From simulated steps to real-world care: AI learns how we walk for neurology

The way we walk — our gait — can reveal much more than we might expect. Subtle characteristics in stride, left-right symmetry, and step timing can be early signs of neurological conditions including Parkinson’s disease, cerebral palsy, cognitive decline, and dementia. But traditional clinical gait assessments rely heavily on visual observation, an approach limited by subjectivity, time, and environment. That’s beginning to change.

In a new collaboration between IBM Research and Cleveland Clinic, a team of scientists and clinicians has developed a new approach to gait analysis that works across a wide range of neurological conditions and sensor configurations — including common devices like smartphone cameras and wearables. The core of the work is a new AI foundation model, informally called GaitFM, which learns the patterns of human movement from a rich, diverse dataset; most of it synthetic. The model is designed for the real-world, where patient data is scarce, and conditions are far from ideal, whether inside or outside the clinic.

“There is tremendous potential for AI-based gait analysis to improve the diagnosis of neurological conditions and track treatment response,” said Dr. James Liao, a neurologist at Cleveland Clinic and co-lead of the project. “Our goal is to bring this out of idealized lab conditions and into everyday clinical practice, helping doctors and patients by providing consistent, objective assessments.” This work is part of the Discovery Accelerator — a joint center that utilizes Cleveland Clinic’s research and clinical expertise, and IBM’s global leadership in computing technologies.

Why gait matters and how synthetic data enables real-world assessment

Gait isn’t just about how fast someone walks. It’s a complex blend of step length, joint coordination, foot placement, and posture. These nuances can offer powerful clinical signals, but many are hard to detect, even for trained professionals. Existing AI models often struggle to generalize across disparate patient populations, diseases, monitoring conditions, and sensing technologies. Even something as simple as a change in camera angle can significantly degrade performance.

“Most existing gait models are trained on small, narrowly defined datasets focused on a single disease, age group, or sensor type,” explained Yasunori Yamada, senior research scientist at IBM Research and lead author of the study. “That makes them hard to scale or apply broadly in clinical practice.”

That’s where synthetic data enters the picture. The researchers harnessed a generative AI model with physics-based musculoskeletal simulations of walking. This data isn’t fake, it’s grounded in biomechanics and designed to emulate the effects of age, neurological disorders, and sensor variability. The team generated thousands of high-fidelity variations in gait, covering a wide spectrum of body types, disease conditions, wearable sensors, and camera viewpoints. These synthetic gaits serve as diverse, scalable, and tunable data, helping AI models learn how to generalize.

Synthetic along with real data is better than either alone

The AI model was pre-trained on synthetic gait sequences, then fine-tuned with additional real clinical data from patients with Parkinson’s disease, cerebral palsy, and dementia. As recently described in Nature Communications, this hybrid approach delivered state-of-the-art or better performance in gait parameter estimation as well as clinical classification and prediction, while using just a fraction of the real-world data typically required.

Two highlights stood out. The first is data-efficient generalization: When pre-trained on synthetic data, the model required only a fraction of the real-world data to match or exceed state-of-the-art current benchmarks. Even without access to clinical data, models trained solely on synthetic gaits were able to estimate clinically relevant movement features across neurological conditions, sometimes outperforming real-data-trained models.

The second is multimodal applicability: Synthetic data was generated for multiple sensor modalities, including video, accelerometer, and electromyography (EMG). This allows deployment across a wide range of clinical and at-home settings. Notably, a model trained solely on synthetic data was able to estimate lower-limb muscle activity using only a single video feed. This flexibility opens the door to detailed motion analysis without specialized equipment.

“This is how digital health becomes real,” said Jeff Rogers, IBM’s global research leader for digital health. “By combining clinical insight with scalable AI tools, we’re moving past the proof-of-concept phase and into actual clinical utility.”

This synthetic-to-real approach holds particular promise for rare or underrepresented conditions, where data scarcity has historically held back AI innovation. By simulating gaits for these populations and fine-tuning with small clinical datasets, the model supports the development of accurate, equitable, and scalable diagnostic tools.

A foundation for gait-based healthcare

The culmination of this work is GaitFM, the team’s foundation model for gait analysis that can interpret videos from a single camera, with no need for lab-grade sensors or motion-capture rigs required. GaitFM was recently honored with the Best Paper Award at the 2025 IEEE International Conference on Digital Health held in July in Helsinki, and it is already being piloted in Cleveland Clinic deployments.

“Once we prove that models trained on synthetic data are clinically reliable, we can begin to scale up real-world use; recording more gaits, improving the model, and building a virtuous cycle,” said Liao. “But synthetic data will always remain essential to achieve the flexibility we need to assess patients inside and outside the clinic, along with covering edge cases, rare diseases, and new devices.”

From simulated steps to real patient care, this collaboration represents a meaningful milestone in how AI can learn from movement and help patients maintain their health.