Random Channel Ablation for Robust Hand Gesture Classification with Multimodal Biosignals

Area of Science:

• Biomedical Engineering
• Human-Computer Interaction
• Machine Learning

Background:

• Biosignal-based hand gesture classification is crucial for human-machine interaction.
• Multimodal biosignal sensing often suffers from data loss due to missing channels, impacting classification performance.
• Developing robust classifiers for incomplete biosignal data is a significant challenge.

Purpose of the Study:

• To propose and evaluate a novel method, Random Channel Ablation (RChA), to enhance the robustness of hand gesture classifiers against missing data channels.
• To assess the effectiveness of RChA in multimodal biosignal classification using ultrasound and force myography (FMG).

Main Methods:

• Acquired multimodal biosignal data (ultrasound and FMG) from the forearm for 12 hand gestures from 2 subjects.
• Implemented Random Channel Ablation (RChA) during the training of a convolutional neural network architecture.
• Compared RChA against baseline, imputation, and oracle methods using 5-fold cross-validation.

Main Results:

• RChA demonstrated significant improvements in gesture classification accuracy compared to the baseline method.
• An average improvement of 12.2% and 24.5% was observed with up to 4 and 8 missing channels, respectively.
• The proposed RChA method showed superior robustness to increasing numbers of missing channels compared to other evaluated methods.

Conclusions:

• Random Channel Ablation (RChA) is an effective strategy for improving classifier robustness in multimodal, multi-channel biosignal-based hand gesture classification.
• RChA offers a promising approach to mitigate the adverse effects of data loss in real-world biosignal acquisition systems.
• The findings highlight the potential of RChA for enhancing the reliability of human-machine interfaces reliant on biosignal interpretation.