DWT-OEFS: discrete wavelet transform based optimized ensemble feature selection for Parkinson's disease severity...

Parkinson's disease (PD) is a cognitive degenerative condition of central nervous system which highly impacts the motor function, resulting in gait dysfunction. Determining the severity of PD is essential for timely and efficient medical management. Doctors often utilize clinical manifestations to grade the severity of PD using Hoehn & Yahr scale where their evaluation is heavily reliant on skill and experience. We propose an optimized ensemble metaheuristic-based feature selection framework by utilizing the signal processing techniques to grade the severity of PD on publicly available Physionet gait Vertical Ground Reaction Force dataset obtained using wearable device. Due to scarcity of medical dataset, the sample size is increased by segmentation of signal. Discrete wavelet transform (DWT) decomposes the signal and a total of 13 features including statistical, frequency and entropy-base are extracted. For an optimum subset of features, three bio-inspired metaheuristic algorithms Binary Grey Wolf Optimization, Binary Whale Optimization and Binary Dragonfly algorithm are used for optimized ensemble feature selection (OEFS) to prevent dimensionality curse thereby improving the classification accuracy. Further, the class imbalance issue is addressed via SMOTETomek and the selected features are then subjected to four best performing classifiers and weighted voting-based classifier. The suggested model is assessed using variety of performance assessment techniques like accuracy, precision, recall, F1-score and Mathew's Correlation Coefficient. The ensemble model achieves the maximum classification accuracy of 98.56% for multiclass classification through weighted voting. Our proposed approach outperforms existing models and individual classifiers, demonstrating its ability to accurately forecast and classify PD severity.