Multi-label speech feature selection for Parkinson's Disease subtype recognition using graph model.

Parkinson's Disease (PD) is the second-most common neurodegenerative disorder. There is a certain pathological connection between PD and dysphonia. Speech signals have been successfully used to identify PD and predict its severity. Moreover, PD has several subtypes, such as tremor, freezing of gait and dysphagia. The recognition of subtypes is of great significance for the diagnosis and treatment of PD. In this paper, we consider PD subtype recognition as a multi-label learning task and try to simultaneously recognize these subtypes using speech signals. In the proposed recognition framework, multiple types of speech data are collected, such as/a/,/pa-ka-la/, etc., and different speech features are extracted from different types of speech data. The features are concatenated as the representation of speech data. Especially, a multi-label speech feature selection algorithm based on graph structure is proposed to choose the key features and followed by a multi-label classifier for PD subtype recognition. The speech samples of 70 PD patients are collected as speech corpus. Experimental results show that the proposed multi-label feature selection method can obtain higher recognition performance than other classical ones in most cases.