Аннотация
Early detection of autism spectrum disorder (ASD) remains a critical challenge in pediatric neurodevelopment. This study introduces a hybrid CNN–LSTM framework for automated ASD detection by analyzing spatiotemporal patterns in facial expressions, hand gestures, and body postures extracted from video recordings. Using the MediaPipe framework, 2D anatomical landmarks and dynamic motion energy metrics were computed to form a 1,639-dimensional feature vector per frame. The proposed model was benchmarked against three other architectures—CNN– BiLSTM, Conv1D–GRU, and CNN–Transformer—on a custom dataset of children’s behavioral recordings. Evaluation metrics demonstrated strong classification performance for the CNN– LSTM model (accuracy = 91.0%, AUC = 0.931), with statistical analysis confirming its robustness and stability. The findings suggest that the CNN–LSTM framework offers a reliable, interpretable, and scalable approach for real-time ASD screening.