Scholarlink

Background <p>Understanding how specific exercises induce musculoskeletal responses is central to sports science and rehabilitation. While X-ray imaging effectively evaluates exercise-induced kinematic changes such as the path of the instantaneous center of rotation (PICR), it requires radiation exposure, limiting its clinical utility. This study investigated whether machine learning models using non-invasive sensor data could accurately predict X-ray-derived PICR changes for individualized exercise selection.</p> Methods <p>Forty-five healthy participants performed three shoulder internal rotation exercises: Belly Press, Lift Off, and Prone Wiper. During the exercises, muscle activity was recorded via surface electromyography (EMG) from four muscles, and muscle thickness was measured using ultrasound. Pre- and post-exercise X-ray imaging determined PICR displacement as the reference outcome. Linear regression and four nonlinear models (including Multilayer Perceptron [MLP] and Support Vector Regression) were evaluated using ninefold cross-validation to predict PICR changes and identify the optimal exercise for each subject.</p> Results <p>Nonlinear models consistently outperformed linear regression in estimating X-ray-derived pre-post changes across all sensing modalities, demonstrating lower mean absolute errors. In subject-specific optimal exercise selection, nonlinear models showed higher agreement with the X-ray reference. Notably, the EMG-only MLP model accurately identified the optimal exercise in 97.8% (44 of 45) of participants, performing equally to models utilizing combined EMG and ultrasound features.</p> Conclusions <p>Nonlinear machine learning models effectively capture subject-specific musculoskeletal responses beyond linear mappings. Relying solely on accessible EMG features, the proposed framework provides highly accurate, individualized exercise selection aligned with X-ray reference measures, offering a promising decision-support tool for personalized rehabilitation without repeated radiation exposure.</p> Trial registration <p>Not applicable. This study is a cross-sectional study evaluating a predictive machine learning model on healthy participants, not a clinical healthcare intervention trial.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Machine learning-based prediction of individualized musculoskeletal responses during shoulder exercises using surface electromyography: a cross-sectional, repeated-measures laboratory study

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月28日

查看全文
  • Donghyun Kim,
  • Chan-Su Lee,
  • Jonghoon Kim

摘要

Background

Understanding how specific exercises induce musculoskeletal responses is central to sports science and rehabilitation. While X-ray imaging effectively evaluates exercise-induced kinematic changes such as the path of the instantaneous center of rotation (PICR), it requires radiation exposure, limiting its clinical utility. This study investigated whether machine learning models using non-invasive sensor data could accurately predict X-ray-derived PICR changes for individualized exercise selection.

Methods

Forty-five healthy participants performed three shoulder internal rotation exercises: Belly Press, Lift Off, and Prone Wiper. During the exercises, muscle activity was recorded via surface electromyography (EMG) from four muscles, and muscle thickness was measured using ultrasound. Pre- and post-exercise X-ray imaging determined PICR displacement as the reference outcome. Linear regression and four nonlinear models (including Multilayer Perceptron [MLP] and Support Vector Regression) were evaluated using ninefold cross-validation to predict PICR changes and identify the optimal exercise for each subject.

Results

Nonlinear models consistently outperformed linear regression in estimating X-ray-derived pre-post changes across all sensing modalities, demonstrating lower mean absolute errors. In subject-specific optimal exercise selection, nonlinear models showed higher agreement with the X-ray reference. Notably, the EMG-only MLP model accurately identified the optimal exercise in 97.8% (44 of 45) of participants, performing equally to models utilizing combined EMG and ultrasound features.

Conclusions

Nonlinear machine learning models effectively capture subject-specific musculoskeletal responses beyond linear mappings. Relying solely on accessible EMG features, the proposed framework provides highly accurate, individualized exercise selection aligned with X-ray reference measures, offering a promising decision-support tool for personalized rehabilitation without repeated radiation exposure.

Trial registration

Not applicable. This study is a cross-sectional study evaluating a predictive machine learning model on healthy participants, not a clinical healthcare intervention trial.