GRAM: Graph-based Attention Model for Healthcare Representation Learning.

Area of Science:

• Artificial Intelligence in Medicine
• Computational Biology
• Biomedical Informatics

Background:

• Deep learning shows promise in healthcare predictive modeling but faces challenges with insufficient data and poor interpretability.
• Existing methods struggle to align learned representations with established medical knowledge, limiting clinical utility.

Purpose of the Study:

• To address data insufficiency and interpretability issues in healthcare deep learning models.
• To propose and evaluate a GRaph-based Attention Model (GRAM) that integrates electronic health records (EHR) with medical ontologies.

Main Methods:

• GRAM supplements EHR data with hierarchical medical ontology information.
• A novel attention mechanism represents medical concepts as combinations of their ontological ancestors.
• Performance was compared against recurrent neural networks (RNNs) in disease prediction tasks.

Main Results:

• GRAM achieved 10% higher accuracy for rare disease prediction compared to RNNs.
• It demonstrated a 3% improved area under the ROC curve for heart failure prediction.
• GRAM required an order of magnitude less training data and produced interpretable, ontology-aligned representations.

Conclusions:

• GRAM effectively addresses data insufficiency and interpretability challenges in healthcare predictive modeling.
• The model's ability to generalize and align with medical knowledge offers a significant advancement.
• GRAM provides a promising approach for developing more robust and clinically relevant AI in medicine.