Sagittal-Plane Knee Flexion Moment Estimation Using a Lightweight Deep Learning Framework Based on Sequential Surface...

Area of Science:

• Biomechanics
• Wearable Technology
• Machine Learning

Background:

• Knee joint moment estimation is crucial for sports, rehabilitation, and human-machine interaction but often requires lab settings.
• Surface electromyography (sEMG) provides a non-invasive method for estimating joint moments.
• Existing deep learning models for sEMG-based estimation are often too computationally intensive for edge deployment.

Purpose of the Study:

• To propose Topo2DCNN-LSTM, a computationally efficient deep learning model for estimating sagittal-plane knee flexion moment.
• To enable on-device inference of knee joint moments using sEMG data.
• To demonstrate the feasibility of personalized knee moment estimation in resource-constrained environments.

Main Methods:

• Developed a lightweight two-dimensional convolutional neural network (2D CNN) integrated with long short-term memory (LSTM) units.
• Utilized a feature-based sequential representation, converting raw sEMG signals into Root Mean Square (RMS) feature frames.
• Trained and evaluated the model on a public walking dataset with synchronized sEMG and kinetic data at varying treadmill speeds.

Main Results:

• The quantized Topo2DCNN-LSTM model achieved a mean RMS Error of 0.088 ± 0.020 Nm/kg at 1.2 m/s and 0.114 ± 0.034 Nm/kg at 1.8 m/s.
• Demonstrated an average inference latency of 28 ms on a SparkFun Thing Plus-SAMD51, with low RAM and flash memory usage.
• Validated the model's effectiveness for on-device, isolated inference in resource-limited conditions.

Conclusions:

• Topo2DCNN-LSTM is a viable, lightweight deep learning solution for estimating knee flexion moment from sEMG.
• The model's efficiency supports its deployment on wearable devices for real-time biomechanical analysis.
• This approach offers a proof of concept for personalized, on-device joint moment estimation in practical settings.