Image by desidesidesi/Shutterstock

Gait assessment plays a critical role in the diagnosis and management of neurological disorders. Although AI-based models offer the potential to enhance the objectivity and accessibility of clinical gait analysis, most existing models are limited by their dependence on specific patient populations and sensor configurations, largely due to the scarcity of diverse clinical datasets. In a recent study, researchers from IBM Research, Cleveland Clinic, and University of Tsukuba addressed this limitation by developing a novel framework that integrates synthetic gait data generated by generative AI with physics-based musculoskeletal simulation. This approach supports the development of scalable and generalizable gait analysis models applicable across heterogeneous patient groups and sensor environments.

Tsukuba, Japan—Gait assessment is critical for diagnosing and monitoring neurological disorders, yet current clinical standards remain largely subjective and qualitative. Recent advances in AI have enabled more quantitative and accessible gait analysis using widely available sensors such as smartphone cameras. However, most existing AI models are designed for specific patient populations and sensor configurations, primarily due to the scarcity of diverse clinical datasets—a constraint often driven by privacy concerns. As a result, these models tend to underperform when applied to populations or settings not well represented in the training data, limiting their broader clinical applicability.

In a study published in Nature Communications, researchers from IBM Research, the Cleveland Clinic, and the University of Tsukuba propose a novel framework to overcome this limitation. Their approach involves generating synthetic gait data using generative AI trained on physics-based musculoskeletal simulations. These simulations incorporate a broad spectrum of musculoskeletal parameters—spanning age groups from children to older adults, and conditions from healthy to pathological—as well as varied sensor configurations. This synthetic diversity enables the development of gait analysis models that are more robust and generalizable across a wide range of patient populations and clinical environments.

The team validated their approach using a large-scale real-world dataset comprising over 12,000 gait recordings from more than 1,200 individuals, including patients with cerebral palsy, Parkinson's disease, and dementia. The evaluation demonstrated two key strengths of the proposed framework:

1. Zero-shot capability: Models trained exclusively on synthetic data achieved performance comparable to—or even exceeding—that of models trained on real-world data. These models accurately estimated clinically relevant gait parameters (e.g., gait speed, step length, step time) and even muscle activity from single-camera video recordings.
2. Data-efficient generalization: Pretraining on synthetic data consistently enhanced model performance across a range of clinical tasks—including disease detection, severity grading, treatment response assessment, and longitudinal prediction of disease progression—under varying disease conditions and sensor configurations. Remarkably, models pretrained on synthetic data and fine-tuned with only limited real-world data outperformed state-of-the-art deep learning models trained entirely on real data.

These capabilities are especially valuable for rare or underrepresented conditions, where access to large-scale clinical datasets is limited. This work highlights the potential of synthetic data-driven approaches to enable scalable, equitable, and generalizable clinical motion analysis.

###
This work was supported in part by the Japan Society for the Promotion of Science, KAKENHI [grant numbers 19H01084 and 21K12153].

Original Paper

Title of original paper:

Utility of synthetic musculoskeletal gaits for generalizable healthcare applications

Journal:

Nature Communications

DOI:

10.1038/s41467-025-61292-1

Correspondence

Professor ARAI Tetsuaki
Institute of Medicine, University of Tsukuba

Related Link

Institute of Medicine