Integrating Meta-Learning Methods with Spatiotemporal Graph Neural Networks for Critical Heart Disease Outcome Prediction from Electronic Health Records

Extracting valuable insights from Electronic Health Record (EHR) data is challenging since they are high-dimensional, sparse, and with evolving temporal patterns. In this paper, an advanced framework that combines Meta-Learning with Spatiotemporal Graph Neural Networks (ST-GNNs) was developed to enhance disease prediction models utilizing EHR data and their robustness against data variability. The novelty of this approach is the utilization of meta-learning on one EHR dataset with one healthcare task for the acquisition of a single model that can generalize to other EHR datasets and healthcare tasks. Finally, to understand the spatiotemporal relationship of patients, the proposed framework leverages ST-GNNs for capturing complex spatiotemporal relationships in patient data and obtaining a comprehensive understanding of disease progression over time. The main goal of this study is to improve disease prediction using model parameters that adapt dynamically using meta-learning and effectively learning temporal dependencies on EHR data using GNNs. The model is trained with an adaptive learning mechanism where it is constantly fine-tuning itself according to the evolving data patterns. The proposed approach attains superior performance in terms of accuracy and training time saved in comparison with state-of-the-art deep learning models. Moreover, the integration of spatiotemporal GNNs significantly improves the interpretability of predicted risk trajectories, consequently supporting better decision-making in personalized healthcare.