A comparison of feature selection methods when using motion sensors data: a case study in Parkinson's disease

Area of Science:

• Biomedical Engineering
• Neurology
• Data Science

Background:

• Parkinson's disease (PD) motor symptom quantification is crucial for treatment monitoring.
• Data-driven modeling using machine learning (ML) offers potential for objective assessment.
• Effective feature selection is vital for optimizing ML model performance in healthcare.

Purpose of the Study:

• To evaluate the impact of different feature selection methods on ML model performance for quantifying PD motor symptoms.
• To compare the efficacy of step-wise regression, Lasso regression, and Principal Component Analysis (PCA) in conjunction with various ML algorithms.
• To assess the validity, reliability, and treatment sensitivity of combined feature selection and ML approaches.

Main Methods:

• Extracted 88 spatiotemporal features from motion sensor data during hand rotation tests.
• Applied step-wise regression, Lasso regression, and PCA for feature selection.
• Utilized Support Vector Machines (SVM), Decision Trees (DT), Linear Regression, and Random Forests for symptom quantification.
• Assessed performance using correlation coefficients, Root Mean Squared Error (RMSE), and clinical rating correlations (Unified PD Rating Scale).

Main Results:

• Step-wise regression improved performance (correlation coefficients, RMSE) for most ML models (except DTs).
• Step-wise regression with SVM showed enhanced treatment sensitivity and higher correlation with clinical PD ratings compared to PCA with SVM.
• ML models showed mixed results in discriminating between PD patients and healthy controls.
• Step-wise regression emerged as the top-performing feature selection method.

Conclusions:

• The selection of feature selection methods is critical for successful data-driven modeling in PD research.
• Step-wise regression is recommended for its superior performance in quantifying PD motor symptoms.
• Optimized feature selection can improve the accuracy and clinical relevance of ML-based PD assessments.