Atrial fibrillation (AF) is associated with an elevated risk of ischemic stroke (IS). The growing global burden of AF and IS demands improved risk prediction and prevention strategies. Advancements of artificial intelligence (AI), particularly machine learning (ML) and deep learning (DL) offer opportunities to leverage vast biological datasets for enhanced risk assessment and personalized medicine. Our research explores the application of AI techniques, especially knowledge graph-based approaches, to process complex relationships within biological data for AF-related IS risk prediction. We present three novel approaches for AF-related IS risk prediction:

An AI approach, leveraging ML techniques, utilizes electronic health records, drug history, and demographic information to calculate IS risk for AF patients. Light Gradient Boosting Machine (LightGBM) was employed as the predictive model, with SHapley Additive exPlanations (SHAP) analysis adapted to provide interpretability. This method outperforms current standard statistical methods in accuracy and interpretability, offering clinicians a more reliable risk assessment tool.

A graph-based learning approach that analyzes two constructed graphs for IS-AF prediction, incorporating protein-protein interactions. The core of this method is an embedding-based deep learning model developed for disease prediction. SHAP analysis was also applied to provide insights into risk calculations. This approach stands out in capturing and integrating complex relationships between biological entities into the risk calculations.

AF-Biological-IS-Path (ABioSPATH), is an interpretable path-based model that combines drug-protein-disease pathways with real-world clinical data. It implements Graph Attention Network (GAT), Bidirectional Long Short-Term Memory (Bi-LSTM) networks, and attention mechanisms to predict IS risk and identify potential pathways. ABioSPATH outperforms previously constructed methods, achieving state-of-the-art performance while providing molecular pathway-level insights.

Experimental results demonstrate that these three methods significantly outperform existing approaches in predicting 1-year IS risk in AF patients. Notably, ABioSPATH achieved a 16% increase in AUROC compared to the current clinical standard CHA₂DS₂-VASc score. Our work presents a comprehensive framework for AF patient analysis and algorithm construction, bridging the gap between AF-IS protein analysis and graph-based deep learning methods. This research offers a systematic solution for IS risk prediction that can benefit healthcare professionals and medical experts, leading to tailored and more efficient AF management strategies.