Utility of Inter-subject Transfer Learning for Wearable-Sensor-Based Joint Torque Prediction Models

Area of Science:

• Robotics
• Machine Learning
• Human-Computer Interaction

Background:

• Generalizability of models across individuals is a major challenge in human subjects research, particularly for wearable-sensor-controlled exoskeleton devices.
• Enabling widespread adoption of these technologies requires models that can generalize across subjects or be fine-tuned for individual users.

Purpose of the Study:

• To investigate the utility of single- and multi-subject parameter transfer on Long Short-Term Memory (LSTM) models for sensor-to-joint torque prediction.
• To evaluate the impact of parameter transfer on task performance and computational resources for network training.

Main Methods:

• Utilized transfer learning techniques with LSTM models for sensor-to-joint torque prediction.
• Investigated parameter transfer from single-subject and multi-subject models to target subjects.
• Assessed performance metrics and computational costs associated with different transfer scenarios.

Main Results:

• Parameter transfer from both single- and multi-subject models demonstrated effective knowledge transfer.
• Performance varied across different source and target subject pairings, indicating nuanced transferability.
• Transfer learning offers potential for reduced model training time and computational expenses.

Conclusions:

• Parameter transfer is a viable strategy for enhancing model generalizability in wearable-sensor exoskeleton applications.
• Further research into the factors influencing performance variance across subject pairings can optimize personalization and minimize data requirements.
• This approach can facilitate the development of personalized wearable-sensor-based joint torque prediction technologies.