Scholarlink

Purpose <p>Subject-specific musculoskeletal models are valuable for studying postoperative complications. These models require CT/MRI to define muscle geometry and force-generating parameters; consequently, their use in routine clinical settings remains limited. Current practice; i.e. anthropometric scaling, introduces substantial error. This study aimed to determine if machine learning could predict subject-specific muscle parameters from readily available data, thereby improving spinal load prediction.</p> Methods <p>Using 250 CT-derived musculoskeletal models, we trained machine learning algorithms (Random Forests, Autoencoders, and Multilayer Perceptrons) to predict muscle geometry and maximum-isometric-force. Inputs were restricted to demographics and 3D vertebral centroids. We integrated these predicted parameters into OpenSim, simulating 11 static postures. The resulting spinal joint forces were compared against CT-derived subject-specific musculoskeletal models (Ground Truth) and anthropometric scaling.</p> Results <p>A hybrid Random Forest–Autoencoder approach performed best. Relative to anthropometric scaling, it lowered maximum-isometric-force RMSE (27 vs. 34&#xa0;N) and geometry RMSE (7.6 vs. 8.8&#xa0;mm). Spinal load predictions were also more accurate: R² between predicted and Ground-Truth spinal forces rose from 0.83 to 0.9 for compressive forces (RMSE 184 to 145&#xa0;N), and shear-force R² increased from 0.67 to 0.8 (RMSE 238 to 178&#xa0;N). Overall, global load-estimation error improved by ~ 18% for both modalities.</p> Conclusion <p>While not a perfect substitute for volumetric imaging, the proposed machine learning workflow significantly reduces the estimation errors inherent in standard anthropometric scaling. This approach offers a superior approximation for generating subject-specific musculoskeletal models in clinical settings where CT or MRI data are unavailable. </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 outperforms anthropometric scaling in predicting muscle parameters and spinal loading: a subject-specific musculoskeletal modeling study

  • 期刊论文

  • 发表日期：

    2026年4月12日

查看全文
  • Nima Ashjaee,
  • John Street,
  • Sidney Fels,
  • Thomas Oxland

摘要

Purpose

Subject-specific musculoskeletal models are valuable for studying postoperative complications. These models require CT/MRI to define muscle geometry and force-generating parameters; consequently, their use in routine clinical settings remains limited. Current practice; i.e. anthropometric scaling, introduces substantial error. This study aimed to determine if machine learning could predict subject-specific muscle parameters from readily available data, thereby improving spinal load prediction.

Methods

Using 250 CT-derived musculoskeletal models, we trained machine learning algorithms (Random Forests, Autoencoders, and Multilayer Perceptrons) to predict muscle geometry and maximum-isometric-force. Inputs were restricted to demographics and 3D vertebral centroids. We integrated these predicted parameters into OpenSim, simulating 11 static postures. The resulting spinal joint forces were compared against CT-derived subject-specific musculoskeletal models (Ground Truth) and anthropometric scaling.

Results

A hybrid Random Forest–Autoencoder approach performed best. Relative to anthropometric scaling, it lowered maximum-isometric-force RMSE (27 vs. 34 N) and geometry RMSE (7.6 vs. 8.8 mm). Spinal load predictions were also more accurate: R² between predicted and Ground-Truth spinal forces rose from 0.83 to 0.9 for compressive forces (RMSE 184 to 145 N), and shear-force R² increased from 0.67 to 0.8 (RMSE 238 to 178 N). Overall, global load-estimation error improved by ~ 18% for both modalities.

Conclusion

While not a perfect substitute for volumetric imaging, the proposed machine learning workflow significantly reduces the estimation errors inherent in standard anthropometric scaling. This approach offers a superior approximation for generating subject-specific musculoskeletal models in clinical settings where CT or MRI data are unavailable.