The performance of domain-based feature extraction on EEG, ECG, and fNIRS for Huntington’s disease diagnosis via shallow machine learning

Abstract

Introduction: The early detection of Huntington’s disease (HD) can substantially improve patient quality of life. Current HD diagnosis methods include complex biomarkers such as clinical and imaging factors; however, these methods have high time and resource demands. Methods: Quantitative biomedical signaling has the potential for exposing abnormalities in HD patients. In this project, we attempted to explore biomedical signaling for HD diagnosis in high detail. We used a dataset collected at a clinic with 27 HD-positive patients, 36 controls, and 6 unknowns with EEG, ECG, and fNIRS. We first preprocessed the data and then presented a comprehensive feature extraction procedure for statistical, Hijorth, slope, wavelet, and power spectral features. We then applied several shallow machine learning techniques to classify HD-positives from controls. Results: We found the highest accuracy was achieved by the extremely randomized trees algorithm, with an ROC AUC of 0.963 and accuracy of 91.353%. Discussion: The results provide improved performance over competing methodologies and also show promise for biomedical signals for early prognosis of HD.