Enhancing Heart Failure Prediction with Temporal Attention and Graph-based Learning

Heart failure (HF) remains a major global health concern, with early risk prediction playing a crucial role in improving patient outcomes and reducing hospitalization rates. Traditional predictive models often struggle to capture the complex and evolving nature of a patient’s health status over time. To address this challenge, we propose a Temporal Risk Pattern Network (TRPN), a novel approach that leverages time-series electronic health records (EHRs) to identify patterns in disease progression. By incorporating a graph-based structure with temporal attention mechanisms, TRPN dynamically models fluctuations in key clinical indicators, allowing for more accurate and personalized risk assessment. Our model constructs a temporal risk graph, which continuously learns from patient data to detect early warning signals of heart failure. This approach enables the identification of both short-term variations and long-term trends in vital signs, laboratory results, and comorbidities. Using real-world EHR datasets, we evaluate TRPN against traditional machine learning models, demonstrating its superior performance in accuracy, recall, and F1-score. Additionally, the interpretability of the network enhances clinical trust, supporting healthcare providers in making timely and informed decisions. By integrating advanced temporal modeling with predictive analytics, TRPN offers a robust framework for heart failure risk prediction, paving the way for proactive patient management and personalized treatment strategies.