An interpretable XGBoost model for differentiating early Alzheimer's disease and bvFTD

Background: Differentiating early Alzheimer's disease (AD) from behavioral variant frontotemporal dementia (bvFTD) is clinically challenging due to overlapping symptoms. While machine learning (ML) can improve diagnostic accuracy, model performance and interpretability are crucial for clinical adoption. Objective: To develop, optimize, and interpret a high-performing ML model to differentiate AD from bvFTD in mild dementia (CDR=1) using data-driven clinical variables. Methods: Using a 29-variable set derived from data of 10,452 NACC participants, we trained eight ML classifiers. Models were evaluated with and without hyperparameter optimization (HPO) via nested cross-validation. The optimal model, selected based on the area under the precision-recall curve (PR AUC), was interpreted using SHapley Additive exPlanations (SHAP). Results: The optimized XGBoost model demonstrated superior discrimination (PR AUC 0.787; ROC AUC 0.957). SHAP analysis revealed lower age and higher behavioral impairment scores as the top predictors for bvFTD, while memory and orientation deficits were key predictors for AD. Conclusion: An optimized and interpretable XGBoost model accurately differentiates these syndromes using only clinical data. The model’s decision-making process, clarified by SHAP, aligns with established clinical knowledge, highlighting its potential as a transparent data-driven tool to support diagnostic decisions.