Feature selection from markerless movement recordings to assess dystonia in children with cerebral palsy K.L. Stolk a, b , M. Schwartz b , M.M. van der Krogt a , S.S. van de Ven a , L.A. Bonouvri´ e a , J. Harlaar a, c , A.I. Buizer a , H. Haberfehlner a a Amsterdam UMC, Rehabilitation Medicine- Amsterdam Movement Sciences, Amsterdam, Netherlands b Gillete Children’s Specialty Healthcare, Center for Gait and Motion Analysis, Saint Paul, USA c TU Delft, Biomechanical Engineering, Delft, Netherlands 1. Introduction Assessment of dystonia in children with cerebral palsy (CP) is commonly performed by observation of movements in a clinical setting [1], which is subjective and time-consuming. Therefore objective and home-based measurements would be preferable. Movement-related features extracted from markerless motion tracking from regular videos may be used in a predicting algorithm. As a frst step towards such a machine learning algorithm to assess dystonia, a subset of rele- vant features needs to be identifed. 2. Research question Which features, extracted from markerless motion tracking from 2D videos, are relevant candidates to construct a machine learning model to assess dystonia in CP? 3. Methods 34 Children with dyskinetic CP (8–22 years old, GMFCS level IV and V) were included in the study [2]. Video recordings of twenty seconds were collected at one to three different time intervals in two resting positions, resulting in 94 videos of lying and 94 videos of sitting. All videos were clinically scored using the Dyskinesia Impairment Scale (DIS) [3]. As sitting and lying result in one total rest dystonia DIS score, 94 scores were calculated, which were analyzed as 94 different samples. The positions of twelve bony landmarks were tracked from the videos using DeepLabCut [4]. The position data was scaled to the upper arm length. From these time series, marker velocity and acceleration, knee and elbow angles, and angular velocities were obtained, and discrete outcome measures were extracted. All 1444 features were converted to Z-scores [5]. An overview of the features, categorized into position, movement, variability, and repeatability is presented in Table 1. A minimal redundancy maximum relevance algorithm [6] was used to rank candidate features for estimating the total rest dystonia score of the DIS [3], leaving out the scores for the eyes, mouth and neck. The top-50 ranked features were selected, as about two samples per feature have been described as suffcient (94/2≈50) [7], and the percentage of features in each category was calculated. Table 1 The division of the features into fve categories and the number of features per category. This table shows the fve categories, the features that are assigned to each category, the number of features and the percentage in different subsets of the Features. Category name Units Feature Total number of feature Number in top-50 Nr %tot %cat Nr %tot %cat Description of Position Pos/angle Mean position 316 21.9 100 16 32 5.06 Pos/angle Range Pos/angle Interquartile range Pos/angle Number of peaks Pos/angle Average height of peaks Stability of hips/shoulders, visibility of hips Description of movement Vel/acc/angvel Mean 552 38.2 100 11 22 1.99 Vel/acc/angvel Range Vel/acc/angvel Interquartile range Vel/acc/angvel Number of peaks Vel/acc/angvel Average height of peaks (continued on next page) Contents lists available at ScienceDirect Gait & Posture journal homepage: www.elsevier.com/locate/gaitpost https://doi.org/10.1016/j.gaitpost.2020.08.075 0966-6362/© 2020 Published by Elsevier Ltd. Gait & Posture 81 (2020) 354–355