InfEHR: Clinical phenotype resolution through deep geometric learning on electronic health records

Area of Science:

• Artificial Intelligence
• Machine Learning
• Clinical Informatics

Background:

• Electronic health records (EHRs) offer rich multimodal data for clinical decision-making.
• Current machine learning (ML) approaches are limited by the need for large labeled datasets.
• Unlocking the full potential of EHR data for ML remains a significant challenge.

Purpose of the Study:

• To introduce InfEHR, a novel framework for automated clinical likelihood computation from EHRs.
• To enable advanced ML analyses on EHR data without requiring extensive labeled training data.
• To demonstrate the efficacy of geometric deep learning for probabilistic inference in clinical settings.

Main Methods:

• Developed InfEHR, a framework applying deep geometric learning to EHR data.
• Converted whole EHRs into temporal graphs to capture dynamic phenotypic information.
• Utilized a few labeled examples for training to compute and revise clinical probabilities.
• Validated InfEHR against physician heuristics using real-world EHR data.

Main Results:

• InfEHR achieved superior performance in identifying neonatal culture-negative sepsis and postoperative acute kidney injury.
• Demonstrated significant improvements in sensitivity for both conditions compared to physician heuristics.
• Maintained high specificity, indicating reliable predictions.
• Showcased particular effectiveness in low-prevalence disease detection.

Conclusions:

• InfEHR effectively computes clinical likelihoods from EHRs using geometric deep learning.
• The framework overcomes the limitation of requiring large labeled datasets for ML.
• InfEHR offers a scalable solution for probabilistic inference in real-world clinical settings.
• Geometric deep learning shows promise for advancing EHR-based clinical decision support.