Assessing Low Back Movement with Motion Tape Sensor Data Through Deep Learning

TL;DR

This work tackles remote assessment of low back movements using a low-cost MT wearable by introducing MT-AIM, a deep learning pipeline that combines data augmentation with conditional generative models and feature augmentation through MT-to-kinematics translation. By generating synthetic MT samples and predicting joint angles conditioned on MT data, then applying DTFT-based features, MT-AIM achieves strong classification performance across six movements, even with a small, noisy dataset. The study demonstrates that MT-based sensing, when paired with advanced generative and feature-enhancement strategies, can rival higher-end MoCap approaches for remote movement analysis, with near-perfect classification accuracy on the tested tasks. The findings suggest substantial potential for remote LBP assessment and PT personalization, while acknowledging the need for larger, diverse clinical datasets and real-world validation to ensure generalizability and practical utility.

Abstract

Back pain is a pervasive issue affecting a significant portion of the population, often worsened by certain movements of the lower back. Assessing these movements is important for helping clinicians prescribe appropriate physical therapy. However, it can be difficult to monitor patients' movements remotely outside the clinic. High-fidelity data from motion capture sensors can be used to classify different movements, but these sensors are costly and impractical for use in free-living environments. Motion Tape (MT), a new fabric-based wearable sensor, addresses these issues by being low cost and portable. Despite these advantages, novelty and variability in sensor stability make the MT dataset small scale and inherent to noise. In this work, we propose the Motion-Tape Augmentation Inference Model (MT-AIM), a deep learning classification pipeline trained on MT data. In order to address the challenges of limited sample size and noise present within the MT dataset, MT-AIM leverages conditional generative models to generate synthetic MT data of a desired movement, as well as predicting joint kinematics as additional features. This combination of synthetic data generation and feature augmentation enables MT-AIM to achieve state-of-the-art accuracy in classifying lower back movements, bridging the gap between physiological sensing and movement analysis.

Figures (8)

• Figure S1: Generative and classification data flow for MT-AIM. Processed MT and MoCap kinematics data are used to train synthetic MT and kinematics generators. The trained models are sampled to create synthetic MT and kinematics data. A DTFT is taken of all MT data. Finally, real, transformed, and synthetic data is all combined before being used to train a classifier.
• Figure S2: Left: The placement of six MTs (black stripes) and retroreflective markers used for 3D optical MoCap (gray circles). Right: The six movements tested with human subjects: standing extension, standing forward flexion, standing lateral flexion (left and right), and seated rotation (left and right).
• Figure S3: Example MT normalized resistance for each movement type, repeated three times.
• Figure S4: Left: Sample of processed MT data showing individual sensors. Right: Sample of corresponding MoCap processed kinematics data showing Euler angles between the lower lumbar segment (L4--L5) and pelvis segments, and between the upper lumbar segment (L1--L3) and lower lumbar segment (L4--L5).
• Figure S5: General data flow and architecture for the kinematics generators. (a) An MT-conditioned Transformer Encoder produces an embedding that is summed with a Gaussian latent representation of the kinematics learned by a feedforward encoder. The combined latent is decoded into interpretable kinematic outputs, and the resulting loss is used for gradient-based optimization. (b) The learned kinematics latent is replaced with randomly sampled Gaussian noise $\epsilon$, which is combined with the MT embedding and decoded to generate kinematics.