Toward a Foundation Model for Healthcare Applications on Single-Camera Gait Analysis

Gait assessment offers valuable insights into disease diagnosis and prognosis in various neurological diseases. Recent advances in deep learning-based pose estimation and time-series analysis have enabled quantitative gait analysis using single-camera videos. However, the scarcity of datasets involving clinical populations due to privacy concerns has hindered the potential of these models and their adoption in healthcare applications. To address this challenge, we introduce GaitFM, a foundation model for single-camera-based gait analysis, pre-trained on simulation-based synthetic gaits to learn both the estimation of clinically relevant gait features and feature representation for a wide range of human gaits including pathological gaits. We evaluated GaitFM by adapting it to multiple types of gait-based downstream tasks related to dementia, including those involving clinical diagnostic status, cognitive impairments in distinct domains, neuropathological changes, and future cognitive decline. Our results demonstrate that GaitFM consistently outperforms baseline models using previous state-of-the-art methods across all downstream tasks. Additionally, GaitFM facilitates data-efficient model adaptation, performing comparably to the best comparison model while using significantly fewer real data samples. These findings demonstrate the potential of GaitFM to facilitate the widespread adoption of gait analysis in healthcare.