2399 - Phenotyping Chronological Accumulation of Comorbidities for Prostate Cancer Patients: A Computational Approach

Materials/Methods: We randomly sampled 2000 patients with PC from a single institutional dataset. We identified the first chronological instance of each patient’s inpatient and outpatient diagnoses, categorized by 3-digit ICD codes to more broadly capture comorbidities managed in the outpatient setting, after their PC diagnosis. Disease clustering (DC) and temporal relationship phenotyping between ICD diagnoses were characterized via NMF following Hassaine et al. Steps include (1) disease matrix generation for every patient by first incidence of disease at age (in years) with number of unique conditions in the patient dataset, (2) disease frequency and inverse patient frequency calculation to correct for bias towards more frequent diseases, (3) Gaussian smoothing to correct for the variance of diagnosis noise of recorded chronic diseases, and (4) matrix concatenation and decomposition via NMF to output 2 matrixes: diagnosis prevalence in a DC and temporal phenotype per DC. Optimal DCs were identified by measuring the cophenetic correlation coefficient to identify the most stable clustering derived from NMF results. DCs were analyzed by joint activation probabilities and chronicity of DC diagnoses by ascendancy analysis.