Automatic motion artifact detection in electrodermal activity data using machine learning. (April 2022)

• Record Type: Journal Article
• Authors: Hossain, Md-Billal; Posada-Quintero, Hugo F.; Kong, Youngsun; McNaboe, Riley; Chon, Ki H.
• Is Part Of: Biomedical signal processing and control. Volume 74(2022)
• Publication Date: 2022-04
• Subjects: EDA -- SCR -- Machine learning -- Motion artifact -- LOSO validation -- Feature selection; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.103483
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: We created an annotated ED) motion artifact database with simultaneous reference EDA signal. The performance of the proposed method is compared with two state-of-the-art methods. The performance of the proposed method is evaluated on an independent dataset for generalizability. The algorithm detects motion artifacts with 94.7% accuracy. Abstract: Background and objective: Electrodermal activity (EDA) has gained popularity in recent years for diverse applications such as emotion and stress recognition; assessment of pain, fatigue, and sleepiness; and diagnosis of depression and epilepsy. However, presence of motion artifacts (MA) hinders accurate analysis of EDA signals. This study presents a machine learning framework for automatic motion artifact detection on electrodermal activity signals. Methods: We extracted several statistical and time frequency features from EDA and investigated machine learning algorithms to automatically detect noisy EDA segments. To avoid incorrect adjudication due to the aperiodic nature of EDA signals, we collected both clean and MA-corrupted EDA from immobile and moving hands, respectively. The MA-corrupted EDA data were annotated by three experts as either MA-corrupted or clean using the criteria recommended in the literature, as well as the correlation between MA and the reference EDA. Results: We performed a subject-independent validation strategy to evaluate the performance of the machine learning models. The best-performing model classified the MA and clean EDA segments with 94.7% accuracy. A comparison of our motion artifact detection approach with two previously published methods showed that our best performing method outperformed them and retained its accuracy on entirely different, unseen data from a separate study, indicating the method's generalizability. Conclusions: The current work can provide accurate and autonomous adjudication of MA-corrupted EDA signals. Given the lack of accurate MA detection methods for EDA signals, this work may lead to more applications of EDA as a physio-marker.