Constrained Tensor Factorization for Cancer Phenotyping and Mortality Prediction

Electronic health records (EHR) enable machine learning methods like tensor factorization to extract computational phenotypes. Using Northwestern Medicine data (2000-2015), we analyzed breast, prostate, colorectal, and lung cancer cohorts to predict five-year mortality. Adding a supervised term, indication filtering, and social determinants of health (SDOH) covariates improved interpretability and performance. AUCs ranged from 0.623-0.694 (breast), 0.603-0.750 (prostate), 0.523-0.641 (colorectal), and 0.517-0.623 (lung). Constrained tensor factorization proves effective for deriving mortality-predictive phenotypes from sparse EHR data.