An Efficient Model based on the k-Nearest Neighbors Algorithm for Parkinson’s Disease Detection

Background & Objectives: Parkinson’s disease (PD) is a neurological disorder characterized by the progressive loss of brain cells, significantly affecting body movement. Early diagnosis not only reduces healthcare costs but also helps prevent adverse outcomes for patients. Researchers are increasingly utilizing intelligent machine learning methods to enhance the accuracy and efficiency of PD diagnosis. Materials & Methods: Although several data mining techniques have achieved reasonable accuracy in diagnosing PD, they often encounter trade-offs between accuracy and execution speed and are sensitive to parameter settings and data outliers. The k-Nearest Neighbors (KNN) algorithm, for example, is valued for its simplicity and speed but suffers from limitations such as sensitivity to neighborhood size and reliance on majority voting, both of which can degrade performance. To address these challenges, this study employs an advanced variant of the KNN algorithm, referred to as Multiple Local Mean Vector-based Nearest Neighbor Classification (MLMV-NNC), alongside a neural network classifier trained using Bayesian backpropagation. The MLMV-NNC method enhances traditional KNN by incorporating multiple local mean vectors, thereby reducing the influence of outliers and improving classification robustness. Results: The proposed diagnostic approach demonstrates superior performance in detecting PD. Specifically, the model achieves an accuracy of 99%, precision of 96%, specificity of 98.6%, and sensitivity of 100%. Furthermore, a comparative analysis with traditional methods, including Support Vector Machines (SVM) and Artificial Neural Networks (ANN), highlights the superior performance of the proposed method. Conclusion: The findings indicate that the combination of MLMV-NNC and a neural network trained via Bayesian backpropagation constitutes a highly effective approach for diagnosing PD. This method not only improves accuracy but also mitigates common challenges such as sensitivity to parameter settings and data outliers, offering a promising alternative to conventional classification techniques.