Abstract: Deep learning has gained popularity in detecting Parkinson's disease (PD) from speech due to its ability to automatically extract meaningful representations from raw data. The most popular approaches are based on Convolutional Neural Network (CNN) models fed with spectrograms. However, the use of these algorithms is constrained due to the cross-dataset accuracy obtained during the validation process. Thus, in this work, we focus on studying the cross-domain effect -specifically due to different databases- for the screening of PD using a CNN-based model and two different speech corpora. To address the cross-domain challenge, we propose the use of domain adversarial (DA) training as a method to obtain discriminant and domain-invariant models. The visualization of the feature space distribution extracted by this model, using t-distributed stochastic neighbor embeddings, along with its divergence and variance by class, indicates a significant improvement in domain adaptation. These initial results provide valuable insights for further model refinement and constitute a proof of concept that domain adversarial methods offer a feasible option for creating a more generalizable speech-based PD detection model.
Keywords: Convolutional Neural Networks, Deep learning, Domain Adversarial, Parkinson’s Disease.
[1] E. J. Ibarra-Sulbaran, J. D. Arias-Londoño, J. I. Godino-Llorente. "DOMAIN ADVERSARIAL CONVOLUTIONAL NEURAL NETWORK FOR PARKINSON’S DISEASE DETECTION FROM SPEECH." In Proc. Models and Analysis of Vocal Emissions for Biomedical Applications (MAVEBA). 13th International Workshop, Firenze, Italy.