Predicting behavioral escalation in students with autism using physiological and movement data

Behavioral escalation in students with Autism Spectrum Disorder (ASD) presents significant challenges in educational and community settings, often leading to reactive interventions after behaviors become observable. This study investigated whether physiological and movement data obtained from wearable biometric sensors could retrospectively identify precursors of escalation and inform proactive support strategies.

Physiological and movement data were collected from 30 students with ASD in naturalistic classroom environments using wearable sensors. Measures included heart rate variability (HRV), galvanic skin response (GSR), skin temperature, and accelerometer data. Behavioral observations were used to identify escalation events, and machine learning models, including Random Forest, Support Vector Machine (SVM), and Long Short-Term Memory (LSTM) networks, were trained on previously collected physiological and behavioral data to classify pre-escalation states. Model performance was evaluated retrospectively rather than through real-time deployment.

Among the models tested, the LSTM network achieved the highest classification accuracy (85.2%) in identifying pre-escalation states. Physiological changes associated with escalation were detectable approximately 2 to 5 minutes before observable behavioral manifestations. HRV and GSR emerged as the strongest predictive features, while inclusion of accelerometer data further improved classification performance. These findings suggest that measurable physiological patterns precede overt behavioral escalation in students with ASD.

The results support the feasibility of using wearable biometric data and machine learning approaches to retrospectively identify antecedent physiological markers of behavioral escalation. Interpreted through frameworks drawn from Applied Behavior Analysis (ABA) and Cognitive Behavioral Therapy (CBT), these physiological signals may function as both antecedents to challenging behavior and indicators of internal emotional or regulatory states. Although real-time implementation was not evaluated, the findings provide a foundation for future predictive monitoring systems designed to facilitate earlier intervention, reduce reliance on reactive or punitive responses, and promote more inclusive and emotionally responsive educational environments for neurodiverse learners.