Transfer Learning for Efficient Intent Prediction in Lower-Limb Prosthetics: A Strategy for Limited Datasets

Area of Science:

• Biomedical Engineering
• Rehabilitation Technology
• Machine Learning in Healthcare

Background:

• Locomotion intent prediction is crucial for advanced prosthetic control in transfemoral amputees.
• Data scarcity poses a significant challenge for developing accurate subject-specific predictive models.
• Existing methods often require extensive user-specific data, limiting their applicability.

Purpose of the Study:

• To develop and evaluate a transfer learning approach for enhancing locomotion intent prediction in transfemoral amputees, especially in data-limited scenarios.
• To compare the efficacy of transfer learning against traditional subject-dependent models.
• To identify optimal sensor configurations for intent prediction.

Main Methods:

• Utilized transfer learning by fine-tuning three pre-trained models (transfemoral amputees, able-bodied, mixed dataset) on new transfemoral amputee data.
• Compared performance against subject-dependent models requiring extensive individual training data.
• Investigated the impact of varying sensor configurations, including thigh IMU and load cell combinations.

Main Results:

• Transfer learning achieved comparable error rates to subject-dependent models with significantly less data.
• Model performance improved with increased subject-specific data availability.
• A combination of thigh inertial measurement unit (IMU) and load cell proved to be a practical and efficient sensor setup.

Conclusions:

• Transfer learning offers a viable and data-efficient solution for improving locomotion intent prediction in transfemoral amputees.
• Pre-existing, pre-trained features are beneficial in data-scarce environments.
• Optimized sensor configurations enhance the practical implementation of these prediction systems.