Collaborative Coarse-to-Fine Disease Learning With Discharge Summary Awareness for EHR Event Prediction.

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Computational Biology

Background:

• Deep learning models are used for electronic health record (EHR) event prediction.
• Existing models face challenges in modeling dynamic disease relationships, leveraging diagnosis code ontologies, and incorporating unstructured clinical notes.

Purpose of the Study:

• To propose a coarse-to-fine disease learning framework with patient notes for enhanced EHR event prediction.
• To capture both dynamic and static disease characteristics effectively.
• To address limitations in current deep learning models for EHR analysis.

Main Methods:

• Constructed a fine-grained dynamic disease graph by analyzing co-occurrence distributions.
• Refined disease embeddings by integrating hierarchical ICD-9-CM code information.
• Utilized gated recurrent units, location-based attention, and soft attention mechanisms.
• Incorporated unstructured discharge summaries and auxiliary patient notes for collaborative learning.

Main Results:

• The proposed model demonstrated superior performance in EHR event prediction compared to nine baseline methods.
• Experiments were conducted on two real-world EHR datasets: MIMIC-III and MIMIC-IV.
• The framework effectively captured dynamic disease relationships and integrated multi-perspective ontological information.

Conclusions:

• The coarse-to-fine disease learning framework offers a significant advancement in EHR event prediction.
• Integrating hierarchical diagnosis codes and unstructured clinical notes improves model accuracy.
• The proposed method provides a robust approach for analyzing complex patient data in healthcare.