Semi-Automated Multi-Label Classification of Autistic Mannerisms by Machine Learning on Post Hoc Skeletal Tracking

• Published in: Autism research Vol. 18; no. 4; p. 833
• Main Authors: Lemler, Christian, Kleber, Solvejg K, Polzer, Leonie, Raji, Naisan, Kitzerow-Cleven, Janina, Kim, Ziyon, Platte, Simeon, Freitag, Christine M, Bast, Nico
• Format: Journal Article
• Language: English
• Published: United States 01.04.2025
• Subjects: Algorithms; Autism Spectrum Disorder - classification; Autism Spectrum Disorder - diagnosis; Autism Spectrum Disorder - physiopathology; Child; Child, Preschool; Female; Humans; Machine Learning; Male; Movement - physiology; Neural Networks, Computer; Stereotyped Behavior - physiology; mannerisms; multi‐label classification; autism spectrum disorder; machine learning; stereotypic behavior
• ISSN: 1939-3806
• DOI: 10.1002/aur.70020

Mannerisms describe repetitive or unconventional body movements like arm flapping. These movements are early markers of restricted and repetitive behaviors (RRBs) in autism spectrum disorder (ASD). However, assessing mannerisms reliably is challenging. Even after extensive training in behavioral observations, inter-rater agreements for mannerism items remain insufficient. The current study used machine learning (ML) to classify mannerisms from videotaped behavioral observations in children with ASD. We developed a classification scheme for mannerisms as ground truth and applied it to videotaped behavioral observations from an early intervention study. ML was used in two steps: First, the OpenPose algorithm post hoc extracted features based on body movements in the videos. Second, a long short-term memory (LSTM) neural network classified the features in a multi-label approach to distinguish between the absence of mannerisms, flapping, jumping, and both flapping + jumping. The trained models achieved 70.2% accuracy (F1 score: 31.8%) using nested cross-validation. The analysis improves on previous videotaped ML classification studies by splitting training and test data subject-wise, highlighting its clinical applicability. The LSTM models are made publicly available for use with other video datasets. Our results show that ML-based classification of mannerisms is a promising tool for enhancing objective diagnostic methods of behavioral observations.