Autoencoder-derived Single Summary Metric to Assess Gait Quality

• Published in: Archives of physical medicine and rehabilitation Vol. 102; no. 10; p. e95
• Main Authors: Tabashum, Thasina, Wang, Shou-Jen, Kruger, Karen, Krzak, Joseph, Albert, Mark V.
• Format: Journal Article
• Language: English
• Published: Elsevier Inc 01.10.2021
• Subjects: Deep Learning; Gait Analysis; Movement Disorders; Gait Analysis; Movement Disorders; Deep Learning
• ISSN: 0003-9993; 1532-821X
• DOI: 10.1016/j.apmr.2021.07.758

To improve holistic interpretation of quantitative gait analysis data through a deep learning autoencoder with a single-valued reduced representation. The tailored single value is another strategy to quantify overall gait quality in individuals with movement disorders. Such overall summary metrics provide a holistic assessment of gait quality for assessing outcomes of therapeutic interventions. Observational Study. The data was collected at the Motion Analysis Centers (MACs) at the Shriners Hospitals for Children - Chicago. 412 subjects with cerebral palsy under the age of 18 and 86 normal subjects (498 total) with collected gait data from the MAC at Shriners Hospitals for Children - Chicago over 20 years. Not applicable. A single summary metric generated by an autoencoder model quantifies the quality of an individual's gait. This metric can be used as a measure of improvement of gait after surgery to improve mobility in conditions such as Cerebral Palsy. The model is trained using high-quality data at the Shriners MAC in Chicago from 2004 to 2020, which includes temporo-spatial parameters (7), lower extremity kinematic (64), and lower extremity kinetic (43) data - a total of 114 features. The learned single metric captures more than 86% variance, which was established using subject-wise cross-validation. This preliminary model can produce an overall summary value for the gait of individuals by incorporating the temporo-spatial parameters, lower extremity kinematics, and lower extremity kinetic data. Such summary metrics can be used to assess the impact of therapeutic interventions. None.