Unsupervised Machine Learning of Automated Brain Volumetry Uncovers Three Volumetric Endophenotypes in Cognitive Impair-ment
TL;DR
The findings suggest that cognitive impairment manifests through heterogeneous patterns of structural brain involvement, with potential implications for prognosis and personalized care, and the high accuracy of a parsimonious four-feature model supports the clinical utility of focused volumetric assessment for risk stratification.
Abstract
Cognitive impairment and dementia represent a heterogeneous group of disorders with variable patterns of brain atrophy. Automated brain volumetry combined with machine learning offers the potential to identify biologically meaningful subgroups beyond traditional clinical classifications. This study aimed to characterize cerebral volumetric profiles and identify distinct volumetric endophenotypes in individuals with cognitive impairment using unsupervised machine learning. This retrospective cross-sectional study included 78 participants with cognitive impairment who underwent T1-weighted three-dimensional magnetic resonance imaging (MRI). Automated brain volumetry was performed using mdbrain software (version 2.2), which provides normalized percentile values for 42 brain regions adjusted for age, sex, and total intracranial volume. Unsupervised k-means clustering was applied to volumetric percentiles to identify distinct endophenotypes. Cluster validation was performed using elbow method, silhouette analysis, Calinski-Harabasz index, and Davies-Bouldin index. Discriminative features were identified using ANOVA with effect size calculations (η²). A multinomial logistic regression model was developed to predict cluster membership. The cohort comprised 48 (61.5%) females and 30 (38.5%) males. Substantial inter-individual variability was observed across all brain regions, with hippocampal percentiles ranging from 0.1 to 99.8 (median: 25.6). Clustering analysis revealed three distinct volumetric endophenotypes: Cluster 1 (Preserved, n=27, 34.6%) characterized by percentiles above the 40th percentile; Cluster 2 (Moderate Atrophy, n=31, 39.7%) with percentiles in the 10th–30th range; and Cluster 3 (Severe Atrophy, n=20, 25.6%) with hippocampal percentiles below the 5th percentile and ventricular percentiles above the 85th percentile. Hippocampal volumes demonstrated the strongest discriminative ability (left hippocampus: F=68.4, η²=0.64; right hippocampus: F=65.2, η²=0.63), followed by temporal (η²=0.60–0.61) and ventricular volumes (η²=0.57–0.58). Interhemispheric asymmetry was minimal (median absolute difference: 3.4–7.2 percentile points). A parsimonious model using four volumetric features (left hippocampus, left temporal, left ventricle, and right frontal percentiles) achieved excellent classification accuracy (87.2%, cross-validated AUC=0.91). Automated brain volumetry combined with unsupervised machine learning identified three distinct volumetric endophenotypes in cognitive impairment, primarily discriminated by hippocampal, temporal, and ventricular volumes. These findings suggest that cognitive impairment manifests through heterogeneous patterns of structural brain involvement, with potential implications for prognosis and personalized care. The high accuracy of a parsimonious four-feature model supports the clinical utility of focused volumetric assessment for risk stratification. Future longitudinal studies with multimodal biomarker integration are warranted to validate these findings and establish their prognostic value.