Stack Ensemble Oriented Parkinson Disease Prediction Using Machine Learning Approaches Utilizing GridSearchCV-Based Hyper Parameter Tuning

Abstract

Since the coronavirus came into existence and brought the entire world to a standstill, there have been drastic changes in people's lives that continue to affect them even as the pandemic recedes. The isolation reduced physical activity and hindered access to non-COVID related healthcare during lockdown and the ensuing months brought increased attention to mental health and the neurological disorders that might have been exacerbated. One nervous system disorder that affects the elderly and needs better awareness is Parkinson's disease. We have machine learning and a growing number of deep learning models to predict, and detect its onset; their scope is not completely exhaustive and can still be optimized. In this research, the authors highlight techniques that have been implemented in recent years for prediction of the disease. Models based on the less redundantly used classifiers-naive Bayes, logistic regression, linear-support vector machine, kernelizing support vector machine, and multilayer perceptron-are initially implemented and compared. Based on limitations of the results, an ensemble stack model of hyper-tuned versions using GridSearchCV out of the top performing supervised classifiers along-with extreme gradient boosting classifier is implemented to further improve overall results. In addition, a convolutional neural network-based model is also implemented, and the results are analyzed using two epoch values to compare the performance of deep learning models. The benchmark datasets-UCI Parkinson's data and the spiral and wave datasets-have been used for machine and deep learning respectively. Performance metrics like accuracy, precision, recall, support, and F1 score are utilized, and confusion matrices and graphs are plotted for visualization. 94.87% accuracy was achieved using the stacking approach.