Enhanced diagnostic interpretation of the MoCA using machine learning.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Diagnostics

Background:

• Traditional interpretation of the Montreal Cognitive Assessment (MoCA) relies on a fixed cutoff score, potentially limiting diagnostic precision.
• Artificial Intelligence (AI) offers a potential for more nuanced analysis of cognitive data.
• Optimizing neurocognitive diagnosis is crucial for timely and accurate dementia subtyping.

Purpose of the Study:

• To evaluate machine learning (ML) models for improved detection of cognitive impairment.
• To assess ML's ability to classify dementia subtypes using detailed MoCA data.
• To compare ML-based diagnosis against the standard MoCA cutoff score.

Main Methods:

• Analysis of 38,746 clinical observations from the National Alzheimer's Coordinating Center database.
• Training of five supervised ML algorithms (XGBoost, Random Forest, SVM, Logistic Regression, KNN) on MoCA subtest scores, demographics, and cognitive metrics.
• Application of nested Repeated Grouped Cross-Validation and Youden Index for model optimization and interpretability via SHAP values.

Main Results:

• ML models consistently outperformed the conventional MoCA cutoff for detecting cognitive impairment, with XGBoost achieving the highest performance (Youden Index 0.61).
• Dementia subtype classification varied, with primary progressive aphasia best classified (Youden ≈ 0.77), followed by Lewy body dementia and Alzheimer's disease.
• Feature importance analysis identified the Cognitive Quotient as a universal predictor and highlighted disease-specific drivers like delayed recall (Alzheimer's) and verbal fluency (primary progressive aphasia).

Conclusions:

• Interpretable machine learning significantly enhances the diagnostic utility of the MoCA, surpassing traditional fixed cutoff methods.
• This AI-driven approach transforms the MoCA into a precision diagnostic aid, enabling better patient triage.
• The findings support the integration of ML for more accurate and individualized neurocognitive assessments in clinical practice.