Domain Knowledge Augmented Contrastive Learning on Dynamic Hypergraphs for Improved Health Risk Prediction

Akash Choudhuri, Hieu Vu, Kishlay Jha, Bijaya Adhikari

Abstract

Accurate health risk prediction is crucial for making informed clinical decisions and assessing the appropriate allocation of medical resources. While recent deep learning based approaches have shown great promise in risk prediction, they primarily focus on modeling the sequential information in Electronic Health Records (EHRs) and fail to leverage the rich mobility interactions among health entities. As a result, the existing approaches yield unsatisfactory performance in downstream risk prediction tasks, especially tasks such as Clostridioides difficile Infection (CDI) incidence prediction that are primarily spread through mobility interactions. To address this issue, we propose a new approach that leverages Hypergraphs to explicitly model mobility interactions to improve predictive performance in health risk prediction tasks. Unlike regular graphs that are limited to modeling pairwise relationships, hypergraphs can effectively characterize the complex high-order semantic relationships between health entities. Moreover, we introduce a new contrastive learning strategy that exploits the domain knowledge to generate semantically meaningful positive (homologous) and negative (heterologous) pairs needed for contrastive learning. This unique contrastive pair augmentation strategy boosts the power of contrastive learning by generating feature representations that are both robust and well-aligned with the domain knowledge. Experiments on two real-world datasets demonstrate the advantage of our approach in both short-term and long-term risk prediction tasks. Our framework obtains gains in performance up to 29.49% for PHOP and 30.64% for MIMIC-IV for MICU transfer prediction, and 13.17% for PHOP and 4.45% for MIMIC-IV for CDI Incidence Prediction.

Introduction

Most existing health risk prediction models treat patients as isolated sequences of data points within an Electronic Health Record (EHR). While deep learning models like RNNs and LSTMs are proficient at capturing a single patient's medical history, they often ignore "mobility interactions"—the web of shared environments and providers that connect patients. In a hospital setting, risks are often communal. If a group of patients shares a room or is treated by the same specialist, their health risks, particularly for infectious diseases like CDI, become intertwined. Standard graphs can model pairwise relationships, but they struggle to represent high-order interactions, such as multiple patients sharing a unit simultaneously. This paper bridges that gap by using dynamic hypergraphs to capture these complex group dynamics over time.

Methodology

The proposed framework consists of a three-part architecture designed to capture high-order interactions and refine them through domain-guided learning.

  1. Dynamic Heterogeneous Hypergraph Construction: For every timestamp, the model builds a hypergraph where nodes represent patients. Hyperedges are created based on shared entities: a "Doctor" hyperedge connects all patients seen by the same doctor, a "Room" hyperedge connects those in the same unit, and a "Medication" hyperedge connects those prescribed the same drug.
  2. Domain Knowledge Augmented Contrastive Learning: Standard contrastive learning often uses random perturbations (like dropping edges) to train models. However, random noise can be detrimental in clinical settings. This approach uses domain knowledge graphs—specifically a hierarchy of medications and a graph of doctor specialties—to guide perturbations. It creates "homologous" pairs by removing interactions between patients and doctors who share a specialty, forcing the model to learn deeper semantic representations rather than just memorizing links.
  3. Temporal Aggregation: After generating robust embeddings for each patient at each time step, the model employs a Transformer-based self-attention mechanism to aggregate information from the previous $k$ days to make a final risk prediction.
Architecture of the proposed framework, illustrating hypergraph construction and domain knowledge-based augmentation.

Experimental Setup

The model was evaluated on two large-scale datasets: PHOP, a proprietary dataset from a large 800-bed tertiary care teaching hospital containing over 349,000 medication interactions, and MIMIC-IV, a widely used open-source dataset containing de-identified EHR data from ICU patients. The tasks included predicting MICU (Medical Intensive Care Unit) transfers (1, 2, and 3 days ahead) and CDI incidence (3 days ahead).

Results and Discussion

The proposed model consistently outperformed state-of-the-art baselines. For MICU transfer prediction, the framework achieved performance gains of up to 29.49% in AUPRC for the PHOP dataset and 30.64% for MIMIC-IV. For the difficult task of predicting CDI infections, the model saw improvements of 13.17% in AUPRC on PHOP and 4.45% on MIMIC-IV.

Ablation studies showing the impact of different hyperedge types (a) and model components (b) on prediction performance.

Ablation studies (Figure 3) revealed that medication interactions were the most critical feature for prediction. Removing medication hyperedges resulted in the most significant drop in ROC scores compared to doctor or room hyperedges. This suggests that shared prescriptions serve as a powerful proxy for both the type of illness and the severity of the patient's condition. Furthermore, the inclusion of contrastive learning and hypergraph structures significantly boosted performance over baseline models, particularly for infectious disease prediction where group dynamics are paramount.

Conclusion

By leveraging dynamic hypergraphs and domain-knowledge-augmented contrastive learning, this framework provides a more holistic view of patient risk in hospital environments. Future work may include increasing interaction granularity by accounting for shared medical equipment and nurses, as well as incorporating dynamic domain knowledge to reflect evolving clinical guidelines.

Summary updated

BibTeX
@inproceedings{choudhuri2025domain,
  author = {Akash Choudhuri and Hieu Vu and Kishlay Jha and Bijaya Adhikari},
  title = {{Domain Knowledge Augmented Contrastive Learning on Dynamic Hypergraphs for Improved Health Risk Prediction}},
  booktitle = {SDM},
  year = {2025},
  doi = {10.1137/1.9781611978520.52},
}
Figure