Analyzing Data Efficiency and Performance of Machine Learning Algorithms for Assessing Low Back Pain Physical Rehabilitation Exercises

TL;DR

This work tackles automated assessment of low back pain rehabilitation exercises using robot-assisted coaching by comparing data-efficient Gaussian Mixture Models and Spatio-Temporal Graph Convolutional Networks. It leverages skeleton data from Kinect v2, OpenPose, and BlazePose across the Keraal and Kimore datasets to evaluate movement quality. Findings indicate that depth sensors offer limited gains, while 3D inputs do not consistently outperform 2D in some datasets; STGCN generally yields higher accuracy with more training data, whereas GMM remains valuable for fast, data-sparse scenarios and real-time use. The study provides guidance on sensor choice, model selection, and annotation quality for at-home rehabilitation monitoring using skeleton-based data.

Abstract

Analyzing human motion is an active research area, with various applications. In this work, we focus on human motion analysis in the context of physical rehabilitation using a robot coach system. Computer-aided assessment of physical rehabilitation entails evaluation of patient performance in completing prescribed rehabilitation exercises, based on processing movement data captured with a sensory system, such as RGB and RGB-D cameras. As 2D and 3D human pose estimation from RGB images had made impressive improvements, we aim to compare the assessment of physical rehabilitation exercises using movement data obtained from both RGB-D camera (Microsoft Kinect) and estimation from RGB videos (OpenPose and BlazePose algorithms). A Gaussian Mixture Model (GMM) is employed from position (and orientation) features, with performance metrics defined based on the log-likelihood values from GMM. The evaluation is performed on a medical database of clinical patients carrying out low back-pain rehabilitation exercises, previously coached by robot Poppy.

Figures (6)

• Figure 1: Skeleton format for the three pose estimation algorithms used in the Keraal dataset. For the Kimore dataset, we only have Kinect data.
• Figure 2: Architecture of spatio-temporal graph convolutional networks. The network chains two spatio-temporal convolutional blocks (ST-Conv blocks) and a fully-connected output layer. Each ST-Conv block contains two temporal gated convolution layers and one spatial graph convolution layer in the middle. Image taken from yu2018spatio.
• Figure 3: Architecture of STGCN for physical rehabilitation assessment. The network combines STGCNs with LSTM layers as suggested in deb2022graph. Image taken from deb2022graph.
• Figure 4: F1 scores of GCN and GMM on the Keraal dataset, with different training sizes used. The data is averaged across exercises or groups. For GMM various sizes of validation sizes (for threshold defining) are deployed.
• Figure 5: Accuracy of STGCN and GMM on the Keraal dataset.