Parkinson's Disease Detection Using Empirical Fourier Decomposition of Speech Signal

The advancement of speech-based automated systems for Parkinson's disease (PD) detection has attracted growing attention due to their potential use in telemonitoring health systems, as they provide a reliable, quick, cost-effective, and noninvasive solution. The objective of this study is to provide an innovative method based on the Empirical Fourier Decomposition (EFD) approach to discriminate between PD patients and healthy controls using speech signals.

Speech signals produced by PD patients and healthy controls (HC) are decomposed into several modes using EFD. Resulting modes are used to automatically discriminate between PD and HC. From each of the decomposed signals, various statistical features, including mean, variance, skewness, and kurtosis, are extracted together with entropy-based measures including entropy, Renyi's entropy, and approximate entropy. Finally, the combination of features is evaluated. All experiments are performed considering isolated words and sustained vowels of the PC-GITA database. An independent test set is used to assess the generalization capabilities of the proposed method.

The Support Vector Machine (SVM) classifier obtained an accuracy of 93% using the word /apto/ and 95% using the vowel /i/ from the PC-GITA dataset. The same system, without any additional optimization or tuning, was used to classify subjects in the independent test set. The results obtained in this scenario yielded accuracies of up to 84% for isolated words. The proposed EFD technique incorporates an improved approach to segmenting the Fourier spectrum, combined with the use of a zero-phase filter bank. The segmentation strategy effectively overcomes the inconsistency issue and the trivial component problem by defining a predetermined number of components for signal decomposition. When compared to other decomposition approaches such as the Empirical Wavelet Transform (EWT), it is found that the EFD technique exhibits greater accuracy in decomposing signals with closely spaced modes.

The experimental outcome shows that our proposed EFD-based features are suitable and resilient for automatic classification between PD patients and HC subjects.