Abstract
Parkinson’s disease is a neurodegenerative disorder that progresses slowly and its symptoms appear over time, so its early diagnosis is not easy. A neurologist can diagnose Parkinson's by reviewing the patient's medical history and repeated scans. Besides, body movement analysts can diagnose Parkinson's by analyzing body movement. Recent research work has shown that changes in speech can be used as a measurable indicator for early Parkinson’s detection. In this work, the authors propose a speech signal-based hybrid Parkinson's disease diagnosis system for its early diagnosis. To achieve this, the authors have tested several combinations of feature selection approaches and classification algorithms and designed the model with the best combination. To formulate various combinations, three feature selection methods such as mutual information gain, extra tree, and genetic algorithm and three classifiers namely naive bayes, k-nearest-neighbors, and random forest have been used. To analyze the performance of different combinations, the speech dataset available at the UCI (University of California, Irvine) machine learning repository has been used. As the dataset is highly imbalanced so the class balancing problem is overcome by the synthetic minority oversampling technique (SMOTE). The combination of genetic algorithm and random forest classifier has shown the best performance with 95.58% accuracy. Moreover, this result is also better than the recent work found in the literature.