Performance of Wearable Artificial Intelligence for Emotion Dysregulation in Autism: A Systematic Review (Preprint)

UNSTRUCTURED Emotion dysregulation (ED) is a core feature of autism spectrum disorder, spanning subtle affective changes to acute behavioral episodes. Wearable AI, referring to artificial intelligence applied to physiological data from on-body sensors, offers a promising approach for continuous, objective ED detection, yet no synthesis has evaluated its performance in this population. This systematic review examined the performance and characteristics of wearable AI systems for detecting emotion dysregulation in autistic individuals. Eight databases were searched; eligible studies applied AI algorithms to wearable sensor data in autistic populations to detect ED-related states. Two independent reviewers conducted study selection, data extraction, and risk of bias assessment using PROBAST-AI, with findings synthesized narratively. Eighteen studies (2012-2025; mean sample size = 22) were included. A three-tier temporal framework was inductively derived from the included literature to organize findings according to intervention priority: baseline affective monitoring (17%), escalating stress detection (44%), and acute dysregulation identification (39%). The Empatica E4 and support vector machines were the most common device and algorithm, respectively. Reported performance was highest for acute dysregulation (accuracy/AUC frequently > 90%), more variable for escalating stress (69-99%), and lowest for baseline states (75-81%). All 18 studies were rated at high or unclear risk of bias in the analysis domain. Although wearable AI shows promise for ED detection in autism, the evidence base remains small and methodologically limited, and further research is needed before clinical translation can be considered.