Transforming Musculoskeletal Disorder Management With Multimodal AI Platform for Early Diagnosis, Self-Screening, and Personalized Care
Med Research
Cite this paper
Plain citation
Pengyuan Xu, Guang Yang, Ge Zhang, Yafang Zhao, Pu Wan, Guiyun Zhou, Mengjie Jiang, Yixin Zhang, Zhaoyang Wang, Yichen Dai, Xiqing Zhao, Ziliang Fu, Ziyu Zheng, Ziyun Xu, Ruisi Dong, Xiaojie Xia, Duo Zhang, Yisheng He, Michael Moran, Qiguang Wang, Tim Woodfield, Khoon Lim, Tiantian Kong, Junnan Tang, Kang Tian, Kai Zhou, Hao Chi, Pengpeng Zhang, Jianhua Yang, Mingliang Zhang, Moyan Li, Yuan Cheng, Xiaolin Cui. Transforming Musculoskeletal Disorder Management With Multimodal AI Platform for Early Diagnosis, Self-Screening, and Personalized Care. Med Research. 2026. https://doi.org/10.1002/mdr2.70073
Research context and source limits
Limited evidence note. Only the source material identified below has been reviewed; this is not a complete Paper GEO V2 assessment.
Reported multimodal design; diagnostic performance and clinical utility remain unverified.
Reported data and clinical task
The paper describes an AI system for osteoarthritis, rheumatoid arthritis and osteoporosis using routine laboratory measurements, demographics and symptom narratives. It reports 58,777 NHANES records, split into 49,523 training records from 1999–2016 and 9,254 temporal-test records from 2017–2018, plus 4,436 records from four Chinese centres. These are reported dataset roles, not independently verified evidence of clinical utility.
Source: Results 2.1–2.2; Figure 3 caption
A specific, interpretable architecture
The described fusion combines three structured-model scores with a 768-dimensional Bio_ClinicalBERT representation, producing 771 features for a classifier. The text encoder uses LoRA adapters; explicit diagnostic fields and disease names are described as removed or masked. This is a supervised prediction architecture with workflow orchestration, not evidence that a generative model independently reasons through diagnosis or improves patient outcomes.
Source: Results 2.4 Multimodal Data Interconnection; Figure 5
Performance claims remain limited
All three printed training AUC estimates lie outside their accompanying confidence intervals. The 42-versus-41 predictor count and the relationship between test-set-based hyperparameter selection and later one-time evaluation also require clarification. The audit consequently does not endorse diagnostic accuracy, the reported multimodal improvement, or prospective self-screening effectiveness.
Source: Results 2.2, Figure 2 caption and Figure 3C caption; Results 2.4
Evidence limits
- Training AUC/CI pairs are internally inconsistent.
- Predictor count and model-selection/evaluation separation are unresolved.
- Reported architecture does not demonstrate prospective clinical benefit.
Provenance
- Article source — Official publisher or PMC full text reviewed locally. Results 2.1–2.2; Figure 3 caption
Source review: 2026-10-27.
Verify publication identity
These records identify the same original work; database inclusion is not evidence of scientific validity or clinical benefit.