Enhancing Joint Motion Prediction for Individuals with Limb Loss Through Model Reprogramming

TL;DR

This work tackles amputee gait prediction under sparse data by applying model reprogramming to repurpose a foundation model trained on able-bodied data, without altering network parameters. It introduces a three-part framework: a foundation module with a shared core $g_s(\cdot; \theta_s)$ and task heads $g_t(\cdot; \theta_t)$, a template mapping that derives a correction input $X_{corr}$ from amputee data, and a refurbish module $h(\cdot; \Theta_h)$ that maps amputee inputs $X_{amp}$ to $X_{corr}$ for accurate predictions of missing-limb motion. The method outperforms cross-mapping and direct-mapping baselines, particularly when training data from amputees are limited, and demonstrates the ability to adapt pre-trained models for prosthetic control without retraining. This approach offers a data-efficient path to improve assistive technologies and amputee mobility by leveraging existing gait models across domains.

Abstract

Mobility impairment caused by limb loss is a significant challenge faced by millions of individuals worldwide. The development of advanced assistive technologies, such as prosthetic devices, has the potential to greatly improve the quality of life for amputee patients. A critical component in the design of such technologies is the accurate prediction of reference joint motion for the missing limb. However, this task is hindered by the scarcity of joint motion data available for amputee patients, in contrast to the substantial quantity of data from able-bodied subjects. To overcome this, we leverage deep learning's reprogramming property to repurpose well-trained models for a new goal without altering the model parameters. With only data-level manipulation, we adapt models originally designed for able-bodied people to forecast joint motion in amputees. The findings in this study have significant implications for advancing assistive tech and amputee mobility.

Figures (2)

• Figure 1: Illustration of computation of the correction input $X_{corr}$ corresponding to the $k$-th input sample $X^k_{amp}$ of the amputee. The able-bodied input $X^i_{ab}$ that produces the most similar output as that of the desired amputee output $y_{amp}^k$ is searched in the input-output space of the trained able-bodied foundation module. Instead of searching based on a single desired motion variable $y_{amp}^k$, a sequence of values $\{y^{k-m}_{amp}, ..., y^k_{amp}, ..., y^{k+m}_{amp}\}$ (marked by the red region in the right) is used and the able-bodied input $X^i_{ab}$ corresponding to the midpoint of the sequence is considered. Further, a neighborhood of radius $\epsilon$ is considered around $X^i_{ab}$ and the correction input $X_{corr}$ is computed as a weighted sum of samples in this neighborhood with weights decreasing (linearly or exponentially) with increasing distance from the center $X^i_{ab}$.
• Figure 2: (Left) The computed correction template $X_{corr}$ and corresponding prediction by the refurbish module. (Right) Performance of models trained with direct mapping and refurbishing. For refurbishing, $2m+1=1$, $n=5$ with linear weighting, $\alpha=1$ and $\beta$ = 20 were selected based on an extensive grid search. where $2m+1$ is the length of the sequence considered for template matching and $n$ is the number of closest neighbors within the $\epsilon$-neighborhood used for computing the correction template.