What question did this study set out to answer?

This research aims to evaluate the efficacy of a deep learning-based motion analysis system in an exercise training program for patients with low back pain.

April 16, 2026

Intervention Effect of Exercise Training Program Using Deep Learning-based Motion Analysis System on Patients with Low Back Pain

Key Points

This research aims to evaluate the efficacy of a deep learning-based motion analysis system in an exercise training program for patients with low back pain.
Conducted a randomized controlled trial with 118 patients
Compared an experimental group using a motion analysis system to a control group with conventional training
Measured outcomes using visual analog scale (VAS), Oswestry disability index (ODI), and electromyography (sEMG)
Assessed system performance via pose estimation accuracy
Experimental group showed significantly improved VAS scores (2.6 vs. 3.9) and ODI scores (18.7 vs. 28.3) compared to control (P < 0.05)
Movement quality score was significantly higher in the experimental group (88.5 points) than in control (76.8 points)
sEMG analysis revealed better core muscle coordination in the experimental group
Higher training adherence rate (92.5% vs. 85.3%) in the experimental group

Abstract

Low back pain (LBP) is a prevalent and disabling musculoskeletal disorder that severely affects patients’ quality of life. Traditional rehabilitation training relies on therapists’ subjective observations, which have issues such as imprecise assessment and untimely feedback. A deep learning-based motion analysis system was developed to deliver an intelligent exercise training program with real-time, precise feedback. Through a randomized controlled trial (single-center design), 118 patients with chronic non-specific LBP were randomly assigned to an experimental group (receiving intelligent training based on the motion analysis system) and a control group (receiving conventional guided training). Outcome measures included the visual analog scale (VAS), Oswestry disability index (ODI), surface electromyography (sEMG) signals of core muscles, and system-generated movement quality scores. System performance tests indicated that the pose estimation model achieved a mean Average Precision (AP@0.5) of 96.7% in keypoint detection, with a mean per joint position error (MPJPE) of 25.3 mm and a pelvic angle estimation error of 3.1°, providing a reliable technical foundation for precise motion analysis. Supported by this system, after the 8-week intervention (T1), improvements in both VAS scores (2.6 ± 1.1 vs. 3.9 ± 1.3) and ODI scores (18.7 ± 6.3 vs. 28.3 ± 7.6) were significantly greater in the experimental group compared to the control group (P < 0.05), and this advantage was maintained at the 1-month follow-up after the intervention ended (T2). The average movement quality score during the entire intervention period was also significantly higher in the experimental group (88.5±4.2 points) than in the control group (76.8±7.5 points). The sEMG results indicated more coordinated activation and less compensation in the core muscles of the experimental group. The training adherence rate was significantly higher in the experimental group (92.5%) compared to the control group (85.3%). The deep learning-based motion analysis system developed in this study is highly accurate and robust, capable of providing precise, real-time feedback for rehabilitation training for patients with LBP. It significantly alleviates pain, improves functional impairments, and optimizes movement patterns and neuromuscular control. This innovative rehabilitation method has good clinical application prospects.

Mark Helpful

Bookmark

Relay