Phe2vec

Phe2vec automates disease phenotyping by deriving embeddings from electronic health records to identify disease cohorts using unsupervised representation learning.

Key Features:

• Unsupervised learning algorithms: Uses Word2vec, GloVe, and Fasttext to learn representations from EHR data without labeled phenotypes.
• Embedding-based phenotyping: Pre-computes embeddings for medical concepts and patients’ longitudinal clinical histories to capture semantic relationships and identify phenotypes by proximity in embedding space.
• Scalability and flexibility: Applies across diverse diseases without requiring disease-specific modifications.
• Validation against expert standards: Assessed against the Phenotype KnowledgeBase (PheKB) with evaluations on ten diseases showing comparable or superior positive predictive values versus PheKB rule-based algorithms.

Scientific Applications:

• Disease cohort identification: Identifies patient cohorts for downstream clinical and translational studies using embedding similarity to phenotype definitions.
• Epidemiological research: Enables large-scale, data-driven phenotyping to support population-level disease prevalence and association studies.
• Drug repurposing: Facilitates identification of patient groups with similar phenotypic profiles for drug-repositioning analyses.
• Precision and personalized medicine: Supports stratification of patients based on learned phenotypic representations for individualized analyses.

Methodology:

Trained on structured EHRs and clinical notes from 1,908,741 Mount Sinai patients covering 49,234 medical concepts; computes embeddings for medical concepts and patients’ longitudinal records using Word2vec, GloVe, and Fasttext; derives phenotypes by selecting a seed concept and its neighboring concepts in the embedding space; evaluates cohort identification against PheKB across ten diseases using positive predictive value comparisons.