Fine Tuned Multitasking Neural Network for Parkinson's Disease Detection from Voice Recordings

Parkinson's disease (PD) is the second most prevalent neurodegenerative disorder in old age. It is characterized by symptoms such as resting tremor, rigidity, and gait disturbances. It also affects the natural production of speech, causing tremors of the voice and imprecise pronunciation, among others. Given the prevalence of speech disorders in PD, analyzing an individual's speech provides a non-invasive, cost-effective means for detection and monitoring. The objective of this paper was to take advantage of the potential of deep learning, specifically a pre-trained convolutional neural network and a multitasking approach, to classify speech recordings from PD patients and healthy controls (HC) from spectral representations. The proposed multitask analysis methodology aimed to evaluate the effectiveness of pre-trained ResNet models, fine-tuned on Spanish, Italian, and German speech databases, for both single-task and multitask classification approaches. The results indicated that multitask learning, which includes additional tasks such as vowel and sex classification, enhances the model's performance compared to monotask learning by taking advantage of shared representations across related tasks. The multitask approach showed an improvement of up to 5% in classification accuracy and the inclusion of the intermediate models for fine-tuning produced up to 10% better classification accuracy with respect to the implemented baseline. In conclusion, this work contributes to the growing body of literature demonstrating the viability of deep learning methods for non-invasive PD detection and highlights the advantages of multitask learning for pathological speech classification.