Learning probabilistic phenotypes from heterogeneous EHR data.

Area of Science:

• Computational biology
• Medical informatics
• Machine learning

Background:

• Discovering computational models of disease (phenotypes) from electronic health records is challenging.
• Integrating heterogeneous clinical data (notes, labs, diagnoses) for disease modeling requires robust methods.

Purpose of the Study:

• To introduce the Unsupervised Phenome Model (UPhenome), a novel probabilistic graphical model for large-scale phenotype discovery.
• To demonstrate UPhenome's ability to jointly model diseases and clinical observations from heterogeneous patient data in an unsupervised manner.

Main Methods:

• Developed UPhenome, a probabilistic graphical model for joint modeling of diseases and clinical observations.
• Applied UPhenome to two distinct patient datasets: intensive care unit (ICU) records and long-term outpatient records.
• Evaluated UPhenome's performance against baseline Latent Dirichlet Allocation (LDA) models.

Main Results:

• UPhenome learned phenotypes that integrated heterogeneous data types more coherently than LDA-based phenotypes.
• Learned phenotypes predominantly represented single diseases, outperforming baselines in disease specificity.
• UPhenome-derived phenotypes showed significant correlation with ground-truth patient disorders on unseen data.

Conclusions:

• UPhenome effectively learns robust computational disease models from diverse, heterogeneous patient data.
• The model demonstrates adaptability across different clinical settings (ICU, outpatient) without task-specific tuning.
• UPhenome offers a powerful tool for large-scale phenotype discovery and clinical data integration.