Addressing voice recording replications for tracking Parkinson’s disease progression

Tracking Parkinson’s disease symptom severity by using characteristics automatically extracted from voice recordings is a very interesting and challenging problem. In this context, voice features are automatically extracted from multiple voice recordings from the same subjects. In principle, for each subject, the features should be identical at a concrete time, but the imperfections in technology and the own biological variability result in nonidentical replicated features. The involved within-subject variability must be addressed since replicated measurements from voice recordings can not be directly used in independence-based pattern recognition methods as they have been routinely used through the scientific literature. Besides, the time plays a key role in the experimental design. In this paper, for the first time, a Bayesian linear regression approach suitable to handle replicated measurements and time is proposed. Moreover, a version favoring the best predictors and penalizing the worst ones is also presented. Computational difficulties have been avoided by developing Gibbs sampling-based approaches.