ICD-10 and SNOMED CT Representations of Cognitive Decline in All of Us

Electronic health record phenotyping of cognitive decline may vary with the ontology used to encode diagnoses and findings. We compared ICD-10 and SNOMED CT representations of cognitive decline in the All of Us dataset. Separate ontology-specific feature sets were constructed and analyzed with k-means clustering as an exploratory method; clusters were summarized by condition burden, unique conditions, span, age, gender, and top-condition prevalence. Principal component analysis was used for visualization, and logistic regression was used to examine whether ontology-shaped features were informative for dementia classification. ICD-10 clusters were dominated by broad symptom-oriented terms, whereas SNOMED CT showed more differentiated groupings, including a normal-cognition cluster and a higher-burden dementia-related cluster. These findings suggest that ontology choice can influence latent phenotype structure and may affect the interpretability and performance of downstream classification in observational health data.