MMSE-Based Dementia Prediction: Deep vs. Traditional Models.

Area of Science:

• Neurology
• Artificial Intelligence
• Gerontology

Background:

• Early dementia diagnosis is crucial for patient outcomes and reducing societal costs.
• Traditional methods using the Mini-Mental State Examination (MMSE) struggle with complex cognitive patterns.
• Limitations exist in capturing nonlinear interactions and subtle decline patterns with current statistical and machine learning approaches.

Purpose of the Study:

• To develop and evaluate a novel deep learning model for dementia prediction using item-level MMSE data.
• To compare the deep learning model's performance against traditional machine learning models (Random Forest, SVM).
• To enhance clinical interpretability through explainable AI (SHAP analysis).

Main Methods:

• A fully connected neural network was trained on item-level MMSE data from 164 participants (cognitively normal, MCI, dementia).
• The model was compared with Random Forest and SVM classifiers using accuracy, F1-score, confusion matrices, and ROC curves.
• SHAP analysis was employed to identify influential MMSE variables.

Main Results:

• The deep learning model achieved superior performance with 90% accuracy and 0.90 F1-score, outperforming Random Forest (86%) and SVM (82%).
• SHAP analysis identified immediate memory (Q11), calculation (Q12), and drawing shapes (Q17) as key predictors.
• The model demonstrated high predictive accuracy and clinical interpretability.

Conclusions:

• Deep learning offers a powerful, interpretable tool for early dementia diagnosis using MMSE data.
• Item-level analysis and explainable AI provide deeper insights than total scores alone.
• Future research requires larger, multi-institutional datasets for broader clinical validation.