Predicting UPDRS Scores in Parkinson’s Disease Using Voice Signals: A Deep Learning/Transfer-Learning-Based Approach

Abstract

In the last years, literature exhibits successful results in the automatic detection of Parkinson’s disease using voice/speech, especially for patients in medium or late stages of the disorder. By contrast, the prediction of the UPDRS scores -used to assess the severity of the disorder or the efficacy of treatments- has been shown to perform mostly poor. These results could be explained by the need of more complex machine learning models compared to the detection case, and the lack of large databases for properly training artificial intelligence models. To analyse possible solutions to these problems, this work will explore the potentiality of Deep Neural Network (DNN) and Convolutional Neural Network (CNN) models, along transfer learning approaches, for the automatic prediction of the UPDRS scores. Experiments are carried out using feature engineering and feature learning methodologies. In particular for feature engineering, a series of well-know features that are used to characterise vocal conditions are employed to train a DNN. Likewise, the feature learning approach is based on transformation of the input speech using Modulation spectra transformations to train a CNN, considering a transfer learning approach. For transfer learning, the networks will be trained using voice signals from patients of databases of organic and functional voice pathologies; following a network architecture that has been proven successful recently for voice quality assessment using the GRB scale. The approach includes the combination of feature learning and feature engineering approaches using a multimodal strategy. The fine-tuning procedure of the last layers in the second network will be carried out using two databases of PD patients. The results present insights about the potential of deep learning along with transfer learning strategies for the prediction of UPDRS score in parkinsonian speechs.