Background: The Clinical Dementia Rating (CDR) is widely used to stage disease severity in Alzheimer’s disease dementia (ADD) and frontotemporal dementia (FTD). However, prior research has predominantly emphasized cognitive and neuropsychiatric domains as predictors, with comparatively limited attention to functional abilities. Consequently, evidence integrating cognitive, neuropsychiatric, and functional dimensions to predict disease severity remains scarce, particularly in Latin American (LA) populations. This study aimed to examine multidimensional predictors of disease severity, as measured by the CDR, in ADD and FTD subjects using machine learning (ML) models in a LA cohort.

Methods: A total of 434 Chilean and Peruvian participants from the ReDLat and Fondecyt DEFEND (Data-drivEn assessment of Functioning in NeuroCognitive disorders) studies were included: 222 cognitively unimpaired (CU), 140 with ADD, and 72 with FTD. Participants were assessed with the CDR and CDR-FTLD (total and sum-of-boxes, or SoB, scores), measures of cognition (global cognition, memory, language, executive function, and visuospatial abilities), neuropsychiatric symptoms (NPI-Q), and functional ability (T-ADLQ, ECog, WHODAS, and BLESSED) (Table 1). For each pathology, ML classification models (XGBoost, Logistic Regression, Random Forest) predicted total scores, and ML regression models (Gradient Boosting, Random Forest, Lasso) predicted SoB scores. Models were trained on 75% of the data and evaluated on the remaining 25% using 5-fold cross-validation.

Results: In ADD, multidimensional models achieved an F1-score of 0.71±0.04 and an R² of 0.59±0.04, while in FTD, performance reached an F1-score of 0.69±0.05 and an R² of 0.42±0.11, using CDR-FTLD total scores for classification and CDR-SoB for regression (Figure 1). Functional measures (T-ADLQ and ECog) emerged as the most robust predictors of disease severity across both conditions (Figure 2).

Conclusion: Across multiple ML models and measures of disease severity, multidimensional measures of functional ability, integrating ADL (T-ADLQ) and daily life cognition (ECog), emerged as the most robust and consistent predictors of dementia severity, outperforming neuropsychiatric variables and most cognitive measures, with the MMSE as the sole cognitive exception. Together, these findings highlight the central importance of multidimensional functional assessments for clinical evaluation and predictive modeling of dementia severity and underscore the need to test their applicability across diverse populations.