Intradialytic hypotension (IDH) is a critical complication of hemodialysis (HD) and is associated with increased morbidity and mortality. However, accurate identification of patients at high risk for IDH remains challenging in routine clinical practice. Most existing prediction approaches rely on static clinical variables or intermittent measurements and provide limited insight into how IDH risk evolves across different stages of a dialysis session. We developed and validated a machine learning (ML) model that integrated wearable electrocardiography (ECG) and photoplethysmography (PPG) signals to enable time-resolved risk stratification of IDH and to identify clinically informative windows for patients undergoing HD. This retrospective study enrolled 401 maintenance HD patients from four Chinese centers. Patients were stratified into three categories: high-risk (IDH during session), moderate-risk (IDH within one-month follow-up), and low-risk (no IDH). Continuous ECG and PPG signals were acquired across three temporal windows: early-stage (0-40 min post-initiation), middle-stage (40-80 min), and late-stage (80-120 min). A Balanced Random Forest (RF) classifier with SMOTE balancing and Bayesian-optimized cost-sensitive framework was validated on 52 independent patients. The Balanced RF model showed optimal performance during the early-stage with an F1-score of 0.537, accuracy of 0.687, and high-risk recall of 0.667. Early-stage monitoring yielded the highest AUC (0.753) and AUPRC (0.446) for high-risk detection. Feature ablation revealed that PPG reduced false positives while ECG enhanced high-risk sensitivity. Early-stage wearable ECG and PPG monitoring enables effective three-level IDH risk stratification with clinically meaningful sensitivity, providing an optimal actionable window for preventive interventions in HD care.

Accurate upper-limb motor assessment is critical for post-stroke rehabilitation but relies on subjective clinical scales. This study proposes the Action-Aware Multimodal Wavelet Fusion Network (AMWFNet), integrating surface electromyography (sEMG) and robotic kinematics for automated Fugl-Meyer Assessment (FMA-UE)-aligned quantification. Continuous Wavelet Transform (CWT) converts heterogeneous signals into unified time-frequency scalograms. A learnable modality gating mechanism dynamically weights physiological and kinematic features, while action embeddings encode task contexts across 18 standardized reaching tasks. Validated on 40 participants (20 post-stroke, 20 healthy), AMWFNet achieved 94.68% accuracy in six-class classification, outperforming baselines by 9.17% (Random Forest: 85.51%, SVM: 85.30%, 1D-CNN: 91.21%). The lightweight architecture (1.27 M parameters, 922 ms inference) enables real-time assessment-training integration in rehabilitation robots, providing an objective, efficient solution.

Reconstructing Arterial Blood Pressure (ABP) signals from non-invasive sensors poses significant challenges for distributed Internet of Things (IoT) healthcare systems, particularly in real-time performance and cross-platform deployment. This study addresses these challenges with a complete, end-to-end framework combining optimized hardware, an intelligent algorithm, and a scalable system architecture. We introduce a compact, low-power IoT sensing platform (5.0 cm & times; 5.0 cm) for synchronized signal acquisition, paired with a novel two-stage algorithm (FACT-Net). The FACT-Net architecture incorporates parallel cross-hybrid modeling (PCHM) and serial hybrid CNN-Transformer (SHCT) processing for optimal spatiotemporal feature extraction. Validation on four independent datasets comprising 2,960 subjects in total demonstrates high accuracy (3.95 mmHg MAE; R2 = 0.89) and robust generalization (R2 = 0.86-0.92 across diverse datasets). The system is highly efficient, achieving 2.60 +/- 0.90 ms latency and 423 k samples/second throughput. Furthermore, structured pruning successfully reduced model parameters from 21.13 M to 10.33 M while maintaining accuracy. To support deployment, we developed the Cross-Platform Multi-Patient IoT framework. This system enables scalable use from individual home monitoring to multi-patient clinical environments, supporting both edge and cloud processing with dual-mode operation for enhanced reliability. This work contributes a complete, validated end-to-end solution for IoTbased physiological signal processing. It addresses critical challenges in real-time performance, scalability, and cross-platform deployment, bridging the gap between algorithmic research and practical healthcare analytics.

Ambulatory hemodynamic monitoring using photoplethysmography (PPG) holds promise for early detection and management of cardiovascular diseases. However, the increasing demand for automatic diagnosis in routine clinicals, coupled with the scarcity of well-labeled datasets, exacerbates significant challenges in developing learning-based tools with accessibility and explainability. To address these issues, this work proposed a reality-inspired pipeline for PPG generation and an Interpretable Hemodynamic pattern recognition framework with Multi-Instance Learning (IHemoMIL). Fluctuation patterns were introduced using anomaly injection and cardiovascular simulation response. This approach produced fine-grained, labeled synthetic PPG in a causality-informed manner for benchmarking. Multi-instance learning, with its aggregation process optimized by ranking-based mechanisms, was first applied to identify hemodynamic patterns and discriminative motifs within PPG waveforms under a weakly supervised paradigm, inherently providing salience-induced explainability. Extensive evaluations across nine datasets showcase IHemoMIL's superior interpretable and predictive performance compared to state-of-the-art methods, with ablation studies validating the effectiveness of key components. The compact yet effective IHemoMIL aligns with human intuition and integrates seamlessly into the diagnostic workflow, thereby contributing to improved patient outcomes and enhanced healthcare delivery. The implementation is available at URL#.

Bioelectrical signals are vital for non-invasive physiological monitoring and disease diagnosis. The application of artificial intelligence (AI) in automatic bioelectrical signal analysis has garnered attention. However, training supervised AI models requires abundant labeled data, which is particularly challenging in medical scenarios. Existing transfer learning (TL) approaches attempt to transfer knowledge from images in the computer vision (CV) domain, but face significant challenges due to dimensional difference and distribution gaps between CV and bioelectrical signal domains. Traditional domain adaptation (DA) assumes domain task invariance and focuses mainly on feature distribution difference, without considering label inconsistency and the complex cross-dimensional scenarios. Our motivation is to address the fundamental challenges in cross-domain knowledge transfer by jointly solving the dimensional gap and task discrepancy between CV and bioelectrical signal domains, thereby facilitating robust learning with limited labeled medical data. In this paper, we propose a cross-dimensional information transfer (CDIT) framework that enables effective information fusion through parallel encoding-decoding modules, which preserve the discriminative characteristics of both two-dimensional (2D) CV images and one-dimensional (1D) bioelectrical signals while mapping them into a shared feature space. Furthermore, we develop a cross-task DA method that synergistically integrates feature and label information to address the label inconsistency challenge in the DA phase of CDITF. We conduct extensive experiments on two representative scenarios using six databases. The experimental results demonstrate that CDITF with cross-task DA (CDITF-CTDA) achieves successful knowledge transfer from CV into bioelectrical signal domains despite their inherent difference, consistently outperforming baseline methods with improvements of 0.02-0.07 in AUC for bioelectrical signal analysis. These results demonstrate the effectiveness of CDITF-CTDA in leveraging CV knowledge for bioelectrical signal analysis under limited-labeled data medical scenarios.

Continuous blood pressure (BP) monitoring is crucial for health management, yet existing methods struggle with efficiency and adaptability in home and clinical environments. To address this, we propose the edge-based blood pressure estimation system (Edge-BP), an ultra-efficient wearable system for real-time, cuffless BP monitoring. First, we present the Cascaded Depthwise Separable Residual Network (CDS-Net), which employs a cascaded depthwise separable residual architecture and attention mechanisms to estimate BP from photoplethysmography (PPG) and electrocardiography (ECG) signals. Furthermore, we propose a progressive distillation-pruning framework, a model compression method that integrates dependency graph-guided structured pruning with dual-teacher knowledge distillation, substantially improving model compression efficiency. We also develop a wearable device for synchronized PPG and ECG acquisition, leveraging cross-database transfer learning to improve adaptability. The optimized model is deployed on a neural processing unit (NPU) and integrated with fourth-generation (4G) communication, enabling remote monitoring and automatic alerts for abnormal BP detection. CDS-Net exhibits superior performance in estimating systolic blood pressure (SBP) and diastolic blood pressure (DBP), with mean absolute error (MAE) 3.52 mm of mercury (mmHg) and 2.00 mmHg, respectively. Even after the model is compressed-reducing parameters by 94.4 % and computational complexity by 93.7 %- it still achieves MAEs 4.12 mmHg and 2.86 mmHg for SBP and DBP estimation. The testing results meet standards set by the British Hypertension Society and the Institute of Electrical and Electronics Engineers. This study provides a comprehensive solution for continuous BP monitoring in both home and clinical settings, paving the way for future advancements in wearable, physiological signal-based health management.

Detecting adverse physiological events within quasi-periodic cardiovascular rhythms is clinically critical, yet existing models remain hindered by black-box reasoning and their reliance on costly expert annotations. We address this gap with a label-free, interpretable model that delivers perceptible explainability to build clinical trust and structural adaptability for interactive collaboration. We propose the Persistence-Informed Score-based Multi Instance Learning (PISMIL) framework, which synergizes motif discovery and anomaly detection by exploiting the topological dynamics inherent to quasi-periodic cardiovascular signals. PISMIL employs persistent homology to extract invariants encoding motif stability across scales, thereby producing unsupervised priors that drive pointwise anomaly scoring. A confidence-weighted pseudo-labeling strategy converts these scores into supervision signals that guide an attention-based MIL detector, yielding continuous, well-demarcated cues that facilitate a verify-then-confirm clinical workflow. Extensive validation across three real-world ECG, two real-world PPG, and two synthetic PPG datasets demonstrates that PISMIL establishes a new benchmark in anomaly detection and motif localization, surpassing well-established baselines on diverse evaluation metrics. Ablation studies further confirm that topological guidance and MIL paradigm jointly enhance decision-boundary robustness, mitigate attention collapse, and preserve discriminative structural stability. By generating anomaly maps and identifying discriminative motifs for pre-annotation, PISMIL resolves the tension between annotation burden and clinical scalability, with the potential for seamless integration into ubiquitous cardiovascular monitoring workflows.

Objective: Multiple limitations exist in current automated ECG analysis, including insufficient feature integration across leads, limited interpretability, poor generalization, and inadequate handling of class imbalance. To address these challenges, we develop a novel dual-branch framework that comprehensively captures spatial-temporal features for cardiac disease diagnosis. Methods: ECG-AuxNet combines a Multi-scale Transformer Attention CNN for spatial feature extraction and a GRU network for temporal dependency modeling. A Dual-stage Cross-Attention Fusion module integrates features from both branches, while a Feature Space Reconstruction (FSR) auxiliary task is introduced as a manifold regularizer to enhance feature discrimination. The framework was evaluated on PTB-XL (15,709 ECGs) and validated in real-world clinical scenarios (SXMU-2k, 1,673 ECGs). Results: For class-imbalanced disease recognition (NORM, CD, MI, STTC), ECG-AuxNet attained 78.34% F1-score on PTB-XL and 82.63% F1-score on SXMU-2k, outperforming 9 baseline models. FSR significantly improved feature discrimination by 11.7%, enhancing class boundary clarity and classification accuracy. Grad-CAM analysis revealed attention patterns that precisely match cardiologists' diagnostic focus areas. Conclusion: ECG-AuxNet effectively integrates spatial-temporal features through auxiliary learning, achieving robust generalizability in cardiac disease diagnosis with interpretability aligned with clinical expertise.

The impressive performance of generalist large language models (LLMs) such as GPT-4 and Claude in healthcare raises a critical question: will domain-specific medical specialist models become obsolete? We argue that the future of medical artificial intelligence (AI) lies not in building monolithic medical foundation models, nor in replacing human expertise, but in orchestrating collaboration among generalist LLMs, domain-specific specialist models, and clinicians. We propose HetMedAgent, a heterogeneous medical multi-agent framework that enables conflict-aware evidence fusion, uncertainty-based clinician intervention triggering, and adaptive threshold calibration. Experiments on three real-world clinical decision-making tasks demonstrate that the synergy between generalist LLMs and domain-specific specialist models significantly outperforms using either type of model alone, validating the irreplaceable value of specialist models in modality-specific analysis. HetMedAgent drives a paradigm shift from building medical LLMs or foundation models to multi-agent collaboration, achieving a balance between general reasoning capabilities and domain-specific precision.

The physiological signals obtained from advanced sensors, combined with deep learning techniques for classification and regression tasks, have become a core driving force in enhancing smart healthcare. Recently, dense prediction tasks for physiological signals—aimed at generating predictions that are closely aligned with the input signal to enable fine-grained analysis—have garnered increasing attention. The UNet family, often combined with sophisticated task-specific customizations, has become a popular choice to improve prediction performance. However, pretrained Transformers have recently revolutionized deep learning due to their powerful transferability and effectiveness. In this work, we aim to harness the power of pretrained Transformers for dense prediction, eliminating the need for extensive task-specific architecture design. We propose a simple yet universal encoder-decoder architecture that utilizes a pretrained Transformer encoder and a lightweight convolutional Restormer decoder for dense prediction on physiological signals. To optimize the trade-off between model performance and computational efficiency, we incorporate knowledge distillation (KD). Our experiments focus on four representative dense prediction tasks: blood pressure waveform (BPW) estimation, PPG-to-ECG (P2E) reconstruction, denoising, and fiducial point localization. The results show that our proposed architecture outperforms state-of-the-art models, validating the potential of pretrained Transformers in enhancing physiological signal processing and medical diagnostics. This approach marks a significant step forward in optimizing both the performance and efficiency of dense prediction tasks.

Therapeutic robotic systems have emerged as reliable tools for physical rehabilitation, providing variableintensity movement assistance to patients with motor impairments. Robot-assisted rehabilitation facilitates the restoration mobility and dexterity, promotes functional neuroplasticity and potentially enables workforce reentry through training-induced cognitive and motor learning. To boost participant engagement and visuomotor coordination, we propose ArmGuider Pro, an advanced upper-limb training system that integrates hand-eye collaboration and gazetriggered assistance within rehabilitation-tailored serious games. The system implements intuitive eye-tracking and visualtriggering strategies to align therapeutic interventions with participants' intentional focus, incorporating immersive gaming elements and adaptive control algorithms. Experimental validation demonstrates significant activation in motor and cognitive cerebral cortex regions, enhanced visual attention concentration in desired target areas (25.92 % improvement), and improved trajectory adherence across sequential sessions (27.27 % improvement). By harnessing visual attention valence, our proposed system could encourage neuroplasticity, supporting its viability for clinical application and widespread adaption in rehabilitation regimens.

Cardiovascular disease remains a leading global health challenge, necessitating precise and continuous monitoring of blood pressure (BP) and cardiac function. Traditional noninvasive measurement techniques often present operational complexities and discomfort, while photoplethysmography (PPG) technology lacks the ability to capture comprehensive hemodynamic parameters. Herein, we introduce an innovative Internet of Medical Things (IoMT) framework for personalized hemodynamic assessment, driven by advanced flexible sensing technologies and multiscale modeling. Specifically, we propose a pulse wave detection system utilizing microfiber Bragg grating (mu FBG) sensors for comprehensive spatiotemporal monitoring. The system leverages a multiscale perception network (MSP-Net) to achieve precise BP estimation, with mean errors (MEs) and standard deviations (SDs) of -0.05 +/- 2.15 mmHg for systolic BP (SBP) and -0.19 +/- 3.48 mmHg for diastolic BP (DBP). The system's performance surpasses traditional PPG-based methods across multiple hemodynamic parameters. Leveraging pretraining and fine-tuning strategies, this study realizes personalized estimation models, furnishing technical support for precision medicine. Furthermore, we utilize a cloud-edge collaborative framework to achieve hemodynamic assessment across multiple IoMT devices via local area network and the RV1126 platform, providing real-time feedback, thereby establishing a closed-loop control system. Our study presents an innovative approach for the comprehensive assessment of cardiovascular function and the realization of personalized medicine, potentially serving as a promising solution for Healthcare 5.0 applications.

Despite the remarkable success of artificial intelligence (AI) in physiological signal analysis, the inherent variability between individuals poses significant challenges to model generalization. Existing personalization approaches typically rely on supervised fine-tuning of pre-trained general models (GMs) using labeled data from unseen subjects, which limits their practical deployment due to labeling costs and scalability issues. To address this challenge, we propose PULSE, a personalized unsupervised domain adaptation framework that enhances model generalization through self-adaptive learning. Our approach incorporates three key components: (1) an Adaptive Channel Selection and Embedding (ACSE) module that optimizes multi-channel signal processing through learnable attention mechanisms, (2) an Embedding-guided Representation Learning (ERL) strategy that enhances intra-class feature consistency during GM pre-training, and (3) a Self-adaptive Pseudo-label Enhancement (SPE) method that generates high-quality pseudo-labels to facilitate alignment between interdomain data distributions during GM fine-tuning. Extensive experiments on large-scale physiological datasets, including cross-database validation, demonstrate that PULSE achieves 2.8%-6.5% improvements in F1 score (from 91.5% to 95.2 to 98.0%) and 2.6%-6.4% improvements in accuracy (from 90.8% to 94.6% to 97.2%). The framework's effectiveness is validated through dynamic electrocardiogram analysis, showcasing its potential for broader applications in physiological signal processing. The code is publicly available at https://github. com/fdu-harry/PULSE.

Background: With the advancements in wearable technology, photoplethysmography (PPG) has emerged as a promising technique for detecting atrial fibrillation (AF) due to its ability to capture cardiovascular information. However, current deep learning-based methods has strict requirements on the quantity of labeled data. To overcome this limitation, we explore the performance of self-supervised contrastive learning in PPG-based AF detection. Methods: Our method initially utilizes 1,209 h of unlabeled PPG data from the VitalDB database, conducting selfsupervised pretraining using two contrastive learning frameworks, SimCLR and BYOL. Subsequently, the weights of the encoder are transferred and fine-tuned on a small amount of labeled PPG data to complete the AF detection task, including the selected MIMIC III, UMass, and DeepBeat datasets. In the realm of contrastive learning, we investigated seven data augmentation operations to explore their composite and preferred combinations, as well as the effects of double-sided and single-sided transformations. Results: Our research ultimately demonstrated that the preferred combination, incorporating single-sided transformation with the Drift operation, is most suitable for PPG data. Notably, even with only 1 %, 20 %, and 1 % of the training data from the three datasets used for fine-tuning, our approach achieves better F1 scores compared to supervised learning on the respective complete training sets. Additionally, on the 0.01 % DeepBeat training set, fine-tuning still showed a clear advantage over supervised learning. Conclusion: Appropriate self-supervised contrastive pretraining effectively leverages a substantial amount of existing unlabeled PPG data, thus reducing the reliance on labeled data for AF detection, and offering a possible solution to address the limitations posed by the scarcity of labels.

Vascular health, indicated by vascular age, is a key clinical marker. Traditional methods like PWV are complex, while PPG-based methods are more convenient but face challenges in generalization and adaptability. In response, this study proposes a Vascular Age Estimation Model (VAM), a parallel cross-modal hybrid network based on multi-source datasets, which integrates CNN and Transformer architectures to optimize the extraction and fusion of both local and global features. Experimental results on the VitalDB dataset demonstrate that the model achieves an average absolute error (MAE) of 7.66±0.39 years, root mean square error (RMSE) of 9.98±0.45 years, and coefficient of determination (R2) of 0.55±0.04. Compared to the baseline model, VAM reduces prediction errors by at least 20%. Furthermore, on an external test set, the model performs particularly well in the younger population (MAE=5.85 years, R2=0.63), with minimal impact of gender and BMI on estimation results (ΔMAE<0.5 years). Through interpretability analysis, this study confirms the crucial contribution of the diastolic peak in vascular age estimation. This research provides an efficient and interpretable solution for vascular age assessment, with significant clinical application potential.

Given the inadequacy of research in the realm of quality assessment (QA) for physiological signals (electrocardiogram and photoplethysmogram), the current reliance on coarse-grained assessment based solely on unimodal signals has become manifestly insufficient. Within the domain of wearable devices, there is a particularly pressing need for refined QA results pertaining to multi-source physiological signals. This paper endeavors to introduce a real-time fine-grained QA algorithm and, through real-time application demonstrations, showcase its efficacy. We have devised an attention network capable of simulating localized to global dependency relationships. By training individually, a corresponding QA model is assigned to each pattern. Fine-grained labels for the signals are allocated through weakly supervised learning. The proposed lightweight model has undergone systematic deployment, accompanied by the development of an interactive interface. The experiments indicate that the proposed fine-grained QA exhibits intuitive consistency with evaluations by human experts. Tests conducted on a mixed database demonstrate that the proposed methodology outperforms existing state-of-the-art methods. Furthermore, the algorithm has been deployed and integrated into practical applications, yielding outstanding results on self-constructed data. This study introduces a real-time fine-grained QA algorithm that exhibits a high level of visual consistency comparable to human experts, thereby enhancing the performance of multi-source information derivation tasks. The proposed system achieves real-time visualization in real-world environments, offering clinical experts a high-quality reference granularity for multi-source information.

BackgroundPathological upper-limb synergistic movement is a prevalent symptom of post-stroke motor dysfunction and pose a significant challenge in the rehabilitation of hemiplegia. However, the underlying mechanisms remain elusive, hindering the development of effective therapeutic strategies.ObjectiveThis study aims to explore the mechanisms underlying pathological synergic movements post-stroke by examining the interrelationship between motor modules associated with upper limb flexion synergy and the repetitive action of "taking food to the mouth."MethodsSurface electromyography (sEMG) was employed to capture the EMG signals of normal elbow flexion, the act of "taking food to the mouth," and post-stroke upper limb flexion synergy. Non-negative matrix factorization (NMF) was employed to compare and analyze the corresponding modular and coefficient matrices derived from these three motor tasks.ResultsThe modular matrix associated with flexion synergy exhibited the most significant correlation with the "taking food to the mouth" task, whereas the coefficient matrix shared the lowest correlation.ConclusionUpper limb flexion synergy after stroke may be related to the activation of the highly excitatory motor module formed by previously repeated movement of "taking food to mouth".

Early detection of widespread undiagnosed sleep apnea is crucial for preventing its severe health complications. However, large-scale diagnosis faces inaccessible monitoring and trust barriers in automated analysis, particularly due to the absence of transparent artificial intelligence frameworks capable of monitoring adaptation. Here, we develop Apnea Interact Xplainer, a transparent system enabling sleep apnea diagnosis through flexible channel analysis across clinical and home settings. Analyzing 15,807 polysomnography recordings from seven independent multi-ethnic cohorts, our system achieves accuracies of 0.738-0.810 for four-level severity classification, with 99.8% accuracy within one severity grade and R-squared of 0.92-0.96 for apnea-hypopnea index prediction on external test cohorts. The system provides multi-level expert-logic interpretable visualization of respiratory patterns enabling transparent collaborative decision-making. Notably, it achieves a sensitivity of 0.970 for early sleep apnea detection using only oximetry signals, while providing nightly risk assessment and intelligent monitoring reports. This study establishes a paradigm shift in advancing early and cost-effective sleep apnea diagnosis through transparent artificial intelligence.

The precise monitoring of hemodynamic parameters is crucial for cardiovascular health assessment and disease prevention. However, current noninvasive hemodynamic monitoring technologies fail to balance performance and efficiency. To overcome these challenges, this study introduces a dual-focus multitask fusion framework built on a sparse multigate mixture-of-experts network (SM(2)oE) and multiview photoplethysmography (PPG) for the accurate estimation of noninvasive hemodynamic parameters. This framework integrates a multiscale fusion architecture with a hybrid attention mechanism and incorporates an uncertainty regression loss function (UR-Loss) to enhance intertask information fusion and overall model performance. Through structured pruning and bidirectional knowledge distillation, the framework effectively reduces computational costs and ensures model efficiency. Cross-database evaluation demonstrates that the proposed method achieves blood pressure (BP) estimation performance falling within the numerical thresholds specified by IEEE, BHS, and AAMI guidelines. Under a lightweight strategy, the LSM(2)oE exhibits significant clinical application potential, with mean absolute errors for systolic BP, mean arterial pressure, diastolic BP, and heart rate being 4.02 mmHg, 2.99 mmHg, 3.00 mmHg, and 2.40 bpm, respectively. The parameter counts and FLOPs are 1.26 M and 0.15 Gmac, respectively, surpassing existing baseline models with only 4.0% of the parameter volume of the teacher model. Additionally, leveraging a cloud-edge collaboration framework, we achieved hemodynamic assessment across multiple Internet of Medical Things (IoMT) devices via a local area network and the RV1126 platform. The system demonstrates cross-scale analysis capability and real-time parameter estimation performance in IoMT environments, showcasing extensive application potential.

BackgroundThe long-term success rate of atrial fibrillation (AF) ablation remains a significant clinical challenge, particularly in patients with persistent atrial fibrillation (Persistent AF, PeAF). The recurrence risk in PeAF patients is influenced by various factors, which complicates the prediction of ablation outcomes. While clinical characteristics provide important references for risk assessment, the predictive accuracy of existing methods is limited and they fail to fully leverage the rich information contained in electrocardiogram (ECG) signals. Integrating clinical features with ECG signals holds promise for enhancing recurrence prediction accuracy and supporting personalized management.MethodsThis study conducted a retrospective analysis of PeAF patients who underwent radiofrequency catheter ablation treatment between 2016 and 2019. A multimodal fusion framework based on a residual block network structure was proposed, integrating preprocedural AF rhythm 12-lead ECG signals, clinical scores, and baseline characteristics of the patients to construct a deep learning model for predicting the risk of postablation recurrence in PeAF patients. A fivefold cross-validation method was used to partition the data set for model training and testing.ResultsThe fusion model was evaluated on a cohort of 77 PeAF patients, achieving good predictive performance with an average AUC of 0.74, and a maximum of 0.82. It significantly outperformed traditional clinical scoring systems and single-modal models based solely on ECG signals. Additionally, the model demonstrated lower variance (0.08), reflecting its robustness and stability with small sample sizes.ConclusionThis study innovatively combines AF rhythm ECG signals with clinical characteristics to construct a deep learning model for predicting the recurrence risk in PeAF patients after radiofrequency catheter ablation. The results show that this method effectively improves prediction performance and provides support for personalized clinical decision-making, with significant potential for clinical application.