Can Unsupervised Machine Learning Support Malnutrition Risk Screening in Older Adults? A Preliminary Study Using Multidimensional Phenotyping and Hierarchical Clustering on Principal Components

Background/Objectives: Malnutrition in older adults constitutes a significant public health concern, reflecting a multifactorial risk and demonstrating the necessity of early identification in geriatric care. This preliminary study aimed to explore whether an unsupervised machine learning (ML) approach based on hierarchical clustering on principal components (HCPC) could help characterize multidimensional phenotypes associated with malnutrition risk in older adults. Methods: This exploratory cross-sectional study included 105 older adults, aged 60–95 years, recruited from community and residential care settings. Anthropometric measurements, body composition, nutritional status, depressive symptoms, physical and functional fitness, frailty, sarcopenia, appetite, and biological and hematological parameters were evaluated. The statistical analyses were performed using ML combining Multiple Factor Analysis (MFA) for dimensionality reduction and HCPC for cluster identification. Results: HCPC identified three distinct phenotypes. Cluster 1 represented a relatively preserved nutritional and geriatric phenotype with lower frailty, residential care, and disease burden. Cluster 2 was a male-dominant group with preserved muscle mass and functional reserve. Cluster 3 was the most vulnerable phenotype, characterized by higher malnutrition risk, frailty, polypharmacy, depressive symptoms, and poorer functional status, with the strongest links observed for body composition, muscle mass, grip strength, malnutrition risk, frailty risk, and functional performance. Conclusions: This preliminary study suggests that unsupervised machine learning may help characterize geriatric and nutritional profiles associated with malnutrition risk in older adults. Higher malnutrition risk was observed in the profile characterized by greater multidimensional vulnerability, warranting cautious interpretation and validation in larger cohorts.