Scalable random forest for plug-and-play myoelectric control

Abstract

Myoelectric control systems translate electromyography (EMG), the non-invasive muscle electrical signals detected through cutaneous electrodes, into control commands for external devices, enabling human-machine interactions across diverse applications. Despite remarkable advancements, myoelectric control systems still face challenges, primarily due to the inherent EMG variability stemming from multiple confounding factors, including inter-user physiological differences, long-term signal drift, dynamic arm positions, and noise contamination. While state-of-the-art decoders have shown excellent performance in controlled settings, their low robustness, high computational complexity, lack of explainability, and heavy reliance on extensive calibration data limit their practical implementation in real-world applications.

The doctoral research presented in this thesis aims to find a practical decoder for myoelectric control that addresses the above-mentioned challenges. Specifically, this thesis presents a systematic investigation of simple but powerful random forest (RF) as an alternative decoder, unlocking the underutilized strengths of RF models to develop robust, adaptable, inter-user generalizable, explainable, easily parallelizable, and computationally efficient myoelectric control systems. We introduce five key innovations: (1) explainable deep forest models for high-density EMG (HD-EMG) processing, combining inherent noise resilience with physiologically consistent model interpretations. (2) a pre-trainable RF framework enabling one-shot personalization by grafting, pruning and appending decision trees; (3) a self-calibrating RF framework that progressively adapts to long-term (up to 5 weeks) EMG variations without labeled calibration data; (4) posture-invariant RF models demonstrating inherent robustness to arm position changes; and (5) scalable RF that achieves ≈500× model size compression with negligible accuracy loss. Extensive experiments including intra-day/long-term testing scenarios, with data from 106 participants, were conducted to validate above innovations.

This research establishes scalable RF as a viable and practical foundation for plug-andplay myoelectric interfaces, simultaneously satisfying four key requirements often considered mutually exclusive in the field: (1) high classification accuracy, (2) computational efficiency, (3) physiological explainability, and (4) minimal calibration demands. The proposed solutions bridge the critical gaps between laboratory demonstrations and real-world settings.