Enhancing diagnosis prediction with adaptive disease representation learning.

Area of Science:

• Medical Informatics
• Machine Learning
• Computational Biology

Background:

• Electronic health records (EHRs) contain valuable data for predicting future diseases.
• Time series models capture disease progression but often miss semantic correlations between diseases.
• Predefined disease co-occurrence graphs can aid diagnosis but may be noisy and lack adaptive semantic understanding.

Purpose of the Study:

• To develop an adaptive graph-driven framework for improved disease prediction using EHRs.
• To learn robust disease representations by dynamically optimizing disease relationships.
• To enhance the accuracy of future diagnosis prediction.

Main Methods:

• Proposed an end-to-end framework named Adaptive Disease Representation Learning (ADRL).
• Introduced an adaptive mechanism for dynamic adjustment of disease relationships via self-supervised perturbations on a global co-occurrence graph.
• Utilized an SVD-based accelerator to reduce computational load.

Main Results:

• The ADRL model effectively learns complex semantic associations between diseases.
• Adaptive learning of disease relationships significantly improved diagnosis prediction performance.
• Experimental results on two real-world EHR datasets demonstrated superior performance compared to existing models.

Conclusions:

• Adaptive Disease Representation Learning (ADRL) offers a powerful approach for enhancing disease prediction from EHRs.
• Dynamically adapting disease relationships is crucial for capturing nuanced semantic correlations.
• The proposed framework shows significant potential for clinical decision support systems.