Transformer-Based Context-Informed Incremental Learning With sDTW Alignment Unlocks Fast and Precise Regression-Based...

Area of Science:

• Biomedical Engineering
• Neuroscience
• Human-Computer Interaction

Background:

• Regression-based myoelectric interfaces offer intuitive control but face challenges with calibration, dynamics, and user consistency.
• Temporal neural architectures can improve controllers by learning user behavior patterns, but require representative closed-loop training data.
• Context-informed incremental learning (CIIL) acquires data online but struggles with temporal deviations between assumed and true user intent.

Purpose of the Study:

• To introduce T-sDTW-CIIL, a novel transformer-based incremental learning framework for myoelectric control.
• To integrate temporal modeling, closed-loop learning, and soft dynamic time warping (sDTW) for tolerant label alignment.
• To evaluate T-sDTW-CIIL's performance against traditional methods in an adaptive cursor-control task.

Main Methods:

• Developed T-sDTW-CIIL, a transformer-based incremental learning framework incorporating temporal modeling and sDTW.
• Recruited twelve participants for a regression-based cursor-control task using static and CIIL variants of MLP and transformer models.
• Assessed performance in a high-precision ISO-Fitts' environment, measuring success rates, throughput, efficiency, and simultaneity.

Main Results:

• T-sDTW-CIIL demonstrated significantly higher success rates, throughputs, efficiencies, and simultaneity gains compared to baseline MLP.
• Achieved 2.0x, 2.4x, and 3.7x higher throughputs for large, medium, and small targets, respectively, versus static MLP.
• Maintained a 98.4% success rate for small targets, while static MLP degraded to 23.4%; reduced contraction intensity by ~10%.

Conclusions:

• T-sDTW-CIIL effectively combines temporal learning with context-informed co-adaptation, overcoming limitations of existing myoelectric controllers.
• The framework enables robust, low-intensity human-computer interaction through improved real-time adaptation and user intent alignment.
• Results highlight the potential of advanced temporal neural architectures for next-generation prosthetic and assistive devices.