Abstract
Electronic Health Records (EHRs) offer valuable insights for healthcare prediction. Existing methods approach EHR analysis through direct imputation techniques in data space or representation learning in feature space. However, these approaches face the following two critical limitations: first, they struggle to model long-term clinical pathways due to their focus on isolated time points rather than continuous health trajectories; second, they lack mechanisms to effectively distinguish between clinically relevant and redundant features when observations are irregular. To address these challenges, we introduce PathCare, a neural framework that integrates clinical pathway information into prediction tasks at the neuron level. PathCare employs an auxiliary sub-network that models future visit patterns to capture temporal health progression, coupled with a neuron-level filtering gate that adaptively selects relevant features while filtering out redundant information. We evaluate PathCare on the following three real-world EHR datasets: CDSL, MIMIC-III, and MIMIC-IV, demonstrating consistent performance improvements in mortality and readmission prediction tasks. Our approach offers a practical solution for enhancing healthcare predictions in real-world clinical settings with varying data completeness.