Abstract
Predicting future health events from Electronic Health Records (EHR) is a cornerstone of personalized medicine, yet traditional models often overlook the evolving clinical relationships between diseases. This paper introduces Chet (Context-aware Health event prediction via Transition functions), a novel framework that models a patient's medical history as a dynamic graph. By leveraging global disease co-occurrence patterns and local patient subgraphs, Chet distinguishes between persistent and emerging conditions. Utilizing a Modified Gated Recurrent Unit (M-GRU) to track disease-level hidden states, the model captures how potential risks evolve into diagnoses. Experimental results on MIMIC-III and MIMIC-IV datasets demonstrate that Chet significantly outperforms state-of-the-art baselines in both diagnosis and heart failure prediction while maintaining high parameter efficiency.
Introduction
Effective health event prediction requires capturing both the global clinical context—how diseases relate across a population—and the local context of an individual patient's trajectory. Current models frequently treat disease codes as independent, static entities, failing to account for the fact that a condition's clinical priority and impact change over time. We propose that medical history is best represented as a dynamic subgraph that evolves across hospital visits, reflecting the transition of diseases from potential risks to active diagnoses.
Methodology
Dynamic Graph Construction
Chet begins by constructing a global co-occurrence graph from the entire dataset. For each patient visit, the model extracts a local subgraph. This subgraph includes the patient's current diagnoses and their "neighbors"—diseases they do not yet have but which frequently appear alongside their current conditions in other patients. This approach allows the model to monitor "neighbor" nodes as indicators of future prognosis.
Disease-Level Temporal Learning and M-GRU
To model the evolution of a patient's state, Chet categorizes diseases into three roles: Persistent (existing conditions), Emerging Neighbors (new conditions previously identified as neighbors), and Emerging Unrelated (new conditions appearing without prior local context).
The framework employs a Modified Gated Recurrent Unit (M-GRU) designed to handle matrix-based hidden states. Unlike standard GRUs that process flat vectors, the M-GRU tracks individual disease trajectories, learning the transition functions that govern how a disease's role changes between visits.
Experimental Evaluation
Datasets
The model was evaluated on two large-scale EHR datasets:
- MIMIC-III: Data from ~7,500 patients with multiple visits between 2001 and 2012.
- MIMIC-IV: A more recent sample of 10,000 patients from the 2013–2019 period.
Key Results
Chet demonstrated superior performance across multiple tasks:
- Diagnosis Prediction: On MIMIC-III, Chet achieved a weighted F1 score of 22.63% and a Recall@20 of 37.87%, significantly outperforming models like G-BERT, GRAM, and RETAIN.
- Heart Failure Prediction: The model reached an AUC of 90.83% on the MIMIC-IV dataset.
- Efficiency: Chet achieved these results with significantly fewer parameters (2.59M) than competitors like G-BERT (7.53M), proving that context-aware architecture can achieve superior results with lower computational overhead.
Conclusion and Future Work
Chet advances health event prediction by treating medical history as a dynamic, relational graph. By explicitly modeling disease-level transitions and leveraging global clinical knowledge, it provides a more accurate and clinically relevant representation of patient health. Future research will focus on incorporating additional data modalities, such as lab results and clinical notes, and enhancing model interpretability to better support clinical decision-making.
Summary updated