Deterministic Overlapping Multimorbidity Phenotypes for Leakage-Safe EHR Modeling of Incident Cognitive Impairment in All of Us.

Background: In electronic health record (EHR) research, multimorbidity is commonly represented by summary indices that may oversimplify disease co-occurrence or by unsupervised cluster labels that may lack stability across samples. This study evaluated a leakage-safe framework for deterministic, overlapping multimorbidity phenotyping. Methods: Data from the All of Us Research Program were used to study 23,435 adults aged 50 years and older, anchored at their first SARS-CoV-2-positive test. Eleven pre-index Charlson component indicators were used as the binary baseline representation. Stable co-occurrence patterns were explored using association rule mining with bootstrap recurrence criteria. K-modes clustering was used as a discovery aid, while final phenotype membership was defined by deterministic rule functions. Results: The framework produced three overlapping phenotype flags and a transparent 8-state overlap structure. The analytic cohort included 1462 incident cognitive impairment cases. Two phenotype flags were associated with higher odds of incident cognitive impairment. However, adding the phenotype flags to baseline models yielded only marginal improvements in discrimination. Conclusions: The framework's primary value lies in providing a reproducible, transparent, and overlap-aware representation of multimorbidity rather than in producing substantial predictive gains. This approach may support interpretable baseline phenotyping and leakage-safe modeling in real-world EHR studies.