Data Augmentation Strategies for Improving the Training Dataset Size of Publicly Available Parkinson’s Datasets

Résumé fourni par la source

Parkinson's disease (PD) is a neurodegenerative illness that severely impairs speech patterns, making analysis of speech vital in its diagnosis. Nonetheless, Parkinson's disease speech datasets accessible in the open domain tend to be hampered by small sample sizes, imbalance in classes, and variability limitations, making the creation of high-quality machine learning models challenging. This study discusses the effects of data augmentation processes, such as the addition of Gaussian noise, interpolation, and scaling, for improving the quality and diversity of publicly available Parkinson's speech datasets.The proposed augmentation approaches impose controlled variation in speech-relevant features to enhance mode generalization and combat overfitting. Experimental performance shows that the methods result in a 4-7% accuracy boost for classification, and underrepresented samples are especially benefitted when compared with existing methods such as random sampling, continuous wavelet transform and normalization method. Comparative comparison validates that interpolation efficiently alleviates class imbalance, Gaussian noise benefits robustness towards real-world fluctuations in speech, and scaling supports maintaining key relationships between features.The statistical parameter such as mean of original feature value and augmented show 1-2% variation, this indicates that the original information is preserved using these augmentation techniques.The results indicate that a hybrid augmentation method can improve the reliability of computer-aided Parkinson's disease detection models significantly; enabling enhanced early diagnosis and patient monitoring.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.