An explainable machine learning approach for Alzheimer's disease classification.

Area of Science:

• Neurology
• Computational Biology
• Biostatistics

Background:

• Early Alzheimer's disease (AD) diagnosis is challenging due to subtle biomarker changes.
• Machine learning (ML) offers potential for identifying AD risk but often lacks explainability.
• High dimensionality and limited datasets pose challenges in AD research.

Purpose of the Study:

• To develop and validate explainable ML models for early AD diagnosis and progression prediction.
• To identify key factors contributing to AD development using interpretable ML approaches.
• To leverage a large-scale dataset for robust AD risk assessment.

Main Methods:

• Utilized a National Alzheimer's Coordinating Center dataset (169,408 records, 1024 features) with feature space reduction.
• Trained Support Vector Machine (SVM) models for binary (NC vs. AD) and multiclass classification and progression prediction.
• Employed rule-extraction techniques (class rule mining, stable and interpretable rule set) and SHAP/LIME for model explainability.

Main Results:

• SVM models achieved high performance: 98.9% F1 score for binary classification and 90.7% for multiclass.
• SVM accurately predicted AD progression (88% F1 for binary, 72.8% for multiclass).
• Explainable AI identified critical factors: MEMORY, JUDGMENT, COMMUN, ORIENT, and the Clinical Dementia Rating tool.

Conclusions:

• Explainable ML models demonstrate high accuracy in AD diagnosis and progression prediction.
• Identified key cognitive and clinical factors crucial for understanding AD development.
• The study highlights the utility of interpretable ML in clinical decision support for Alzheimer's disease.