Balanced vs. Imbalanced Data in Parkinson's Disease Detection: A Machine Learning, Ensemble Machine Learning and Deep Learning Perspective Using SMOTE Based on Clinical Data

The study uses the UCL clinical dataset to examine the efficacy of several DL, machine learning, and ensemble learning classification models in identifying Parkinson's ailment. A pair of discrete methodologies were utilized: one that leveraged balanced data obtained by the method known as Synthetic Minority Over-sampling (SMOTE) and the other that employed unbalanced data. Based on balanced data, the K-Nearest Neighbors (KNN) model fared better than other algorithms in both methods, with a noteworthy accuracy of 98.31%, along with metrics for precision, recall, and F1-Score of 98.36%, 98.31%, and 98.31%, respectively. By contrast, the KNN model produced an accuracy of 94.87% on data that was not balanced. These results provide important new insights into Parkinson's ailment prediction techniques and highlight the role that data balance plays in improving model performance.