Advanced Multidimensional Speech Fusion for Early Parkinson's Diagnosis

This paper proposes a novel early detection model for Parkinson's disease based on multidimensional speech feature fusion. To overcome the limitations of single-feature or traditional machine learning methods in accuracy and stability, the study utilizes a non-invasive approach that integrates individual patient information with multimodal speech feature optimization strategies. Using a publicly available Parkinson's speech dataset consisting of 756 samples and 755 -dimensional features, a three-stage feature selection method-comprising patient ID grouping with mean aggregation, dual correlation filtering, and chi-square testing-was employed to reduce the feature space to 31 key dimensions. Data preprocessing using MinMaxScaler and PCA further enhanced the training efficiency by mitigating feature magnitude differences and redundancy. The model was developed by integrating logistic regression, XGBoost, and support vector machine classifiers, with a RandomOverSampler applied to address the imbalance between healthy and patient samples. Experimental results demonstrated that the logistic regression model achieved the best overall performance, attaining an accuracy of 88% on the validation set and outperforming other models in terms of AUC, F1-score, precision, and recall. These findings provide a promising technical framework for the development of a lowcost, high-precision portable screening tool for early Parkinson's disease diagnosis and may be extended to the early detection of other neurodegenerative disorders.