Personalized Control for Lower Limb Prosthesis Using Kolmogorov-Arnold Networks

TL;DR

This work addresses turn-intent recognition for a semi-active lower-limb prosthesis by evaluating Kolmogorov-Arnold Networks (KAN) and their deep-learning variant FKAN against conventional models (MLP) and CNNs, using IMU-derived data from five amputees. It investigates two questions: whether learnable activation functions improve turn-prediction performance, and whether subject-specific vs pooled training data yields better macro-average F1-scores. Across the study, learnable activations did not significantly outperform static activations; however, subject-specific training improved performance for conventional ML models, while DL models achieved comparable results with pooled data. CNN/FKAN generally dominated in several cases, but inference speed and data size limited practical deployment for KAN-based approaches. The findings guide model selection and data strategy for prosthesis controllers, suggesting that data pooling can be viable for DL-based approaches, while careful consideration of task complexity and real-time constraints is crucial for deploying adaptive activation functions in embedded devices.

Abstract

Objective: This paper investigates the potential of learnable activation functions in Kolmogorov-Arnold Networks (KANs) for personalized control in a lower-limb prosthesis. In addition, user-specific vs. pooled training data is evaluated to improve machine learning (ML) and Deep Learning (DL) performance for turn intent prediction. Method: Inertial measurement unit (IMU) data from the shank were collected from five individuals with lower-limb amputation performing turning tasks in a laboratory setting. Ability to classify an upcoming turn was evaluated for Multilayer Perceptron (MLP), Kolmogorov-Arnold Network (KAN), convolutional neural network (CNN), and fractional Kolmogorov-Arnold Networks (FKAN). The comparison of MLP and KAN (for ML models) and FKAN and CNN (for DL models) assessed the effectiveness of learnable activation functions. Models were trained separately on user-specific and pooled data to evaluate the impact of training data on their performance. Results: Learnable activation functions in KAN and FKAN did not yield significant improvement compared to MLP and CNN, respectively. Training on user-specific data yielded superior results compared to pooled data for ML models ($p < 0.05$). In contrast, no significant difference was observed between user-specific and pooled training for DL models. Significance: These findings suggest that learnable activation functions may demonstrate distinct advantages in datasets involving more complex tasks and larger volumes. In addition, pooled training showed comparable performance to user-specific training in DL models, indicating that model training for prosthesis control can utilize data from multiple participants.

Figures (5)

• Figure 1: VSTA positioned below the socket provided variable torsional stiffness. Simulated IMU at proximal end of VSTA.
• Figure 2: Illustration of a single-step 90° spin turn, showing the planted foot on the inside of the turn. The right leg represents the prosthesis and the left leg represents the intact limb without sensor.
• Figure 3: Model optimization and statistical analysis for hypothesis testing. Hypothesis 1 compares the 10 test data divisions between KAN and MLP (left panel). Hypothesis 2 compares the 10 test data divisions using pooled training from a single model (e.g., KAN) (right panel) with the 10 divisions using specific training from the same model (left panel).
• Figure 4: Violin plots show the distribution of macro F1 scores for participant-specific models (Hypothesis 1). The left plot compares MLP and KAN, while the right compares CNN and FKAN. Horizontal bars indicate the median and interquartile range. p-values assess statistical significance for each participant.
• Figure 5: Comparison of Subject-Specific vs. Pooled Models Across Models (Hypothesis 2). Violin plots show the distribution of macro F1 scores for each participant across different models. Each subplot compares participant-specific training with pooled training (yellow) for MLP, KAN, CNN, and FKAN. Horizontal bars indicate the median and interquartile range. p-values assess statistical significance for each participant.