[home] [Personal Program] [Help]

---

Session: Poster session 2 (Odd numbers)
Session starts: Friday 27 January, 10:00
Presentation starts: 10:00

---

Wei Tang (University Medical Center Groningen)
Peter van Ooijen (University Medical Center Groningen)
Deborah Sival (University Medical Center Groningen)
Natasha Maurits (University Medical Center Groningen)

Abstract:
Early Onset Ataxia (EOA) and Developmental Coordination Disorder (DCD) exhibit significant phenotypic similarities that may make differentiation challenging. Gait in ataxia can be clinically assessed semi-quantitatively using the Scale for the Assessment and Rating of Ataxia (SARA). However, such clinical evaluation is subjective and requires significant clinical competence. Wearable inertial sensors might provide an objective and feasible alternative. This approach, however, needs a special setup before usage which takes time and is not very patient-friendly. Because of the widespread use of video cameras in outpatient clinics and recent advances in pose estimation, we proposed a 2D skeleton-based machine learning technique for the classification of EOA, DCD, and healthy controls from freehand single camera recordings. Eighty-five children (31 EOA, 20 DCD, and 34 controls) were asked to walk in a straight line at their own pace, towards and away from a single 2D camera. Recordings were taken at 1280 × 720 resolution and 30 frames per second. Seventeen skeleton keypoints were extracted using the AlphaPose model pretrained on the MSCOCO dataset. The PoseFlow framework was then utilized to match the skeleton to the same subject in a recording. We employed a normalization approach using the average Euclidean distance from left shoulder to right hip and from right shoulder to left hip (ASH) to derive distance-based features. To improve the classification accuracy, we employed the overlapping sliding window (OSW) method to decompose the skeleton time series into windows of 45 frames with 30 frames overlap. We used 20 combined features as obtained from principal component analysis on statistical characteristics (mean, std, cov, skewness…) of ankle, knee, hip, elbow, shoulder and wrist distances. Classification results were validated by the XGBoost model and leave-one-patient-out cross-validation. The XGBoost classifier obtained a 67% f1 score on the original data before ASH normalization and OSW data augmentation. After ASH normalization, the f1 score improved to 70%. After ASH normalization and OSW data augmentation, the classifier finally achieved an f1-score of 78% for our dataset, with 63% for EOA, 76% for DCD, and 88% for healthy control children at group-level. Our approach demonstrates a promising application for the classification of movement disorders in a practical and objective way. Additionally, the ASH normalization method and the OSW data augmentation improved the classification result. We will incorporate other tasks, assessing upper body movement, in the future to develop further improved algorithms for clinical diagnostic support.