Abstract
Background: Parkinson's disease (PD) is a neurodegenerative disorder that affects millions of people worldwide. Speech analysis has emerged as a non-invasive tool for automatic PD detection; however, the scarcity and homogeneity of available datasets often limit the generalization capability of machine learning models, motivating the use of data augmentation strategies to improve robustness. Methods: This study presents a data augmentation-based methodology for speech-based classification between PD patients and healthy control subjects. A deep learning model trained from scratch on Mel spectrograms is evaluated using augmentation techniques applied at both the waveform and time-frequency levels. Multiple training and model selection strategies are analyzed and model performance is assessed through internal validation as well as using an independent dataset Results: Experimental results show that carefully selected data augmentation techniques improve classification performance with respect to the non-augmented counterpart, achieving gains of up to 3% in accuracy. However, when evaluated on an independent dataset, these improvements do not consistently translate into better generalization. Conclusions: These findings demonstrate that, while data augmentation can effectively enhance model performance within a single dataset, this apparent robustness is not sufficient to guarantee generalization on independent speech corpora for PD detection.