Cognitive Profiles of People Living with Dementia: PCA and Clustering Analysis

Introduction: Dementia presents heterogeneous cognitive profiles that often transcend standard diagnoses. However, clinical classifications tend to overlook subtle differences in memory, language, executive, and visuospatial function. Data-driven subtyping may better capture this cognitive variability to inform personalized interventions. We aimed to identify, characterize, and evaluate the clinical relevance of distinct cognitive subtypes among patients with dementia using a dimensionality-reduction and clustering approach on TorCA data. Methods: We analyzed cross-sectional data from 721 individuals (TDRA cohort) with dementia who completed a battery of neuropsychological tests covering memory, executive, language, and visuospatial domains. After data pre-processing (multicollinearity and missing values), we performed principal component analysis with multiple retention criteria to reduce dimensionality. Ward’s hierarchical clustering was applied on the selected principal components to derive cognitive subgroups. Cluster validity and profiles were assessed via silhouette scores and visualizations (biplots, radar charts). Results: Four principal components were retained, reflecting overall severity and major cognitive domains. Hierarchical clustering based on these components revealed three robust cognitive subtypes: (1) predominant executive and visuospatial impairment; (2) broad language/fluency impairment; and (3) marked memory and visuospatial deficits. These subtypes cut across traditional diagnostic categories (e.g. Alzheimer’s, vascular dementia), underscoring heterogeneity within diagnoses. In an exploratory seven-cluster solution, clusters differed significantly in age, suggesting early-onset vs. late-onset cognitive profiles. Discussion: A PCA-guided clustering approach identified clinically meaningful dementia subtypes that are not captured by conventional diagnoses. These data-driven subtypes highlight the importance of assessing multidomain cognitive patterns and offer potential for tailoring clinical interventions and improving prognostic accuracy.