Objective: To analyze videos of gait to: 1) identify spinocerebellar ataxia types 1 and 3, and 2) predict scores on the gait task section of the Scale for the Assessment and Rating of Ataxia (SARA).

Background: There is growing interest in developing automated platforms to identify and assess ataxia. Gait is a notable target for analysis, as its disruption is a common early sign of the disorder [1].

Method: Videos of participants completing the SARA gait task at annual visits across multiple clinical sites were collected. The videos were accompanied by corresponding task scores provided by raters. We used a convolutional neural network and applied a heuristic to isolate participants from their surroundings. Via recursive feature elimination with cross-validation, specific features of interest were extracted that reflected common characteristics of ataxic gait. A random forest classification model was built to identify at-risk participants (those with scores greater than 0 on the SARA gait task). A random forest regression model was built to predict scores on the SARA gait task. As the majority of provided scores were less than 3 in the dataset, we restricted model predictions to continuous values between 0 and 3.

Results: Across 11 clinical sites in 8 different US states, we gathered 155 unique videos from 65 participants with spinocerebellar ataxia types 1 or 3 and 24 controls [table1]. The random forest classification model achieved an accuracy of 83.1% [6.8%] and F1 score of 80.2% [9.2%] when using ten-fold cross validation to distinguish between at-risk individuals and controls. The random forest regression model achieved a mean absolute error of 0.62 [0.01] and Pearson’s correlation coefficient of 0.73 [0.01] when using ten-fold cross validation to predict scores on the SARA gait task. Per analysis of Shapley values, features capturing wider steps, step variation, slower walking speed, and imbalance most significantly influenced our models’ predictions; these are consistent with previously identified clinical features of ataxic gait [2, 3, 4].

Conclusion: Our results on data collected from multiple clinical sites demonstrate the robustness and generalizability of our models. Efforts to evaluate additional SARA tasks aside from gait are warranted and may be integrated into a holistic scoring pipeline. This platform could then be adapted for use in non-clinical settings.

References: [1] Ataxia – Symptoms and causes. Mayo Clinic. Published 2018. https://www.mayoclinic.org/diseases-conditions/ataxia/symptoms-causes/syc-20355652
[2] Buckley E, Mazzà C, McNeill A. A systematic review of the gait characteristics associated with Cerebellar Ataxia. Gait Posture. 2018 Feb;60:154-163. doi: 10.1016/j.gaitpost.2017.11.024. Epub 2017 Dec 1. PMID: 29220753.
[3] Schmitz-Hübsch T, du Montcel ST, Baliko L, Berciano J, Boesch S, Depondt C, Giunti P, Globas C, Infante J, Kang JS, Kremer B, Mariotti C, Melegh B, Pandolfo M, Rakowicz M, Ribai P, Rola R, Schöls L, Szymanski S, van de Warrenburg BP, Dürr A, Klockgether T, Fancellu R. Scale for the assessment and rating of ataxia: development of a new clinical scale. Neurology. 2006 Jun 13;66(11):1717-20. doi: 10.1212/01.wnl.0000219042.60538.92. Erratum in: Neurology. 2006 Jul 25;67(2):299. Fancellu, Roberto [added]. PMID: 16769946.
[4] Vyšata O, Ťupa O, Procházka A, Doležal R, Cejnar P, Bhorkar AM, Dostál O, Vališ M. Classification of Ataxic Gait. Sensors. 2021; 21(16):5576. https://doi.org/10.3390/s21165576

« Back to 2022 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/analyzing-gait-videos-to-identify-and-evaluate-spinocerebellar-ataxia-types-1-and-3/