Real-Time Knee Angle Prediction Using EMG and Kinematic Data with an Attention-Based CNN-LSTM Network and Transfer Learning Across Multiple Datasets

Abstract

Predicting joint angles from Electromyography (EMG) signals is critical for the control of lower-limb exoskeletons. This paper presents a lightweight, attention-based CNN-LSTM network designed for real-time knee angle prediction. By implementing a multi-stage transfer learning strategy—moving from healthy subjects to those with gait abnormalities—the model achieves high accuracy with minimal subject-specific data. Our approach addresses common challenges in the field, including data scarcity and the "lazy learning" bias toward kinematic data, demonstrating robust performance across three distinct datasets.

Introduction: The Challenge of Real-Time Prediction

Accurately predicting joint angles from muscle activity (EMG) is essential for intuitive control of assistive robotics. However, three primary hurdles persist: data scarcity for pathological gaits, the need for real-time processing without "future-looking" filters, and the complexity of EMG signals which often causes models to over-rely on simpler kinematic patterns. This study proposes a solution that leverages transfer learning and a specialized neural architecture to overcome these limitations.

Methodology

The proposed system utilizes a hybrid architecture to process multi-modal inputs. Spatial features are extracted from EMG channels using Convolutional Neural Networks (CNNs), while temporal dependencies in kinematic data are captured via Long Short-Term Memory (LSTM) units.

Expand

Source, p. 7

Network Architecture

As shown in the architecture diagram, an attention mechanism is integrated to dynamically weight muscle channels, which is particularly effective for the variable signals found in abnormal gait. The model is intentionally lightweight, containing between 72k and 90k parameters, making it suitable for embedded hardware in wearable devices.

Transfer Learning Strategy

The research employs a sophisticated pipeline to ensure generalization across different subjects and conditions:

1. Primary Training: Initial optimization on the Georgia Tech dataset (healthy subjects).
2. SIC to DIC Transition: To prevent "lazy learning" where the model ignores EMG in favor of predictable joint angles, the network is first optimized using only EMG data (Single-Input Configuration) before introducing kinematic and force data (Dual-Input Configuration).
3. Fine-Tuning: Adaptation to new subjects, including those with pathologies, using as few as 3 to 7 gait cycles.

Expand

Source, p. 5

Experimental Datasets

The model's robustness was validated using three datasets:

• Georgia Tech Dataset: 22 healthy subjects providing a baseline for source training.
• UCI Lower-Limb EMG Dataset: Includes 11 healthy and 11 subjects with knee pathologies to test transferability to abnormal gait.
• SMLE Dataset: Data from a physical exoskeleton in "zero-force" mode, incorporating interaction force data.

Results and Discussion

The model demonstrated exceptional accuracy across all test scenarios. For abnormal subjects in the UCI dataset, the model achieved a 6.8% Normalized Mean Absolute Error (NMAE) using EMG alone. When historical knee angles were included, the error for healthy subjects dropped to 3.1% NMAE. The most impressive results were seen in the SMLE exoskeleton integration, reaching 1.09% NMAE for one-step-ahead prediction.

A key finding was the necessity of the attention mechanism for pathological gait; the increased variability in these signals requires the model to focus more precisely on specific muscle activations. Furthermore, the "EMG-first" training strategy proved essential in forcing the network to learn nuanced physiological mappings rather than relying solely on periodic kinematic patterns.

Conclusion

This research provides a viable framework for personalized, real-time control of rehabilitation robotics. By combining a lightweight CNN-LSTM-Attention architecture with a strategic transfer learning approach, the system achieves high precision even with limited data from subjects with walking abnormalities. Future work will explore extending this framework to more complex activities like stair climbing and active exoskeleton assistance.