Automatically Predicting Severity of Parkinson's Disease Using Model Based on XGBoost from Speech

Self-service assessment model for measuring severity of Parkinson's disease (PD) has recently received considerable attention due to the inconvenience and cost of physical visits to the medical clinic for PD patients. Previous works, however, mostly focus on the feature extraction and classifier design without considering notable intrinsic differences between patients with Parkinsonism (PWP), which has an adverse impact on generalization of model detecting PD. This paper introduces a novel PD-detected model based on Extreme Gradient Boost (XGBoost), where prior knowledge including gender and age are used to predict the severity of PD through unified Parkinson's disease rating scale (UPDRS). Firstly, using gender and age as a prior knowledge decomposes the prediction model of UPDRS to get new sub-model. Secondly, we obtain a regressor trained by XGBoost on the respective sub-model. Finally, we can get the final prediction about UPDRS using trained model. The experimental results show that our method significantly improves performance on the remote Parkinson dataset.