Post–COVID Cognitive Risk via Unsupervised AI Phenotyping in the All of Us Cohort
Year: 2026
Type: article
Abstract: Objective/Purpose: To establish and validate an artificial intelligence–driven, unsupervised phenotyping approach that synthesizes electronic health record and patient-reported survey data into a multidimensional risk score for early identification of cognitive impairment following SARS-CoV-2 infection in older adults. Methodology: We leveraged data from 23,435 participants aged 50 years or older enrolled in the All of Us Research Program with laboratory-confirmed COVID-19 between January 2020 and December 2022. Using association rule mining, we mined the cognitive-impairment subgroup (n = 1,424) for 30 high-frequency binary indicators spanning medical comorbidities (via Charlson categories), functional status (daily activities, self-rated health), and social support. K-Modes clustering (k = 5) grouped individuals into five distinct phenotypes, producing an ordinal “severity score.” We tested the score’s prognostic value in multivariable logistic regression—adjusting for age, sex, race/ethnicity, education, income, insurance, and comorbidity—and in propensity-score–matched analyses (1:1 nearest neighbor; caliper 0.2×SD logit PS). Predictive performance was further evaluated with an XGBoost classifier balanced by SMOTE, calibrated by Platt scaling, and interpreted through SHAP values. We quantified clinical utility using decision-curve analysis and assessed model performance via subgroup AUCs and parity metrics across sex and racial/ethnic categories. Results: Increasing severity score predicted cognitive impairment within 30 days post-infection: each phenotype increment was associated with a 2.10-fold increase in adjusted odds (95% CI [1.98, 2.22]; p < .001), with an age interaction (OR 1.02 per year; p = .03). Matched analyses confirmed a 4.5% absolute risk elevation (ATT 0.045; 95% CI [0.032, 0.058]), robust to Rosenbaum bounds (Γ ≤ 1.5). The XGBoost model achieved an AUC of 0.82 ± 0.02, calibration slope of 1.00, and Brier score of 0.039; decision-curve analysis demonstrated net benefit improvements of 2%–4% at 5%–20% risk thresholds. Model audits showed consistent AUCs (0.80–0.83) and parity differences < 0.03. Continuous net reclassification improvement was 0.12. Conclusions: Our unsupervised AI-phenotyping pipeline generates a robust multidimensional risk score for post–COVID cognitive impairment that outperforms traditional single-factor models in prediction, calibration, and subgroup performance. Implications Implementation of this phenotyping framework within clinical decision-support platforms could facilitate targeted cognitive screening and early intervention, optimizing care pathways and resource allocation for older adults in the post–COVID era.
Authors: Rahemi, Zahra, Omidi, Meisam
Institutions: Clemson University, Marquette University
Language: English
Subfield: Clinical Neurology
Field: Medicine
Domain: Health Sciences
Open Access status: green
Funder: National Institutes of Health
Award: G9000064959
OpenAlex ID: W7154090911