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

Abstract

Multimodal perioperative data from patients undergoing atrial fibrillation (AF) ablation offer valuable insights for stratifying recurrence risk, yet remain underutilized in prediction models. This multicenter retrospective study included 2508 patients who underwent AF ablation at five Chinese centers: The Fourth Affiliated Hospital of Zhejiang University School of Medicine (Jan 2016-Mar 2024; Training Cohort), Taizhou Hospital of Zhejiang Province (Jan 2015-Jan 2024; Training Cohort), The Affiliated Hospital of Yunnan University (Jan 2016-Jan 2024; Validation Cohort), Jinhua People's Hospital (Jan 2020-Jan 2024; Test Cohort), and Ningbo Beilun Hospital (Jan 2020-Jan 2024; Test Cohort). We developed a dual-branch deep learning model to predict AF recurrence, in which structured data were processed via a 1D ResNet and textual data were encoded using four large language models (LLaMA-7B, Phi2-2.7B, Mistral-7B, and MedGemma-27B). The model incorporating MedGemma for text feature extraction performed best, achieving areas under the curve of 0.934 (95% CI: 0.921-0.946), 0.928 (95% CI: 0.904-0.950), and 0.911 (95% CI: 0.878-0.941) on the training, validation, and test sets, respectively. Our model integrates multimodal perioperative data from AF ablation patients, effectively identifies high-risk individuals, and may facilitate targeted interventions to reduce relapse.

Ähnliche Arbeiten