Voice feature description of Parkinson’s disease based on co-occurrence direction attribute topology

Abstract

The incipient diagnosis of Parkinson’s disease (PD) helps to control the disease as early as possible and dysphonia is one of the early symptoms of PD. Therefore, the construction of effective voice features is of great significance. To distinguish PD patients from healthy controls, combined with the idea of formal structure analysis, a voice feature description method of co-occurrence direction attribute topology (CDAT) is proposed in this paper. Firstly, the formal context is established according to the direction information statistically obtained in the sub-region of the spectrogram to describe the correspondence between energy points and their direction attributes. Then the CDAT is constructed to obtain the coupling information between the direction attributes in the formal context. Finally, the number of connected domains in the CDAT indicating the degree of nodal coupling is extracted as structural features and input to multiple classifiers for validation. Test and cross-corpora experiments are conducted on two different Parkinsonian sustained vowel datasets, CPPDD and SPDD, achieving the best average classification accuracies of 95.84% and 93.90% with random forest, respectively. The proposed features highlight the variation of energy direction derivatives in the spectrograms through the attribute topology. The results indicate that the proposed method has good classification accuracy on different native language datasets and cross-corpora experiments, which outperforms or is comparable to the performance of latest methods used for PD classification.