove power supply interference; the amplitude of typical power frequency interference can reach 20-30% of the original signal. Outlier detection: removing spikes caused by electrode loosening or electromyography interference using median filtering or an outlier removal algorithm.

[0024] The normalization sub-strategy includes calculating the mean and standard deviation of the ECG signal to make the amplitude of the ECG data lie in a uniform data range to obtain a standard ECG signal; obtaining each sub-data in the follow-up ECG data to obtain a mapping sub-vector from a preset corresponding embedding table, concatenating the mapping sub-vectors to generate a domain embedding vector, and concatenating the domain embedding vector and the standard ECG signal to generate the standard input tensor.

[0025] Then, based on the average of that segment... with standard deviation Amplitude standardization: This z-score normalization places signals from different devices within a uniform numerical range, facilitating neural network processing. Signals from different gain settings are mapped to the range [-3, 3], accelerating network convergence and improving feature consistency across devices, which is crucial for multi-center clinical applications.

[0026] Simultaneously, follow-up metadata is read, and discrete information such as device model, lead type, and sampling rate are mapped into fixed-length domain vectors through an embedding matrix. Specifically, each metadata attribute corresponds to an embedding table, and its ID is mapped to a 128-dimensional continuous vector. Then, the vectors of each attribute are concatenated or summed to obtain the final domain embedding vector. .

[0027] The implementation of the domain embedding function g can be: in This represents the value of the j-th metadata attribute. It is the embedding function of this attribute. This design allows the model to explicitly perceive different acquisition conditions, thus taking these factors into account in feature extraction and drift detection.

[0028] The standardized ECG waveform is concatenated with the domain vector to form an input tensor with a unified structure. This process is in Figure 2 The diagram illustrates this process. Through this method, ECG signals from different devices and at different times are transformed into inputs of the same dimension, facilitating subsequent processing and enabling the model to learn device-invariant representational features.

[0029] The feature recognition module is configured with an electrocardiogram (ECG) representation network, which is used to identify the representation vectors in the standard input tensor to generate follow-up feature representations. The network construction module includes a temporal convolutional encoder, a rhythm-sensitive filter, and an adaptive aggregation strategy. After unified encoding is completed, the input tensor... The data will be fed into a pre-trained ECG representation network. This network consists of the following modules: Temporal Convolutional Encoder: Contains 3-4 layers of dilated convolutional blocks. Each block includes: Dilated Convolutional Layer: Kernel size 3, dilation factor increases layer by layer to increase the receptive field and capture features at different time scales. Batch Normalization Layer: Stabilizes the training process and accelerates convergence. Activation Function: ReLU. Dropout Layer: Regularization, dropout rate 0.2 to prevent overfitting. Rhythm-Sensitive Filter: A dedicated module inserted between convolutional blocks, learning the following ECG-related features: QRS Complex Detection: Learns high-frequency (100-200Hz) spike features. RR Interval Estimation: Learns low-frequency periodicity (0.5-3Hz). P-wave and T-wave Morphology: Learns morphological and amplitude features. The parameters of these filters are learned during pre-training, allowing them to automatically adapt to different patients. Adaptive Temporal Aggregation: After convolutional encoding, adaptive average pooling or attention mechanisms are used to aggregate the entire time series into a single vector, capturing global rhythmic features. Key features such as atrioventricular conduction delay, interval irregularities, and waveform morphology changes can be extracted from continuous electrocardiogram (ECG) waveforms. At the end of the network, this invention utilizes a pre-trained ECG representation network. Enhanced input A unified encoding is performed, where θ represents the set of learnable network parameters of the ECG representation network. The system obtains the representation vector through temporal pooling and linear mapping. This is used to describe the rhythm state of the currently followed node.

[0030] The final representation is further obtained through a lightweight linear mapping: Where W is the shape of The learnable linear transformation matrix, where For the output dimension of the pre-trained network, is the final representation dimension. The role of W is to project the high-dimensional intermediate representation of the pre-trained network onto a more compact representation space, which facilitates subsequent drift detection and classification. For shape A learnable bias vector is used to represent the translation transformation. The intermediate output of the pre-trained network contains rich representational information from the original electrocardiogram signal. This serves as the final electrocardiogram representation vector, used for subsequent recurrence risk assessment and drift detection.

[0031] The training process is as follows: 1. Data Preparation: Collect labeled ECG data from multiple medical centers, including ECG segments of different rhythm types, totaling no less than 100,000 samples. The data should cover different sampling rates, gain settings, and device models to ensure the generalization ability of the pre-trained network.

[0032] 2. Self-Supervised Pre-training: Pre-training is performed using a contrastive learning approach. Random data augmentation is applied to each original ECG signal, generating two augmented versions as positive sample pairs. The network is trained to make ECG recordings from the same patient at different times appear closer together in the representation space, while recordings from different patients appear farther apart. This self-supervised approach does not rely on detailed rhythm annotation, only patient ID correspondence, significantly reducing annotation costs. Encoding: Contains 3-4 temporal convolutional layers, each using dilated convolutions to increase the receptive field. The kernel size is 3, the stride is 1, and the number of channels increases from 64 to 256 layer by layer. Rhythm-sensitive filtering: Rhythm-specific filtering modules are inserted between convolutional layers. These modules learn ECG-related features such as QRS detection and RR interval estimation, using a bandpass filter design with learnable parameters to adapt to different patients' features. Pooling and aggregation: Adaptive average pooling is used to aggregate temporal sequences into a single vector, or an attention mechanism is used to highlight key moment features. Finally, a fully connected layer is used to obtain a representation vector of dimension d. .

[0033] 3. Loss Function and Optimization: A contrastive learning loss was used, with Adam as the gradient optimizer and an initial learning rate of 1e-3, which decayed by a factor of 0.1 when no improvement was made on the validation set. Training was conducted for at least 200 epochs until convergence. This learning rate was chosen based on: (a) the configuration of standard ECG pre-training such as ECGFM, (b) commonly used settings on medical data, and (c) testing at 1e-2, 1e-3, and 1e-4 during the initial debugging phase, with 1e-3 showing the best performance.

[0034] The network building module also includes consistency constraints.

[0035] To reduce feature bias caused by different centers or different devices, this invention incorporates cross-center consistency constraints during the training phase. In a batch containing samples from multiple medical centers, and these samples are in the same clinical state, consistency constraints are imposed on their representations. The purpose of constraints is to enable cross-domain alignment of the feature space during the training phase. Specifically, when two patients are in the same clinical state (e.g., both 3 months post-surgery and neither relapse-free), even if they come from different medical centers and use different equipment, their representation vectors... and It should also be close. Through this constraint, the model learns to ignore irrelevant information caused by differences in equipment and centers, while retaining essential features related to the relapse state.

[0036] The innovation of this feature alignment strategy in this invention lies in its dynamic consistency constraint after the learning of representations, rather than a fixed alignment at the input or convolutional layers. This allows the alignment strategy to adapt as the network learns. Furthermore, this constraint remains active during long-term follow-up learning, ensuring that newly learned features remain within a unified space compared to historical features. The cross-center alignment constraint ensures that the underlying representation space possesses a certain degree of invariance when the model is transferred from medical center A to center B, significantly reducing the risk of negative transfer.

[0037] Maintaining a relatively stable distribution structure of ECG signals of the same rhythm category in the representation space improves the robustness of the model during practical deployment. Cross-center alignment is particularly important in multi-center clinical applications because different centers may have different equipment, operators, and patient populations, and this constraint enables the model to learn robust representations of these differences. Figure 3 This demonstrates the overall process of aligning the feature extraction structure with cross-center features.

[0038] The update triggering module is configured with a dynamic threshold and a drift calculation sub-strategy. The drift calculation sub-strategy is used to calculate the current characterization drift based on the patient's follow-up characteristics. If the characterization drift is greater than the dynamic threshold, the update calculation module will work. The dynamic threshold is set as follows: Postoperative atrial structural remodeling follows a biological time process, with early stages mainly involving inflammatory responses and scar formation, mid-stages involving fibrosis development, and late stages tending towards stabilization. The drift detection mechanism Δ(t) of this invention directly corresponds to atrial electrical changes on this timescale, ensuring that parameter updates correspond to the clinical pathological process. Early threshold τ=0.5: corresponding to the rapid postoperative drift period (Δ(t) is approximately 0.8-1.2), frequently capturing signal changes caused by atrial remodeling.

[0039] The intermediate threshold τ=0.3 corresponds to the fibrosis development period (Δ(t) is approximately 0.4-0.6), and is updated appropriately to track gradual rhythmic changes.

[0040] The late threshold τ=0.2 corresponds to the stable period (Δ(t) is about 0.15-0.3), which keeps the model stable without excessive updates. The dynamic threshold corresponds to the clinical pathological process, avoiding updates at pathologically irrelevant time points, which is something that existing general ML methods have not considered at all.

[0041] The drift calculation sub-strategy includes averaging the representation vectors of all patients at the current follow-up node to obtain the follow-up mean, and calculating the difference between the current follow-up mean and the follow-up mean at the previous time point to generate the representation drift. The electrocardiogram (ECG) signals of atrial fibrillation patients exhibit significant periodic disruption characteristics, manifested as highly irregular RR intervals. This invention, by triggering updates based on changes in the mean of the ECG representation vector during feature drift detection, rather than a simple gradient threshold, enables the model to capture the dynamic changes in rhythm regularity recovery at different postoperative time points. This ECG-specific design ensures the model's persistent recognition of early rhythm patterns (such as subtle changes in QRS morphology).

[0042] After the representation vector of each follow-up node is calculated, this invention monitors feature drift during the follow-up process by calculating the change in the mean representation value between adjacent time nodes. The specific implementation method is as follows: At each follow-up node For all patients, the current representation vector ( (For patient indexing) Calculate the mean to obtain a statistical representation of that time point: in This refers to the total number of patients at this follow-up point (usually the number of patients in the batch, e.g., 10-50). Calculating the mean of the statistical characteristics, rather than the characteristics of individual patients, can eliminate the influence of individual differences, making the test results more stable and reliable.

[0043] The difference between this value and the mean of the previous follow-up node is then calculated as a measure of drift. This Euclidean distance intuitively quantifies the distance the population representation center has moved. The time series of representation drift reflects the evolution trend of the atrial rhythm state after surgery. When the drift remains small across multiple dimensions (Δ(t) between 0.1 and 0.3), it indicates that the rhythm structure is relatively stable and the model can continue to be used; when the drift increases significantly at a certain point... This indicates that the overall atrial rhythm has changed, and the old model may no longer be suitable for the current data distribution.

[0044] Figure 4 The diagram illustrates the temporal relationship of drift calculation and the change of drift amount with follow-up time, clearly presenting the drift characteristics at different postoperative periods.

[0045] The update calculation module includes a parameter learning strategy, which is used to update the learned parameters of the ECG representation network. The parameter learning strategy includes a parameter update representation, which is as follows: Where η is the preset learning rate and λ is the preset regularization weight. For the first The orthogonal components corresponding to the term parameters, For the first The significance value of the parameter in the previous time step. For the first The parameter value of the item. For the first The parameter value of the item at the previous moment.

[0046] The parameter learning strategy also includes a regularization term protection sub-strategy and an orthogonal component calculation sub-strategy. The regularization term protection sub-strategy includes calculating the importance values ​​of the parameters and constructing parameter protection regularization terms. In the previous follow-up phase (t-1), the importance of each parameter was calculated using all data from that phase. For large-scale parameters, the Fisher Information matrix was used as an approximation. Where N is the total number of samples in the previous stage (usually 20-50), and n is the sample index. This represents the relapse prediction task loss calculated for the nth follow-up sample, specifically in the form of binary cross-entropy loss. This calculation can be completed in the previous stage and stored as a parameter importance matrix, without needing to retain the original data.

[0047] The orthogonal component calculation sub-strategy includes calculating the subspace matrix and calculating the corresponding orthogonal components based on the subspace matrix. Based on the gradient vector from the previous stage, the parameter subspace matrix of the old knowledge is calculated. Specifically: Collect the gradient vectors of all training samples from the previous stage: each This is the complete gradient vector. Perform singular value decomposition on the gradient matrix: , where the columns of G are gradient vectors (n). The first k singular vectors are selected to form... (k is usually taken as 50%-80% of the total vector dimension. For example, if the total dimension is 10000, then k=5000-8000). The column vectors spanned become the key directions for historical learning, corresponding to the most important gradient directions in the early follow-up data. This SVD-based method is computationally intensive, but only needs to be executed once when an update is triggered. In the new follow-up stage t, the current gradient vector g is orthogonally projected: The geometric meaning of this projection operation is: removal The projection onto the column vector space P preserves the components orthogonal to P. Computationally, get Coordinates in the P-line space It is the result of projecting back into the original space. These are orthogonal components.

[0048] The update calculation module includes a parameter learning strategy, which is used to update the learned parameters of the ECG representation network. The parameter learning strategy includes a parameter update representation, which is as follows: Where η is the preset learning rate (recommended to be 0.05-0.1 times the original training learning rate 1e-3, i.e., 1e-4), and λ is the preset regularization weight (recommended 0.4-0.6). For the first The orthogonal components corresponding to the term parameters, For the first The significance value of the parameter in the previous time step. For the first The parameter value of the item. For the first The parameter value of the term at the previous time step. This update rule considers two protection mechanisms at the same time: (1) the gradient orthogonal projection g^⊥ protects the key directions of early learning, and (2) the regularization term protects important parameters.

[0049] To prevent the model from forgetting diagnostically significant structural information from early follow-up phases during updates, this invention strengthens the protection of historical key parameters through parameter importance weights, ensuring they remain close to their original values ​​during updates and preventing the model from losing previously learned recurrence discrimination paths. Simultaneously, to reduce gradient conflicts between new follow-up nodes and old tasks, this invention employs gradient orthogonalization, projecting the gradient of the current follow-up node onto a subspace orthogonal to the gradient of the old task, allowing the model to adapt to new data with minimal interference.

[0050] In this process, the model only needs to use a small amount of data (20-50 records) from the current node to complete the update, without having to revisit the original historical ECG signals, which greatly reduces the difficulty of data management and is particularly beneficial for complying with regulations on patient privacy protection. Figure 5 This demonstrates the overall structure of parameter importance protection and gradient orthogonalization updates. This enables the model to continuously adapt to pathological changes with minimal cost during follow-up. Figure 6 This illustrates how the model updates on demand based on drift at different follow-up stages, clearly demonstrating the dynamic and adaptive characteristics of the entire system.

[0051] The risk prediction module is used to generate based on follow-up characteristics. .

[0052] At the output layer, this invention utilizes linear prediction head pairs for characterization. Perform recurrence risk estimation and calculate the predicted probability: Where σ is the sigmoid function. The prediction head is a simple linear layer, which facilitates rapid adjustment during incremental learning.

[0053] Binary cross-entropy is used as the task loss: in ∈ {0, 1} is a sample The labels are 0, which indicates no recurrence during the postoperative follow-up period, and 1, which indicates recurrence.

[0054] The overall optimization objective consists of the following three weighted combinations: Here, α = 0.5 controls the strength of the feature consistency constraint, ensuring feature alignment of data from different centers during initial training; β = 0.3 controls the strength of the parameter protection regularization term, balancing the knowledge fusion of new and old tasks. All three parameters participate in optimization during the initial training phase. During the incremental learning phase, when new follow-up data arrives, only... and constrained Participate in optimization, and The weights can be selectively reduced to α = 0.2, because the feature space is already basically aligned at this point, and no mandatory constraints are needed.

[0055] Of course, the above are just typical examples of the present invention. In addition, the present invention may have many other specific embodiments. All technical solutions formed by equivalent substitution or equivalent transformation fall within the scope of protection claimed by the present invention.

Claims

1. A recurrent learning prediction system for atrial fibrillation ablation postoperative recurrence based on multi-source electrocardio data, characterized in that: The system comprises a data collection module, a data processing module, a feature recognition module, and a risk prediction module. The data collection module is configured to collect follow-up electrocardiogram data. The data processing module comprises a noise processing sub-strategy and a normalization sub-strategy. The feature recognition module is configured with an electrocardiogram representation network. The risk prediction module is configured to generate prediction recurrence information based on the follow-up feature representation.

2. The recurrent learning prediction system for atrial fibrillation ablation postoperative recurrence based on multi-source electrocardio data according to claim 1, wherein: The system further comprises an update trigger module and an update calculation module. The update trigger module is configured with a dynamic threshold and a drift calculation sub-strategy. The update calculation module comprises a parameter learning strategy.

3. The recurrent learning prediction system for atrial fibrillation ablation postoperative recurrence based on multi-source electrocardio data according to claim 2, wherein: The network construction module comprises a time series convolutional encoder, a rhythm-sensitive filter, and an adaptive aggregation strategy.

4. The recurrent learning prediction system for atrial fibrillation ablation postoperative recurrence based on multi-source electrocardio data according to claim 3, wherein: The network construction module further comprises a consistency constraint condition.

5. The recurrent learning prediction system for atrial fibrillation ablation postoperative recurrence based on multi-source electrocardio data of claim 2, wherein: The drift calculation sub-strategy comprises calculating the mean of the representation vectors of all patients at the current follow-up node to obtain a follow-up mean, and calculating the difference between the current follow-up mean and the follow-up mean at the previous time to generate the representation drift.

6. The recurrent learning prediction system for atrial fibrillation ablation postoperative recurrence based on multi-source electrocardio data of claim 1, wherein: The parameter learning strategy comprises a parameter update representation, and the parameter update representation is: wherein, η is a preset learning rate, λ is a preset regularization weight, the important value of the i-th parameter at the previous moment, the parameter value of the i-th parameter, the important value of the i-th parameter at the previous moment, the parameter value of the i-th parameter, the parameter value of the i-th parameter, the parameter value of the i-th parameter at the previous moment, the parameter value of the i-th parameter at the previous moment, the parameter value of the i-th parameter at the previous moment.

7. The recurrent learning prediction system for atrial fibrillation ablation postoperative recurrence based on multi-source electrocardio data of claim 6, wherein: The parameter learning strategy further comprises a regularization term protection sub-strategy and an orthogonal component calculation sub-strategy.

8. The recurrent learning prediction system for atrial fibrillation ablation postoperative recurrence based on multi-source electrocardio data of claim 1, wherein: The follow-up electrocardiogram data comprises device type sub-data, collection parameter sub-data, collection content sub-data, and environment state sub-data.

9. The recurrent learning prediction system for atrial fibrillation ablation postoperative recurrence based on multi-source electrocardio data of claim 8, wherein: The noise processing sub-strategy comprises filtering the electrocardiogram signal in the collection content sub-data through a high-pass filter to obtain a first filtered signal, filtering the first filtered signal through a notch filter to obtain a second filtered signal, and removing outliers in the second filtered signal through an outlier removal algorithm to obtain denoised follow-up electrocardiogram data.

10. The recurrent learning prediction system for atrial fibrillation ablation postoperative recurrence based on multi-source electrocardio data of claim 9, wherein: The normalization sub-strategy comprises calculating the mean and standard deviation of the electrocardiogram signal to normalize the amplitude of the electrocardiogram data within a unified data range to obtain a standard electrocardiogram signal, obtaining each sub-data in the follow-up electrocardiogram data to obtain a mapping sub-vector from a pre-set corresponding embedding table, and concatenating the mapping sub-vectors to generate a domain embedding vector, and concatenating the domain embedding vector and the standard electrocardiogram signal to generate the standard input tensor.

Images