XGBoost-SHAP-based interpretable diagnostic framework for alzheimer's disease.

Area of Science:

• Machine learning applications in neurodegenerative disease research.
• Development of interpretable artificial intelligence models for healthcare.

Background:

• Class imbalance in Alzheimer's disease (AD) progression poses challenges for machine learning (ML)-based auxiliary diagnosis.
• Existing ML models exhibit low diagnostic performance due to data imbalance issues.

Purpose of the Study:

• To develop an interpretable framework, XGBoost-SHAP, to address class imbalance in AD progression.
• To achieve accurate multiclassification of normal cognition (NC), mild cognitive impairment (MCI), and AD.
• To identify key predictive features for clinical decision-making in AD diagnosis.

Main Methods:

• Utilized patient data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) and National Alzheimer's Coordinating Center (NACC) databases.
• Employed extreme gradient boosting (XGBoost) with sample weight adjustments for imbalanced data.
• Integrated Shapley additive explanations (SHAP) for model interpretability and feature analysis.

Main Results:

• The XGBoost-SHAP framework demonstrated superior classification performance compared to other ML models.
• Achieved high accuracy (87.57% on ADNI, 80.52% on NACC) and AUC (0.91 on ADNI, 0.88 on NACC).
• Identified top features (e.g., CDRSB, ADAS13, ventricle volume) with varying associations with AD risk.

Conclusions:

• The interpretable XGBoost-SHAP framework effectively handles imbalanced data for AD multiclassification.
• Provides valuable clinical decision-making guidance through an optimal feature subset.
• Offers new research directions for AD prevention and treatment strategies.