Subitem-level multi-scale assessment and machine learning for three-class cognitive status classification in...

Area of Science:

• Neurology
• Artificial Intelligence
• Biomedical Informatics

Background:

• Cognitive impairment is common in Parkinson's disease (PD), necessitating accurate early classification for effective intervention.
• Distinguishing between normal cognition (PD-NC), mild cognitive impairment (PD-MCI), and dementia (PDD) is crucial for patient management.

Purpose of the Study:

• To develop and validate a machine learning framework for classifying cognitive states in Parkinson's disease patients.
• To enhance model interpretability using SHapley Additive exPlanations (SHAP) for clinical relevance.

Main Methods:

• Utilized data from the Parkinson's Progression Markers Initiative (PPMI) cohort.
• Developed a two-stage ensemble model (XGBoost and MLP) with SMOTE-Tomek for class imbalance.
• Employed strict hold-out validation and SHAP for feature importance analysis.

Main Results:

• Achieved strong and balanced classification performance across all three cognitive subgroups (PD-NC, PD-MCI, PDD).
• Demonstrated high accuracy in identifying cognitively impaired individuals, with an AUC exceeding 0.85 for three-class discrimination.
• SHAP analysis confirmed the clinical significance of predictors like MoCA scores and activities of daily living assessments.

Conclusions:

• The proposed explainable two-stage model offers robust cognitive stratification for Parkinson's disease patients.
• The model shows potential as a scalable tool for early dementia risk identification and clinical decision support in neurology.
• External validation on diverse cohorts is recommended prior to widespread clinical implementation.