Behavioral escalation in autism is typically managed reactively. Educators and caregivers respond to fidgeting, vocal outbursts, task refusal, aggression, self-injury, or attempts to leave the classroom only after these behaviors become visible, which the researchers argue is often too late for effective de-escalation. Traditional approaches rooted in Applied Behavior Analysis rely on functional behavior assessment and observation of antecedents, behaviors, and consequences, but these methods capture internal states poorly, particularly for students with limited verbal communication. The study’s premise is that the autonomic nervous system, which governs stress and emotional regulation through interactions among the hypothalamus, amygdala, and prefrontal cortex, produces measurable changes in heart rate variability, electrodermal activity, and peripheral temperature before distress becomes behaviorally obvious. If those changes can be detected reliably, they may serve as objective early-warning markers that current observation-based practice cannot provide.

The participant sample consisted of students aged 6 to 14, with a mean age of 10.2 years, recruited from an initial screening pool of 37 students. Four did not meet inclusion criteria and three declined or could not tolerate the monitoring procedures, leaving 30 enrolled participants who completed the four-week data collection period with parental consent and student assent. The group included 16 male and 14 female students, with diagnosis levels distributed across DSM-5 severity levels, and verbal ability ranging from one non-verbal student to 11 fluent speakers. Twenty-six students tolerated the wearable sensor for the full monitoring period, while four required periodic adjustments or brief removals that produced partial data loss. Exclusion criteria screened out cardiac conditions and motor control disorders that could confound physiological recordings.

Each student wore a BIOPAC Research Ring on one finger during school hours from 9:00 AM to 3:30 PM over four weeks. The device recorded electrodermal activity at 32 hertz as a proxy for sympathetic arousal, photoplethysmography-based heart rate variability at 64 hertz, skin temperature at 1 hertz, and tri-axial accelerometer data at 100 hertz. Trained research staff applied the ring each morning and removed it at the end of the school day. Simultaneously, trained observers logged behavioral incidents using a standardized rubric categorizing events by severity, from mild agitation and repetitive movements to verbal outbursts, aggression, self-injury, and elopement. Interrater reliability, assessed on 20 percent of observations coded independently by two raters, yielded a Cohen’s kappa of 0.87, indicating strong agreement. Event timestamps were synchronized with physiological streams to the nearest second using a common system clock.

Across the study, 200 behavioral events were documented. Mild agitation accounted for 24 percent and repetitive behaviors for 18 percent of incidents, while higher-severity events such as aggression, self-injury, and elopement attempts were less frequent but required immediate staff intervention. The researchers then processed the physiological data in 30-second overlapping windows with 50 percent overlap, extracting time-domain and frequency-domain heart rate variability features such as RMSSD, SDNN, low-frequency and high-frequency power, and the LF/HF ratio. Galvanic skin response signals were filtered with a zero-phase Butterworth low-pass filter and baseline correction, then skin conductance responses exceeding 0.01 to 0.05 microsiemens with 1-to-5-second rise times were detected and characterized by peak amplitude, rise time, and recovery time. Skin temperature was adjusted for ambient classroom conditions, and accelerometer data yielded activity index, jerk, and entropy measures. Data loss during processing ranged from 6.2 percent for heart rate variability to 9.2 percent for galvanic skin response.

Descriptive statistics revealed systematic physiological differences between pre-escalation and calm periods. Windows within five minutes before an escalation showed lower heart rate variability, with mean RMSSD of 24.8 milliseconds compared with 36.5 milliseconds during non-escalation periods, and an elevated LF/HF ratio of 2.84 versus 1.61, consistent with increased sympathetic activation. Galvanic skin response peak amplitude nearly doubled, averaging 0.22 microsiemens before escalations versus 0.11 microsiemens at baseline. Skin temperature showed modest reductions, from 31.3 degrees Celsius to 30.8 degrees Celsius, reflecting stress-related vasoconstriction, and accelerometer-derived movement variability rose from 0.91 g to 1.48 g. Together, these patterns describe a shift from parasympathetic-dominant regulation toward sympathetic arousal in the minutes preceding observable dysregulation.

The Long Short-Term Memory network, a recurrent architecture designed to capture long-range temporal dependencies in sequential data, achieved the strongest performance: 85.2 percent accuracy, 83.7 percent precision, 86.5 percent recall, an F1-score of 85.1, and an area under the receiver operating characteristic curve of 0.90, all computed on a held-out 30 percent test set with stratified sampling. The Random Forest classifier reached 81.4 percent accuracy and the Support Vector Machine with a radial basis function kernel 78.9 percent. The model detected pre-escalation patterns two to five minutes before observable behaviors, with classification accuracy declining by roughly 3 percent for each additional minute of lead time. In leave-one-subject-out cross-validation, in which the model was trained on all participants except one and tested on the excluded student, average accuracy remained approximately 82.6 percent, suggesting reasonable generalization to individuals not represented in training data.

Feature importance analysis identified heart rate variability as the strongest predictor, followed by galvanic skin response. Adding galvanic skin response features alone to models trained only on heart rate variability and accelerometer data increased recall by 12 percent, while integrating accelerometer features lifted the F1-score by 3.3 percentage points. Ablation testing showed that models using heart rate variability and galvanic skin response alone achieved an F1-score of 81.2 percent, rising to 85.1 percent with accelerometer data and 85.2 percent when skin temperature was included. The researchers also simulated an alert system on the retrospective data, estimating a false-positive rate of approximately 0.18 alerts per student-hour when confidence thresholds were tuned for recall, a figure they note must be validated under operational conditions to assess educator workload and alert fatigue.

The study’s theoretical contribution lies in bridging Cognitive Behavioral Therapy and Applied Behavior Analysis. The authors treat physiological signals as somatic markers of emotional arousal that function both as antecedents within the behavioral antecedent-behavior-consequence framework and as objective indicators of the internal states that cognitive-behavioral accounts describe as private events. This framing is particularly relevant for students who cannot verbally report distress, for whom traditional self-report-based assessment is inaccessible. The researchers also emphasize privacy safeguards: data were encrypted, de-identified before analysis, and alerts were designed to convey anonymized predictions rather than sensitive biometric details, with the explicit intent of supporting rather than replacing professional judgment and avoiding labeling or over-surveillance of students.

Significant limitations temper the findings. Because the system was evaluated retrospectively, the study cannot determine whether live alerts would improve educator response, reduce escalation frequency, or sustain predictive accuracy amid classroom variability, sensor artifacts, and real-time operational constraints. The sample was confined to structured elementary school classrooms, so generalization to home, clinical, or community settings remains untested, and the heterogeneity of autism, including differences in medication status, sensory sensitivities, and co-occurring conditions, may affect model stability across broader populations. The authors also caution that the design cannot establish causal relationships between physiological markers and escalation; the biomarkers should be interpreted as correlates and predictors of risk rather than causes of behavior. Prospective validation across multiple schools, geographic regions, and educator teams, together with individualized alert thresholds and standardized response protocols, will be necessary before such predictive monitoring can be considered an operational tool for proactive behavioral support.

Subject of Research: Using wearable physiological and movement sensors with machine learning to predict behavioral escalation in students with autism

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

Article References: Lamb, R., Owens, T., Choi, I., Kundu, S., Nookkarapu, S., Ray, P., & Kavner, A. (2026). Predicting behavioral escalation in students with autism using physiological and movement data. Frontiers in Psychiatry, 17, Article 1850879. https://doi.org/10.3389/fpsyt.2026.1850879

Image Credits: AI Generated

DOI: 10.3389/fpsyt.2026.1850879

Keywords: autism spectrum disorder, wearable sensors, heart rate variability, galvanic skin response, machine learning, LSTM networks, behavioral escalation, classroom monitoring, autonomic nervous system, applied behavior analysis, cognitive behavioral therapy, predictive modeling