Context-Aware Health Event Prediction via Transition Functions on Dynamic Disease Graphs

Chang Lu, Tian Han, Yue Ning

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.

Figure 1: Illustration of the global disease combination graph and the extraction of local subgraphs for sequential visits, highlighting the transition of neighbors to diagnoses.

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.

Figure 2: The Chet architecture, showing the integration of dynamic graph learning, role-based disease aggregation, and the M-GRU transition mechanism.

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

Abstract

With the wide application of electronic health records (EHR) in healthcare facilities, health event prediction with deep learning has gained more and more attention. A common feature of EHR data used for deep-learning-based predictions is historical diagnoses. Existing work mainly regards a diagnosis as an independent disease and does not consider clinical relations among diseases in a visit. Many machine learning approaches assume disease representations are static in different visits of a patient. However, in real practice, multiple diseases that are frequently diagnosed at the same time reflect hidden patterns that are conducive to prognosis. Moreover, the development of a disease is not static since some diseases can emerge or disappear and show various symptoms in different visits of a patient. To effectively utilize this combinational disease information and explore the dynamics of diseases, we propose a novel context-aware learning framework using transition functions on dynamic disease graphs. Specifically, we construct a global disease co-occurrence graph with multiple node properties for disease combinations. We design dynamic subgraphs for each patient's visit to leverage global and local contexts. We further define three diagnosis roles in each visit based on the variation of node properties to model disease transition processes. Experimental results on two real-world EHR datasets show that the proposed model outperforms state of the art in predicting health events.
BibTeX
@inproceedings{lu2022context,
  author = {Chang Lu and Tian Han and Yue Ning},
  title = {{Context-Aware Health Event Prediction via Transition Functions on Dynamic Disease Graphs}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.1609/aaai.v36i4.20380},
}
Figure