A new study gives clinicians a practical blueprint for automating video-based detection of autism-related self-stimulatory behaviors.
Researchers trained two types of recurrent neural networks — long short-term memory (LSTM) and gated recurrent unit (GRU) models — on pose-derived features from the Self-Stimulatory Behavior Diagnosis dataset. Both outperformed prior convolutional neural network baselines, which topped out at 62-76% accuracy. The sweet spot was sampling one frame every 15 frames: LSTM hit 97.5% accuracy there, while GRU reached 98.75%. The team also ran ten data augmentation strategies through an ablation study on a separate transfer-learning pipeline and found that horizontal flipping helped most, while dropping upsampling hurt most.
Autism spectrum disorder affects over 75 million people worldwide, yet remote behavioral screening tools that scale remain scarce. Concrete guidance on which architecture to pick, how densely to sample video, and which augmentation tricks are worth the trouble is exactly what separates a research result from something a clinical software team can actually ship.
Personalized models trained and tested per subject produced consistent predictions, which is encouraging — though a mean loss of 1.84 on a small dataset is not the same thing as a ready-to-deploy screening tool.