Automatic motion artifact detection in electrodermal activity signals using 1D U-net architecture.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Electrodermal activity (EDA) is crucial for assessing sympathetic nervous system function.
• Motion and noise artifacts (MNA) frequently corrupt EDA signals, especially in wearable and ambulatory systems.
• Inaccurate MNA detection can lead to erroneous physiological assessments and diagnoses.

Purpose of the Study:

• To develop and validate a deep learning-based method for automated detection of motion and noise artifacts (MNA) in electrodermal activity (EDA) signals.
• To address limitations in generalizability and real-time implementation of existing MNA detection algorithms.
• To create a computationally efficient and memory-light model suitable for embedded systems.

Main Methods:

• A one-dimensional U-Net architecture was employed for MNA detection in EDA signals.
• Spectrograms of EDA signals were utilized as input features for the deep learning model.
• The method incorporated data augmentation and was trained and validated on four distinct datasets, including two independent test sets.

Main Results:

• The proposed 1D U-Net model achieved balanced accuracies of 80.0 ± 13.7% and 75.0 ± 14.0% on two independent test datasets.
• Performance was superior or comparable to five other state-of-the-art MNA detection methods.
• The model demonstrated significantly lower computation time and required only 0.28 MB of memory, outperforming other deep learning approaches.

Conclusions:

• The developed deep learning method provides accurate and efficient automated detection of MNA in EDA signals.
• Its low memory footprint and computational requirements enable real-time implementation in resource-constrained embedded systems.
• This approach enhances the reliability of EDA signal analysis for sympathetic function assessment in various applications.