A novel feature extraction method based on TQWT and instantaneous energy variation for Parkinson’s disease detection

Abstract

Voice disorder is an early symptom of Parkinson’s disease (PD), and extracting effective voice features is of great significance for PD detection. The voice features based on tunable Q-factor wavelet transform (TQWT) have promising performance in PD detection only using conventional energy and entropy features. To further develop the potential of TQWT, a novel feature extraction method named instantaneous energy variation based on TQWT (IEV-TQWT) is proposed. Firstly, TQWT is adopted to perform multi-scale decomposition to obtain the oscillation information of the voice signals. Next, IEV-TQWT features are obtained by calculating instantaneous energy variation vectors from the spectrograms of sub-signals. Finally, the IEV-TQWT features are input to multiple classifiers to verify the PD classification performance. The results have shown that IEV-TQWT features with explicit physical information have a significant improvement in classification performance. In addition, IEV-TQWT features outperform other advanced features on the CPPDD dataset and reach mainstream level on SPDD dataset, indicating that IEV-TQWT features can be used as an effective assisted diagnosis tool of PD.