Empirical Bayes Transfer Learning for Uncertainty Characterization in Predicting Parkinson’s Disease Severity

Abstract Parkinson’s disease (PD) is a common neurodegenerative disorder disease. Identifying an accurate model to predict severity level is critical to prevent severe suffering for PD patients. However, the existing research does not consider the heterogeneity among patients, falls short for prediction uncertainty characterization and typically adopts models that involve tedious parameter tuning processes. We propose to incorporate transfer learning and sparse learning under a Hierarchical Bayesian framework for tractable estimation of parameter posterior distribution and prediction uncertainty quantification. Specifically, we develop an empirical Bayes transfer learning (ebTL) model that accounts for patient heterogeneity and meanwhile allows for knowledge transfer between the modeling processes of different patients. ebTL is also featured for automatic hyper-parameters estimation without a tedious tuning process. Finally, we present an application of predicting PD severity level by using features extracted from speech signals across PD patients. The model could achieve better prediction accuracy compared with the other two competing methods and enable reasonable quantification of prediction intervals.