Comparative Analysis of Acted-Out and Real Patient Gait Parameters for Sensing Applications

Neurological conditions greatly affect patients’ mobility and can considerably shorten their lifespan. However, the shortage of neurologists can limit access to essential evaluations, particularly in underserved areas. On the other hand, alternative sensing methods (e.g., videos, WiFi, mmWave), based on machine learning, can provide cost-effective solutions, and have thus gained tractions in recent years. However, these systems need large/diverse datasets for development/evaluation, collection of which can be considerably challenging from real patients. This has then made recruiting actors to replicate gait disorders a common practice, raising questions about the similarity of their gait patterns to those of real patients and the generalizability of the subsequent findings. In this paper, we methodically investigate the extent to which actors can accurately replicate gait disorders, and the corresponding impact on gait disorder classification systems, marking the first study of its kind to the best of our knowledge. Specifically, we collect videos of real patients, largely based on collaboration with a neurology center, and further recruit actors to closely act them out. By statistically analyzing key gait parameters across the two datasets, we then reveal interesting underlying disparity patterns. We further develop gait disorder classifiers to illustrate the impact of this disparity on the classification performance. Our results show interesting trends, e.g., mild gait conditions are classified more accurately than what would have been with real patients, while moderate cases are classified less accurately. Overall, the paper provides a platform for interpreting the results when recruiting actors, thus paving the way towards equitable healthcare.