Genetic Algorithm-Optimized support vector machine for MRI-based multiple sclerosis diagnosis from white matter...

Area of Science:

• Medical imaging analysis
• Machine learning in healthcare
• Neurological disorder diagnostics

Background:

• Multiple sclerosis (MS) diagnosis relies on identifying white matter lesions.
• Developing accurate and efficient diagnostic tools for MS is crucial.
• Machine learning offers potential for improving diagnostic accuracy and efficiency.

Purpose of the Study:

• To evaluate a genetic algorithm-assisted support vector machine (GA-SVM) framework for multiple sclerosis (MS) diagnosis.
• To assess the performance of GA-SVM using pre-extracted white matter lesion features.
• To compare the GA-SVM framework with other machine learning models for MS classification.

Main Methods:

• Utilized a publicly available dataset (n=709) with five pre-extracted white matter lesion features.
• Employed repeated nested stratified 10x5 cross-validation for robust model assessment.
• Optimized hyperparameters within inner training folds and reported aggregated out-of-fold predictions.

Main Results:

• The tuned Radial Basis Function-SVM (RBF-SVM) achieved high performance metrics: AUC=0.993, accuracy=0.987, sensitivity=1.000, specificity=0.973, precision=0.976, and F1-score=0.988.
• Precision-recall analysis (AP≈0.99) and calibration analysis (Brier=0.012) indicated strong discrimination and reliable probability estimates.
• Comparative benchmarking demonstrated competitive performance while maintaining model simplicity and interpretability.

Conclusions:

• The GA-SVM framework is a feasible approach for MS classification using compact radiomics.
• Rigorously validated classical machine learning models can effectively classify MS.
• Further external multi-site validation is necessary to confirm the generalizability of these findings.