Enhanced diagnostic interpretation of the MoCA using machine learning.

Introduction: Artificial Intelligence (AI) is increasingly being integrated into clinical practice to optimize diagnosis in neurocognition. By capturing distinct cognitive signatures, this approach may offer a more precise alternative to the traditional interpretation of the Montreal Cognitive Assessment (MoCA) which often relies on a fixed cutoff score (26/30). We aimed to evaluate whether machine learning models, by integrating detailed MoCA subtest scores, demographic variables, and cognitive chart-derived metrics, can improve the detection of cognitive impairment and classification of dementia subtypes. Methods: We analyzed 38,746 clinical observations (17,188 unique individuals) from the National Alzheimer's Coordinating Center database. Five supervised learning algorithms, Extreme Gradient Boosting (XGBoost), Random Forest, Support Vector Machine (SVM), Logistic Regression, and k-Nearest Neighbors (KNN), were trained using detailed MoCA subtest scores, demographic variables, and cognitive chart-derived metrics as predictors. To ensure generalizability of results and prevent data leakage, we applied a rigorous nested Repeated Grouped Cross-Validation strategy. Decision thresholds were optimized via the Youden Index on independent calibration sets, and model interpretability was ensured through SHAP value analysis. Results: Machine learning models consistently outperformed conventional approach. For the global detection of cognitive impairment, XGBoost achieved the best performance (Youden Index 0.61 vs. 0.54 for the standard cutoff). Regarding subtype classification, models demonstrated variable discriminative capacity depending on clinical homogeneity: primary progressive aphasia was best classified (Youden ≈ 0.77), followed by Lewy body dementia and Alzheimer's disease, while vascular dementia remained more challenging to isolate. Feature importance analysis highlighted the Cognitive Quotient as a robust universal predictor, while pinpointing disease-specific drivers such as delayed recall for Alzheimer's disease and verbal fluency for primary progressive aphasia. Conclusion: Our findings suggest interpretable machine learning enhances diagnostic utility of the MoCA, yielding superior accuracy compared to a fixed cutoff. By synthesizing individualized subtest profiles within a transparent framework, this approach offers a clinically actionable solution. It transforms the MoCA from a simple screening tool to a precision diagnostic aid, optimizing patient triage in the era of disease-modifying therapies.