Optimizing Video Analysis for Autism Behavioral Screening

Researchers investigated methods for the automated detection of autism-related self-stimulatory behaviors from video, aiming to develop scalable computational screening tools. The study focused on two key areas: identifying the most effective sequence-based neural network architecture and optimal temporal sampling rate, and characterizing data augmentation strategies for training on limited behavioral datasets. For the first objective, Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) models were trained on pose-derived features. Both architectures significantly surpassed previous convolutional neural network baselines, with GRU achieving a peak accuracy of 98.75% at a sampling interval of every 15 frames. Regarding data augmentation, the study found that horizontal flipping was highly effective, and upsampling was essential for complex behavioral video augmentation. A personalized machine learning approach also yielded consistent predictions, providing concrete guidance for practitioners in data-scarce clinical domains.