From Observation to AI-Powered Gait Analysis

The way a person walks can reveal as much about brain health as a cognitive test. Stride length, symmetry, coordination, and rhythm all encode neurological signals. Subtle irregularities may be early markers of Parkinson’s disease, cerebral palsy, dementia, or cognitive decline.

Traditionally, gait evaluation has relied on visual observation – a process that is subjective, limited to controlled lab settings, and often misses early warning signs.

Now, a collaboration between IBM Research and Cleveland Clinic is changing this landscape. Their AI foundation model, GaitFM, can interpret gait across multiple conditions and sensors—from smartphone cameras to wearable accelerometers. The breakthrough lies in how it was trained: with synthetic data.

How Synthetic Data Trains AI for Gait Analysis

Most AI systems in neurology face a common limitation: small, narrow datasets. For example, training might include only a few dozen patients of one age group, one disease type, or one recording setup. When applied in a real-world setting—say, switching from a lab camera to a smartphone—the accuracy often drops.

To solve this, researchers used physics-based musculoskeletal simulations of walking. These simulations generated thousands of synthetic gait patterns that mimic real variations in:

  • Age and body type
  • Neurological conditions such as Parkinson’s or cerebral palsy
  • Sensor variability (camera angle, wearable devices, home vs. lab)

Unlike artificial “fake” data, these models are rooted in biomechanics, creating realistic variations in stride and movement. Pre-training on these simulations gave GaitFM a broad understanding of human gait before exposure to clinical data.

Real + Synthetic Data: A Winning Combination

When GaitFM was fine-tuned with patient datasets from Parkinson’s, cerebral palsy, and dementia cases, its accuracy matched or exceeded state-of-the-art benchmarks—even when real data was limited.

Key findings included:

  1. Data Efficiency – Pre-training with simulated gaits reduced dependence on scarce real-world datasets.
  2. Multimodal Adaptability – The model worked across video, accelerometer, and electromyography (EMG) inputs. One version even estimated muscle activity from a single video feed.

This approach means AI can remain robust in different environments—whether in a clinic, a lab, or a patient’s home.

Clinical Applications Beyond the Lab

For neurologists and rehabilitation specialists, the implications are significant:

  • Early Diagnosis – Detect subtle gait abnormalities before they become disabling.
  • Disease Tracking – Monitor progression of conditions like Parkinson’s or dementia objectively.
  • Treatment Measurement – Evaluate whether therapies or medications are improving mobility.
  • Remote Monitoring – Run on simple smartphone videos, making gait analysis accessible even in resource-limited regions.

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

Already, the model has earned recognition with the Best Paper Award at the 2025 IEEE International Conference on Digital Health and is being piloted at Cleveland Clinic.

Why Patients and Caregivers Should Care

While this research is highly technical, its impact is very human.

  • For patients: Easier access to neurological assessments without needing specialized labs.
  • For caregivers: Early detection tools that may help intervene before decline accelerates.
  • For healthcare systems: Scalable, cost-effective solutions for large populations.

As project co-lead Dr. Liao explained: “Synthetic data will always remain essential—covering rare cases and enabling flexible assessment both inside and outside the clinic.”

Key Takeaways

  • AI can analyze walking patterns (gait) to detect neurological conditions.
  • Synthetic, physics-based simulations allow the AI to learn before clinical deployment.
  • Combining real and synthetic data boosts accuracy and efficiency.
  • Technology opens the door to remote, video-based monitoring, potentially making gait analysis as simple as recording a clip on a smartphone.

Must Read:

AI in Healthcare: Can ChatGPT-4 Really Compete with Neurologists?

&

How AI and Digital Biomarkers Are Transforming Stroke Cognitive Monitoring

FAQs

1. What is gait analysis in neurology?
Gait analysis is the study of how a person walks. Neurologists use it to detect movement patterns linked to brain or nerve disorders such as Parkinson’s or dementia.

2. How does AI improve gait analysis?
AI models can objectively measure stride, balance, and coordination—detecting changes too subtle for the human eye and tracking them over time.

3. Why is synthetic data important in healthcare AI?
Synthetic data fills gaps where real-world datasets are too small or inconsistent, helping AI models generalize better across patients and environments.

4. Can gait AI be used at home?
Yes. With tools like GaitFM, gait assessment can run on smartphone videos, making it more accessible for patients in remote areas.

5. Which conditions can be detected with gait AI?
Parkinson’s disease, cerebral palsy, dementia, and cognitive decline are among the primary conditions where gait abnormalities provide early clues.

References

  1. Lord S, Howe T, Greenland J, Simpson L, Rochester L. Gait variability in older adults: a structured review of testing protocol and clinimetric properties. Gait Posture. 2011;34(4):443–50.
  2. Yamada Y, Liao J, Rogers J, et al. Foundation models for gait analysis using synthetic and clinical data. Nat Commun. 2025;16:2154.
  3. Seth A, Hicks JL, Uchida TK, Habib A, Dembia CL, Dunne JJ, et al. OpenSim: Simulating musculoskeletal dynamics and neuromuscular control to study human and animal movement. PLoS Comput Biol. 2018;14(7):e1006223.
  4. Qian Z, Biancardi B, Yamada Y, et al. Synthetic data generation for biomechanics and clinical AI applications. IEEE Trans Neural Syst Rehabil Eng. 2024;32(5):978–88.
  5. Zhou J, Chen Y, et al. Hybrid AI models combining synthetic and real gait data for neurological monitoring. Sci Rep. 2025;15:14062.
  6. Mirelman A, Bernad-Elazari H, Thaler A, Giladi-Yacobi E, Gurevich T, Gana-Weisz M, et al. Arm swing as a potential new prodromal marker of Parkinson’s disease. Mov Disord. 2016;31(10):1527–34.