Core Concepts
Integrating EHR data with survival analysis using MixEHR-SurG improves mortality prediction and identifies phenotype topics associated with patient outcomes.
Abstract
MixEHR-SurG is a novel model that integrates electronic health records (EHR) data with survival analysis to predict patient mortality. The model, MixEHR-SurG, combines EHR topic inference with Cox proportional hazards likelihood and patient-specific topic hyperparameters using PheCode concepts. It achieves competitive mortality prediction and generates meaningful phenotype topics for in-depth survival analysis. The study evaluated MixEHR-SurG using simulated datasets and real-world EHR datasets, demonstrating superior performance in predicting mortality risk among patients. The model identified severe cardiac conditions as significant mortality risk factors in the Quebec Congenital Heart Disease dataset and critical brain injuries in the MIMIC-III dataset. Overall, MixEHR-SurG offers an interpretable approach to distill phenotypic concepts relevant to patient outcomes.
Stats
MixEHR-SurG achieved a mean AUROC score of 0.89 in the simulation dataset.
MixEHR-SurG achieved a mean AUROC of 0.645 on the CHD dataset.
Quotes
"MixEHR-SurG integrates heterogeneous EHR data and models survival hazard effectively."
"MixEHR-SurG provides competitive mortality prediction and meaningful phenotype topics."