PhenoNMF: A novel multi-layer matrix factorization framework for age-stratified comprehensive phenotypic similarity analysis

Résumé fourni par la source

BACKGROUND: Electronic health record (EHR) data enable deep phenotyping for risk prediction and treatment evaluation; however, many existing approaches lack interpretability and are difficult to interpret and validate. Methods that rely on a single modality and disregard age stratification may obscure differences across life stages and weaken model traceability and generalizability. METHOD: We developed PhenoNMF, a multimodal EHR phenotyping framework based on joint nonnegative matrix factorization. PhenoNMF learns sparse multimodal patterns from diagnoses, laboratory tests, and medications using modality-specific sparsity penalties. By excluding age from the decomposition and reintroducing it through age-weighted contribution projections, PhenoNMF captures multimodal co-occurrence within each common pattern module and compares these patterns across age groups. Within each common pattern module (CPM), we define an Age Network Coupling Score (ANCS) to rank diagnosis, laboratory, and medication triads supported by both age-weighted contribution and cross-modality association evidence. RESULTS: In a large critical-care cohort, PhenoNMF showed improved clustering stability and stronger cross-modal structure than multiple unsupervised baselines, yielding CPMs with coherent clinical themes across the lifespan. High-scoring triads prioritized by the ANCS highlighted age-specific multimodal patterns and were used to construct age-stratified survival analyses. These analyses revealed patient subgroups in which laboratory abnormalities modified drug-associated survival associations, summarized using hazard ratios and absolute risk differences. CONCLUSION: PhenoNMF provides an interpretable framework for multimodal EHR phenotyping that separates phenotype learning from age effects and reconnects them through ANCS. By linking age-stratified CPMs and multimodal triads to outcome-based risk measures, the framework supports more precise characterization of patient subgroups and age-specific treatment and risk interactions. CLINICAL TRIAL NUMBER: Not applicable.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.