ABSTRACT In response to the challenges of emotional behavior warning research for children with autism, which relies on data from neurotypical populations and lacks specific multimodal signal resources, resulting in poor model generalization and difficulty in coping with the highly heterogeneous physiological and behavioral patterns of children with autism spectrum disorder (ASD), this study aims to build a personalized, edge‐deployable real‐time early warning system based on smart wearable devices to improve the precise identification of and timely intervention for emotional outbursts in children with ASD. The data acquisition solution uses an Empatica E4 and an Inertial Measurement Unit (IMU) to acquire multimodal physiological and behavioral signals simultaneously. High‐precision spatiotemporal alignment is achieved through hardware triggering and Network Time Protocol (NTP) fine‐tuning algorithms. Subsequently, an individual dynamic baseline model is established, combined with adaptive extraction of heterogeneous features driven by an attention mechanism, to identify the most sensitive 10–15‐dimensional features for each child. Finally, a lightweight Attention‐Long Short‐Term Memory (Attention‐LSTM) model is deployed at the edge on an ESP32‐S3 microcontroller, enabling localized real‐time inference. Experiments demonstrate that behavioral warnings for children aged 7–9 with autism achieve a 92.5% true‐positive rate and an average early‐warning time of 53.5 s. All models converge within 5 days. This research provides a high‐precision, low‐latency, and highly privacy‐protected solution for emotional‐behavior intervention in children with autism, with strong potential for clinical translation.
Mingcui Shen (Sun,) studied this question.