CURENet: combining unified representations for efficient chronic disease prediction.

Area of Science:

• Biomedical Informatics
• Artificial Intelligence in Healthcare
• Clinical Data Science

Background:

• Electronic health records (EHRs) contain diverse data types (notes, labs, visits) crucial for patient health assessment.
• Current predictive models often fail to integrate multimodal EHR data effectively, limiting prediction accuracy.
• Capturing temporal patterns and interactions across data modalities is essential for robust clinical prediction.

Purpose of the Study:

• To develop and evaluate CURENet, a multimodal model for predicting chronic diseases using integrated EHR data.
• To address the limitations of existing models in handling complex interactions within EHR data.
• To enhance the reliability of chronic illness prediction through multimodal data fusion.

Main Methods:

• CURENet integrates unstructured clinical notes, lab tests, and time-series visit data using large language models (LLMs) and transformer encoders.
• LLMs process clinical text and textual lab results, while transformers analyze longitudinal patient visit data.
• The model was evaluated on the MIMIC-III and FEMH datasets for multi-label chronic condition prediction.

Main Results:

• CURENet achieved over 94% accuracy in predicting the top 10 chronic conditions.
• The model demonstrated capability in capturing intricate interactions between different clinical data modalities.
• Successful validation on both public (MIMIC-III) and private (FEMH) datasets.

Conclusions:

• Multimodal EHR data integration using CURENet significantly enhances chronic disease prediction.
• The model's ability to process diverse data types improves the reliability of predictive analytics in healthcare.
• Findings suggest CURENet's potential to advance clinical decision-making and improve patient outcomes.