This project aims to develop and validate a multimodal deep learning model for atrial fibrillation (AF) risk prediction in cardiovascular patients, while simultaneously investigating underlying pathological mechanisms through multi-omics analysis.
Scientific Rationale: Current AF prediction tools demonstrate limited accuracy by focusing solely on clinical parameters, neglecting the synergistic potential of multi-omics data integration. While recent studies have explored single-modal AI approaches (ECG-based models), no existing framework comprehensively combines clinical, imaging, genomic, proteomic, and metabolomic data – a critical gap this study addresses.
Objectives:
Primary: Develop a multimodal deep learning model (1-5 year prediction horizon) integrating:
Clinical data (electronic health records, NT-proBNP); Imaging (12-lead ECG, cardiac MRI); Multi-omics (proteomics/Olink, metabolomics, PRS).
Secondary:
Identify cross-modal interaction effects (e.g., PRS-protein synergies); Discover novel AF biomarkers through differential omics analysis.
Research Questions:
1. Which multimodal feature combinations optimally predict AF onset?
2. How do proteomic/metabolomic profiles differ between AF/non-AF groups?
3. Can we identify mechanistically interpretable pathways (e.g., inflammatory/fibrotic) driving AF risk?
Methodology:
Data: UK Biobank cohort with the data mentioned above.
Modeling: Architecture: Hybrid neural network (1D-CNN for ECG, Transformer for omics, FC for clinical); Validation: 5-fold CV, external validation (CKB cohort); Interpretation: SHAP values, attention mechanisms.
Analysis:
Prediction: Time-dependent AUC, Network benefit; Mechanisms: FDR-corrected omics differences, KEGG pathway enrichment.
This research will yield a clinically deployable AF prediction tool with superior accuracy and novel insights into AF pathophysiology through identified biomarkers and pathways, enabling targeted prevention strategies.
