Atrial fibrillation postoperative recurrence prediction method fusing electrocardiosignals and clinical features

Abstract

The invention provides an atrial fibrillation postoperative recurrence prediction method fusing electrocardiosignals and clinical characteristics. The method comprises the following steps: acquiring data of a patient before an ablation operation, and carrying out resampling, denoising and normalization preprocessing and data segment segmentation on an electrocardiosignal; extracting spatio-temporal features by using a deep network containing a residual convolutional block and a long and short term memory module; screening high-discrimination clinical baseline features through statistical analysis and a machine learning model; extracting time-frequency domain and nonlinear features of short-time heart rate variability; designing a cross-modal attention fusion module to carry out feature adaptive weighted fusion; and outputting a recurrence probability through a multi-layer perceptron based on the fusion features. The method improves the prediction precision through feature complementarity, facilitates the recognition of high-recurrence-risk patients, is suitable for sinus heart rhythm signals or atrial flutter and atrial fibrillation signals, and has a certain application value in the field of cardiovascular precision medical treatment. The method can be popularized to all prediction researches based on the electrophysiological signals.

Description

Technical Field

[0001] The present invention relates to a method for predicting the recurrence of atrial fibrillation after surgery by integrating electrocardiogram signals with clinical characteristics. Background Art

[0002] Atrial fibrillation (AF) is one of the most common complex cardiac arrhythmias in clinical practice, and its global prevalence is increasing significantly with the aging population. Although catheter ablation is currently the mainstay of AF treatment, its long-term success rate is limited, placing a heavy burden on healthcare systems and patients' families.

[0003] Existing methods for predicting postoperative recurrence still face many technical bottlenecks. Traditional prediction methods mostly rely on a single indicator, which makes it difficult to accurately reflect the risk of recurrence. In addition, the significant individual differences between patients in clinical practice (such as age, underlying diseases, living habits, etc.) increase the difficulty of building a unified prediction model. The clinical scoring system is currently a commonly used tool for assessing the risk of atrial fibrillation recurrence in clinical practice, but due to the lack of capturing dynamic features and the lack of full integration of electrophysiological characteristics (such as the spatiotemporal information of ECG signals), it is difficult to comprehensively assess the patient's recurrence risk. Although baseline characteristics can reflect the overall condition of the patient, the prediction model that relies solely on such static indicators has insufficient predictive performance and is difficult to meet the needs of precision medicine.

[0004] Surface electrocardiogram (ECG), a key noninvasive diagnostic tool, is widely used in predicting recurrence after atrial fibrillation ablation. With the rapid development of artificial intelligence (AI), deep learning models have demonstrated superior performance in processing complex data compared to traditional analytical methods. End-to-end approaches can capture dynamic information about atrial electrical remodeling and excel in complex pattern recognition. However, the single nature of the data modality leads to inadequate information utilization and overlooks the long-term impact of clinical variables on a patient's recurrence risk.

[0005] Multimodal fusion uses multiple modalities to observe the same phenomenon, capturing complementary information that cannot be presented by a single modality. In recent years, it has also shown great potential in the diagnosis and risk assessment of cardiovascular diseases. However, traditional fusion strategies often use simple feature splicing or decision-level voting, without designing efficient interaction mechanisms for the heterogeneity of different modalities. Direct fusion of discrete variables in clinical baseline data (such as medical history classification) with ECG signals is prone to introducing noise and leading to the curse of dimensionality. In addition, existing decision-level fusion methods (such as weighted averaging) cannot dynamically calibrate the contribution of each modality to the prediction results, resulting in the inability to fully explore the complementary information of electrophysiological characteristics and clinical indicators. Summary of the Invention

[0006] The purpose of this invention is to propose a method for predicting postoperative recurrence of atrial fibrillation that integrates electrocardiogram (ECG) signals with clinical features. This method utilizes a one-dimensional convolutional neural network to model long-term dependencies in ECG signals, fully extracting characteristic information related to postoperative recurrence. Statistical analysis and machine learning models are then used to screen for highly discriminative clinical baseline features. A hierarchical attention mechanism is then employed to integrate electrophysiological signals, clinical baseline features, and short-term heart rate variability indicators to accurately predict the risk of recurrence after atrial fibrillation ablation, providing quantitative support for the development of personalized treatment plans.

[0007] The present invention proposes a method for predicting recurrence of atrial fibrillation after surgery by integrating electrocardiogram signals and clinical characteristics. The specific steps are as follows:

[0008] (1) Obtaining the patient's pre-ablation data, including original ECG signals and clinical baseline information, wherein the original ECG signals are surface ECG signals, and the acquisition time is not less than one minute;

[0009] (2) Preprocessing the original ECG signal obtained in step (1), taking into account the different configurations of the acquisition equipment, resampling the original ECG signal to a fixed frequency, removing myoelectric interference and baseline drift through denoising, completing amplitude normalization processing, and obtaining an ECG signal after resampling, denoising and normalization processing; segmenting the original ECG signal into data segments through a sliding time window, thereby obtaining clean ECG signal segments;

[0010] (3) For the ECG signal after resampling, denoising and normalization in step (2), the Pan-Tompkins algorithm is used to detect and locate the R wave peak, and the time interval between consecutive R wave peaks is obtained to obtain the RR interval sequence; the R wave peak detection and positioning selects the standard II lead when the ECG signal contains multiple leads;

[0011] (4) For the ECG signal segments obtained in step (2), a convolutional neural network including residual connections and long short-term memory modules is constructed to extract spatiotemporal features and output ECG feature vectors; if the sample size of the effective ECG signal segments used for model training is insufficient to support parameter optimization, a data enhancement strategy is implemented for the ECG signal segments;

[0012] (5) For the clinical baseline data obtained in step (1), candidate features associated with postoperative recurrence are obtained through univariate statistical analysis, feature importance is calculated using a machine learning model, the top N clinical features are retained, and an N-dimensional baseline feature vector is generated, where N is 20%-50% of the total number of candidate features;

[0013] (6) For the RR interval sequence obtained in step (3), extract the time domain, frequency domain and nonlinear heart rate variability (HRV) features to generate an HRV feature vector;

[0014] (7) The ECG feature vector obtained in step (4), the N-dimensional baseline feature vector obtained in step (5), and the HRV feature vector obtained in step (6) are fused using a hierarchical attention mechanism, and the feature weights of each modality are calculated through the intra-modality attention layer:

[0015]

[0016] Among them: F m represents the eigenvector of the mth mode, W m is the trainable weight matrix, U m is a trainable mapping matrix;

[0017] Then establish feature association through the cross-modal interaction layer:

[0018]

[0019] Among them: F i 、F j 、F k Represent the eigenvectors of the i-th, j-th, and k-th modes, respectively, k = 1, 2, ..., M, M is the total number of modes, is the interaction weight matrix, Represents a d×d-dimensional matrix space, where d is the feature dimension; the final output is the weighted fused feature vector;

[0020] (8) After obtaining the weighted fusion feature vector in step (7), the network parameters are optimized by five-fold cross-validation among patients, and the optimal model is determined using the AUC value as the evaluation index, providing an optimized model basis for subsequent training;

[0021] (9) The fused feature vector obtained in step (7) is input into the multilayer perceptron to complete the end-to-end training and output the recurrence probability.

[0022] In the present invention, step (2) is to remove other interferences in the process of ECG signal acquisition, laying a good foundation for subsequent signal analysis and feature extraction.

[0023] In the present invention, the network structure combining the residual connection and the long short-term memory module in step (4) can alleviate the gradient vanishing problem, enhance the model stability, have stronger feature learning ability, and is suitable for processing complex one-dimensional electrocardiogram signals.

[0024] In the present invention, the baseline characteristics screened in step (5) include but are not limited to atrial fibrillation course, hypertension, age, coronary heart disease and hyperlipidemia, which are closely related to the occurrence of atrial fibrillation events and adverse outcomes after atrial fibrillation ablation.

[0025] In the present invention, the short-term heart rate variability feature extracted in step (6) is more suitable for rapid and efficient prediction needs in clinical scenarios and can reflect the patient's autonomic nervous system regulation function.

[0026] In the present invention, the hierarchical attention mechanism fusion in step (7) can further mine feature information related to postoperative recurrence of atrial fibrillation, realize dynamic weighted fusion of multi-dimensional features, effectively suppress redundant information (such as non-specific clinical indicators) and strengthen complementary information, and realize the deep fusion of electrophysiological features and clinical features.

[0027] In the present invention, a five-fold cross-validation method is used in step (8) to optimize the network hyperparameters, ultimately forming an efficient prediction model. Each training set is used to generate and maintain the inference model, and the validation set is used to reflect the performance indicators of the multimodal fusion model on an unknown data set.

[0028] The present invention has the following beneficial effects:

[0029] 1. Compared with traditional clinical scoring systems, this invention breaks through the bottleneck of single modality performance through multimodal information complementarity, which is conducive to solving the problem of insufficient prediction accuracy under complex pathological mechanisms.

[0030] 2. The ECG signal used in the present invention is suitable for sinus rhythm signals or atrial flutter and atrial fibrillation signals. At the same time, it shortens the HRV signal detection time from the traditional five minutes or more to one minute, which helps to improve clinical detection efficiency and reduce the difficulty of patient cooperation in examinations.

[0031] 3. The present invention adopts the attention mechanism to dynamically evaluate the contribution of each modality. Doctors can judge the impact of electrophysiological abnormalities, disease duration, and autonomic nervous system imbalance on the recurrence of specific patients through weight distribution, providing an explainable basis for the formulation of individualized ablation strategies.

[0032] 4. The method of the present invention supports the dynamic expansion of the fusion framework of other modalities (such as echocardiography, intracardiac electrogram, etc.), and can quickly adapt to the newly added data types by adjusting the input dimension of the attention mechanism.

[0033] 5. The technical solution of the present invention has cross-scenario universality and can be extended to all predictive research fields based on electrophysiological signals, providing a methodological reference for similar research. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 Schematic diagram of the original surface electrocardiogram (ECG) signal and the preprocessed ECG signal; (a) is the original signal collected by the device; (b) is the clean ECG signal after preprocessing.

[0035] Figure 2A schematic diagram of a system for predicting recurrence of persistent atrial fibrillation after surgery by fusing electrocardiogram signals and clinical characteristics in Example 1 of the present invention is shown.

[0036] Figure 3 A neural network architecture for electrocardiogram signal feature extraction in Example 1 of the present invention is provided, which includes residual connections and long short-term memory modules.

[0037] Figure 4 It is a visual depiction of HRV feature extraction; (a) is the distribution of RR intervals, (b) is the frequency domain power spectral density distribution, and (c) is the Poincare scatter plot.

[0038] Figure 5 The ROC curves obtained by the recurrence prediction model in Example 1 of the present invention under different modal inputs are shown. DETAILED DESCRIPTION

[0039] The method and application of the present invention are further described below with reference to the accompanying drawings and examples. These embodiments do not limit the present invention; any structural, methodological, or functional changes made by a person skilled in the art based on these embodiments are within the scope of protection of the present invention.

[0040] Example 1: The method for predicting recurrence of atrial fibrillation after ablation that integrates ECG signals and clinical characteristics of the present invention is applied to patients with persistent atrial fibrillation. This example is based on a cohort of 75 patients with persistent atrial fibrillation admitted to a tertiary hospital from 2018 to 2020, including 28 patients with postoperative recurrence (recurrence is defined as the occurrence of atrial fibrillation, atrial flutter or atrial tachycardia for ≥30 seconds 3 months after ablation). The data includes preoperative 12-lead atrial fibrillation ECG signals and clinical baseline data. The specific steps of the method for predicting recurrence of atrial fibrillation after ablation that integrates ECG signals and clinical characteristics of the present invention are as follows:

[0041] (1) Figure 1 The original ECG signal shown in (a) is first resampled to 500Hz to eliminate the frequency deviation caused by equipment differences; then the wavelet transform is used to remove the noise such as electromyographic interference and baseline drift in the original ECG signal; then zero-mean normalization is applied; finally, the ECG signal is segmented using the sliding window method, with a window length of 10 seconds and a step size of 50% overlap (i.e., 5 seconds) to obtain a clean and evenly segmented ECG signal, as shown in Figure 1. Figure 1 (b)

[0042] (2) Using the Pan-Tompkins algorithm, the R wave peaks in the preprocessed ECG signal (selecting standard lead II) are detected and located, and the time intervals between consecutive R peaks are obtained to obtain the RR interval sequence;

[0043] (3) Extract the feature representation related to postoperative recurrence from the clean ECG signal fragment obtained in step (1) and output the ECG feature vector; the feature extraction model mainly includes two modules, a convolution module based on residual connection and a long short-term memory network module. The model framework for extracting ECG signal features is as follows: Figure 3 As shown, the 16,5×1 residual module means that the convolution kernel K size is 5 and the number of channels is 16 in the residual module;

[0044] (4) Perform univariate statistical analysis on clinical baseline data, and then use the random forest model to calculate feature importance, retain the top ten features, and generate a baseline feature vector, including atrial fibrillation duration, age, etc. In order to alleviate the problem of sample imbalance in the dataset, the SMOTE algorithm is used for oversampling;

[0045] (5) Using the RR interval sequence obtained in step (2) to extract short-term HRV features and generate an HRV feature vector; Figure 4 This is a visualization of HRV feature extraction. (a) is the distribution of RR intervals. Time domain analysis directly measures and statistically analyzes RR intervals in time series, reflecting the overall variability of heart rate. Frequency domain analysis converts the time series of RR intervals into the frequency domain through methods such as fast Fourier transform for analysis. Figure 4 (b) is the power spectral density distribution in the frequency domain. Figure 4 (c) shows a Poincare scatter plot. Entropy is a measure of signal complexity and is widely used as a diagnostic tool in biomedicine.

[0046] (6) The ECG feature vector output from step (3) of this embodiment, the baseline feature vector obtained from step (4), and the HRV feature vector obtained from step (5) are fused, and the contribution of each modality to the final prediction is dynamically calibrated by introducing a learnable attention weight module. An attention unit based on a gating mechanism is used to calculate the correlation weight between modalities, thereby achieving adaptive weighted fusion at the feature level, effectively capturing the intrinsic correlation between different data and suppressing the interference of redundant information;

[0047] (7) In the decision-making stage, the optimized fusion representation is input into a multi-layer perceptron containing batch normalization and dropout layers, and the accurate prediction of the risk of recurrence after surgery for persistent atrial fibrillation is completed in an end-to-end manner. The recurrence prediction accuracy, F1 score and AUC of this embodiment on the persistent atrial fibrillation patient dataset are shown in Table 1. The higher these indicators are, the better the performance is. The results were obtained through a five-fold cross-validation method. Single-modality and dual-modality inputs were compared as ablation experiments, and the results showed that multimodal fusion can achieve higher performance indicators.

[0048] Table 1 is a comparison table of the performance of postoperative recurrence prediction on the dataset of patients with persistent atrial fibrillation. The evaluation indicators are accuracy, F1 score and AUC.

[0049] Table 1:

[0050]

[0051] (8) The performance test results of this embodiment are as follows Figure 5 As shown: Figure 5 The figure is an ROC curve diagram under different modal inputs. The results shown in the figure are the average curves of the test set obtained based on five-fold cross-validation. The horizontal axis represents the average false positive rate, the vertical axis represents the average true positive rate, and the AUC represents the area under the receiver operating characteristic curve. The results show that the prediction model constructed by fusing ECG signals and clinical characteristics in this embodiment not only improves the recurrence prediction ability, but also better balances sensitivity and specificity. This embodiment shows that the method of the present invention, by fusing data from different modalities before atrial fibrillation ablation, fully considers various factors that affect postoperative recurrence, and has a significant improvement in prediction performance compared to a single modality.

Claims

1. A method for predicting recurrence of atrial fibrillation after surgery by integrating electrocardiogram signals and clinical characteristics, characterized in that: The specific steps are as follows: (1) Obtaining the patient's pre-ablation data, including original ECG signals and clinical baseline information, wherein the original ECG signals are surface ECG signals, and the acquisition time is not less than one minute; (2) Preprocessing the original ECG signal obtained in step (1), taking into account the different configurations of the acquisition equipment, resampling the original ECG signal to a fixed frequency, removing myoelectric interference and baseline drift through denoising, and completing amplitude normalization processing to obtain an ECG signal after resampling, denoising and normalization; The original ECG signal is segmented by sliding the time window to obtain clean ECG signal segments; (3) For the ECG signal after resampling, denoising and normalization in step (2), the Pan-Tompkins algorithm is used to detect and locate the R wave peak, and the time interval between consecutive R wave peaks is obtained to obtain the RR interval sequence; the R wave peak detection and positioning selects the standard II lead when the ECG signal contains multiple leads; (4) For the ECG signal segments obtained in step (2), a convolutional neural network including residual connections and long short-term memory modules is constructed to extract spatiotemporal features and output ECG feature vectors; if the sample size of the effective ECG signal segments used for model training is insufficient to support parameter optimization, a data enhancement strategy is implemented for the ECG signal segments; (5) For the clinical baseline data obtained in step (1), candidate features associated with postoperative recurrence are obtained through univariate statistical analysis, feature importance is calculated using a machine learning model, the top N clinical features are retained, and an N-dimensional baseline feature vector is generated, where N is 20%-50% of the total number of candidate features; (6) For the RR interval sequence obtained in step (3), extract the time domain, frequency domain and nonlinear heart rate variability (HRV) features to generate an HRV feature vector; (7) The ECG feature vector obtained in step (4), the N-dimensional baseline feature vector obtained in step (5), and the HRV feature vector obtained in step (6) are fused using a hierarchical attention mechanism, and the feature weights of each modality are calculated through the intra-modality attention layer: Among them: F m represents the eigenvector of the mth mode, W m is the trainable weight matrix, U m is a trainable mapping matrix; Then establish feature association through the cross-modal interaction layer: Among them: F i 、F j 、F k Represent the eigenvectors of the i-th, j-th, and k-th modes, respectively, k = 1, 2, ..., M, M is the total number of modes, is the interaction weight matrix, Represents a d×d-dimensional matrix space, where d is the feature dimension; the final output is the weighted fused feature vector; (8) After obtaining the weighted fusion feature vector in step (7), the network parameters are optimized by five-fold cross-validation among patients, and the AUC value is used as the evaluation index to determine the optimal model, providing an optimized model basis for subsequent training; (9) The fused feature vector obtained in step (7) is input into the multilayer perceptron to complete the end-to-end training and output the recurrence probability.

2. The method according to claim 1, characterized in that The clinical baseline data described in step (1) include age, gender, BMI, duration of atrial fibrillation, smoking history, drinking history, history of comorbidities, and EHRA score, HAS-BLED score, and CHA2DS2-VASc score. The history of comorbidities can be collected in a structured manner through an electronic medical record system.

3. The method according to claim 1, characterized in that The denoising method in step (2) includes one or more of filter denoising, wavelet transform, empirical mode decomposition, independent component analysis or adaptive filtering.

4. The method according to claim 1, wherein The residual connection structure of the convolutional neural network in step (4) is: l+1 =x l +F(x l , W l ), where x l is the input of the lth layer, F(x l , W l ) is the output of the convolutional layer, and the gradient disappearance is avoided by identity mapping.

5. The method according to claim 1, wherein The data enhancement strategy in step (4) is one or more of random shift, random scaling, cutout or adding Gaussian noise.

6. The method according to claim 1, characterized in that The machine learning model in step (5) is one of random forest, XGBoost, LightGBM, AdaBoost, logistic regression or decision tree; the statistical analysis method uses independent sample t test or Mann-Whitney U test for normality for continuous variables, and uses chi-square test for categorical variables.

7. The method according to claim 1, characterized in that The time domain features in step (6) are SDNN (standard deviation of RR intervals), RMSSD (root mean square of adjacent RR intervals), and pNN50 (the proportion of adjacent RR interval differences greater than 50ms); the frequency domain features are LF (low-frequency spectrum power), HF (high-frequency spectrum power), and LF / HF (ratio of low and high frequencies); and the nonlinear features are SD1 (short-axis standard deviation obtained through the Poincare map), SD2 (long-axis standard deviation), ApEn (approximate entropy), and SampEn (sample entropy).

Images