A deep learning model integrating structured data and clinical text for predicting atrial fibrillation recurrence.

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Informatics

Background:

• Multimodal perioperative data from atrial fibrillation (AF) ablation patients are valuable for risk stratification but underutilized in prediction models.
• Accurate prediction of AF recurrence post-ablation is crucial for optimizing patient management and treatment strategies.

Purpose of the Study:

• To develop and validate a deep learning model integrating multimodal perioperative data for predicting AF recurrence after ablation.
• To assess the performance of different large language models in processing textual data for AF recurrence prediction.

Main Methods:

• A multicenter retrospective study involving 2508 patients undergoing AF ablation.
• Development of a dual-branch deep learning model utilizing 1D ResNet for structured data and large language models (LLaMA-7B, Phi2-2.7B, Mistral-7B, MedGemma-27B) for textual data.
• Model performance evaluated using Area Under the Curve (AUC) on training, validation, and test sets.

Main Results:

• The deep learning model incorporating MedGemma-27B for text feature extraction achieved high predictive performance.
• AUC values were 0.934 (training), 0.928 (validation), and 0.911 (test set), demonstrating robust generalization.
• The model successfully integrated multimodal data to identify patients at high risk of AF recurrence.

Conclusions:

• A novel deep learning model effectively predicts atrial fibrillation recurrence using multimodal perioperative data from AF ablation patients.
• The model's ability to identify high-risk individuals facilitates targeted interventions, potentially reducing relapse rates.
• This approach highlights the potential of integrating advanced AI techniques with comprehensive patient data in cardiovascular medicine.