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

Summary

InfEHR converts whole EHRs into temporal graphs and uses deep geometric learning to infer clinical likelihoods with minimal labeled data. In neonatal culture-negative sepsis and postoperative AKI, it markedly improved sensitivity over physician heuristics while maintaining high specificity. This demonstrates scalable, label-efficient sepsis phenotyping with potential for earlier recognition and risk stratification.

Key Findings

• Temporal-graph representation of EHRs enabled high-performance probabilistic inference with few labels.
• Sensitivity improved substantially for culture-negative neonatal sepsis (0.60 vs 0.04) and postoperative AKI (0.71 vs 0.20), with preserved specificity.
• Outperformed physician heuristics across two independent health systems, highlighting scalability.

Clinical Implications

Hospitals could deploy InfEHR-like tools to flag culture-negative neonatal sepsis or predict postoperative AKI earlier, guiding timely diagnostics and preventive interventions without extensive manual labeling.

Limitations

• Exact cohort sizes and prospective clinical impact (e.g., time-to-treatment reduction) were not reported.
• Algorithm performance compared to comprehensive machine learning baselines beyond heuristics was not detailed.

Future Directions

Prospective, interventional studies to test whether InfEHR-triggered alerts improve time-sensitive outcomes; integration with explainability and clinician-in-the-loop workflows; broader validation across diseases and health systems.