Attila Zoltán Jenei, Réka Ágoston, and István Valálik
A Siamese-based Approach to Improve Parkinson’s Disease Detection and Severity Prediction from Speech Using X-Vector Embedding
Parkinson’s disease is incurable and is considered one of the most common neurological diseases. It is a progressive disease, which highlights the importance of early detection. Machine learning-based diagnostic support is desirable since the diagnosis is based on history, visual inspection, and drug tests. Speech is presumed to be one of the promising biomarkers that can predict the state of the disease. Combining speech data with deep learning feature extraction in Siamese-based architecture may improve the detection compared with direct regression with acoustic and prosodic features. Read text-based speech samples were acquired from 98 patients with Parkinson’s disease and 107 healthy participants. Feature vectors were extracted with pre- trained x-vector embedding and were used directly with a support vector regressor in a nested cross-validation setup (baseline approach). Furthermore, pairs were allocated, and difference vectors were calculated. These difference vectors were then used to train support vector regressor models in nested crossvalidation (Siamese-based approach). Severity predictions and classification were performed with the outcomes. The Siamesebased setup outperformed the baseline approach both in regression and classification metrics. The relative improvement in root mean square error is 14.4%, and the Pearson correlation is 12.5% at best. After the classification, the relative improvement is 6.0% in sensitivity, 3.0% in specificity, and 4.5% in accuracy. Furthermore, comparing the test sample to not only one but multiple others decreases the average standard deviation of the predicted severity by 16.5% in relative value. Changing only the architecture of the traditional examination setup to a Siamesebased approach may increase the performance of the models.
Reference:
DOI: 10.36244/ICJ.2025.1.9
Please cite this paper the following way:
Attila Zoltán Jenei, Réka Ágoston, and István Valálik, "A Siamese-based Approach to Improve Parkinson’s Disease Detection and Severity Prediction from Speech Using X-Vector Embedding", Infocommunications Journal, Vol. XVII, No 1, March 2025, pp. 76-81., https://doi.org/10.36244/ICJ.2025.1.9