Wearable Biosensors Predict Aggressive Behavior in Youths

The Physiological Logic

BOSTON, MASSACHUSETTS — in a study published in JAMA Network Open in December 2023, researchers led by Tales Imbiriba and Ahmet Demirkaya at Northeastern University showed that machine learning applied to wearable biosensor data could predict imminent aggressive behavior in 70 psychiatric inpatient youths with autism roughly three minutes before it occurred. The best-performing classifier was logistic regression, with a mean area under the receiver operating characteristic curve of 0.80. The work was supervised by Matthew Goodwin, a Northeastern professor whose lab has been pursuing this question across more than a decade of studies and several earlier smaller samples.

The physiological premise is straightforward. Before a behavioral escalation, the sympathetic branch of the autonomic nervous system typically activates: heart rate rises, skin conductance increases, and movement patterns shift. Those signals are measurable on the wrist. Machine learning models trained on each child’s historical data can learn the individual signature of approaching distress, which is the part that matters in a clinical population where minimally verbal participants and intellectual disability are common. In the JAMA cohort, 32 of the 70 participants were minimally verbal and 30 had a documented intellectual disability.

AUROC of 0.80 is not the same as 80 percent accuracy. It means the classifier ranks a randomly chosen pre-aggression window above a randomly chosen non-aggression window 80 percent of the time. That is meaningful in research, and it is not yet a clinical alarm. The earlier 2019 prototype work from the same lab, with 20 participants, was sometimes reported in press coverage as 84 percent accuracy at one-minute lead time. The 2023 replication is the larger and more carefully reported result, and the headline numbers from the two studies are not interchangeable.

“Three minutes is enough time to do something.” — Matthew Goodwin, Bouvé College of Health Sciences, Northeastern University, in Northeastern Global News (January 2024)

The hardware behind almost all of this work has been the Empatica E4 wristband. Empatica is an MIT Media Lab spinoff with corporate offices in Cambridge, Massachusetts and a research and engineering arm in Milan, Italy. The E4 captures blood volume pulse, electrodermal activity, skin temperature, and three-axis acceleration, and it has been validated against clinical electrocardiography for heart rate variability metrics under non-movement conditions in a 2020 study by Schuurmans and colleagues in the Journal of Medical Systems.

The E4 itself, however, has been retired. Empatica sunset the device in 2024 and now sells EmbracePlus, a successor wearable that received FDA 510(k) clearance for continuous physiological monitoring in November 2022 and powers Empatica’s FDA-cleared EpiMonitor seizure-detection system, introduced in February 2024. New ABA research budgets buying hardware in 2026 are buying EmbracePlus, not the E4. The published literature, including most of the autism behavior-prediction work cited in this article, was built on the older device.

KeepCalm and the School-Based Pilots

The clearest example of biosensor data moving from prediction into intervention is the KeepCalm application, developed by Heather Nuske and colleagues at the Penn Center for Mental Health at the University of Pennsylvania. KeepCalm pairs with consumer-grade heart rate trackers, the MioFuse wristband or the Polar H7 chest strap, and not the Empatica E4. The app monitors a child’s heart rate, detects elevated physiological stress, and alerts the educational team in real time with individualized strategy suggestions drawn from the team’s past data on that child.

The framing matters. KeepCalm is designed for educational teams, preschool and elementary teachers and aides, supporting students on the autism spectrum in school settings. It is not currently a clinic product, and it is not a tool an ABA practice would deploy in a center-based program out of the box. A 2025 Journal of Pediatric Psychology paper by Isha Kaur and colleagues described the co-design process across five testing cycles with 73 participants and rated the app highly on acceptability, feasibility, and usability. A separate 2024 paper in MDPI Sensors examined 226 strategy interventions and found that heart rate reduction predicted strategy effectiveness, which is the closer-to-validation result.

What KeepCalm illustrates is the shape of a near-term ABA-adjacent product. A wearable provides a real-time stress signal. A coach-facing app translates that signal into a suggested strategy from a curated library. The interventionist, whether teacher, RBT (registered behavior technician), or BCBA, retains the clinical decision. The wearable is not making the call; it is flagging that a call may be needed sooner than human observation alone would have noticed.

Measuring the Therapist’s Stress

The 2026 frontier in ABA is not just about the child’s physiology. In a January 2026 Journal of Applied Behavior Analysis technical report, Emily Sullivan, Tara Fahmie, and Jamie Gehringer at the University of Nebraska Medical Center’s Munroe-Meyer Institute used the Empatica E4 to measure electrodermal activity in three therapists implementing functional analyses of challenging behavior. Functional analyses by design expose therapists to aggressive, self-injurious, and disruptive behaviors as part of identifying maintaining contingencies, and the work is among the most physiologically demanding clinical activity in ABA.

The Sullivan paper is preliminary. It is a technical report on three therapists, not a controlled trial. What it documents is that acute physiological arousal indicators are present in the data and trackable around the moments challenging behavior occurs. The implication, if the work scales, is that the same biometric platforms used to predict client escalation could also be used to monitor therapist physiological load. That has implications for staff burnout, retention, and the long-running question of how heavy the human cost of severe behavior caseloads actually is.

UNMC’s Munroe-Meyer Institute has run a parallel publication on the same group of therapists in the Journal of Organizational Behavior Management in May 2025, focused on different outcomes. The line between clinical research on clients and occupational health research on staff is thinning, and biometric data is the bridge.

Where the Field Sits Now

A January 2026 scoping review published in Research on Child and Adolescent Psychopathology by Catharina Bergwerff and colleagues at Leiden University and Vrije Universiteit Amsterdam mapped 85 studies of wearable technology in pediatric mental health. The bulk of that work, 54 studies, used wrist-worn actigraphy to monitor sleep, particularly in youth with attention-deficit/hyperactivity disorder, autism, or internalizing problems. Nineteen studies examined autonomic nervous system responses, eight covered motor activity, and five used brain activity or eye-gaze sensors. Only ten focused on externalizing behaviors or youth in forensic and residential care.

The review’s framing is instructive. Wearables are increasingly used for assessment in pediatric mental health, but rarely for intervention. The implementation literature is thin, the high-risk populations most relevant to ABA are underrepresented, and the field has not yet produced large-scale clinical trials. The capability the JAMA 2023 paper demonstrated has not yet been replicated at scale or in routine clinical settings, and almost none of the available work pairs the prediction signal with an intervention that has been shown to change outcomes downstream.

Across the autonomic-nervous-system studies, heart rate variability has consistently emerged as the most informative class of features. RMSSD (root mean square of successive differences) and SDNN (standard deviation of normal-to-normal intervals) are the two HRV parameters most often cited as predictive. Algorithms range from logistic regression and support vector machines to random forests, gradient-boosted trees, and convolutional neural networks. Reported emotion-classification accuracies span 65 to 91 percent depending on the algorithm, the feature extraction method, the labeling scheme, and the participant pool. The variance is not noise; it reflects how unsettled the methodology still is.

Most wearable research in autism still tracks sleep through actigraphy. The autonomic nervous system work is the smaller pile, and it is the one most likely to land in front of a BCBA in the next five years.

Ethics and Practical Limits

Continuous biometric monitoring of autistic children sits inside a longer disability-rights argument that the ABA industry has not finished. Disability rights advocates, including the Autistic Self Advocacy Network (ASAN), have pushed back for years against tracking, surveillance, and behavior-control technologies applied to autistic individuals, and against frameworks that treat autistic distress as primarily a clinical management problem rather than a communication signal. The biometric wearable case sits awkwardly inside that frame: a device that predicts aggression can also be described as a device that monitors a child’s emotional state with a fidelity no neurotypical child is subjected to.

Sensory profile adds a practical layer. Many autistic individuals have heightened responses to touch, pressure, and unfamiliar materials on the skin. The E4’s electrodes press constantly against the wrist, and several published studies note participants who found the device uncomfortable or unwearable. Earlier KeepCalm pilot data, on the other hand, reported that 85 percent of children on the autism spectrum wore the consumer-grade trackers without complaint. The device choice, the band material, and the time of day the child is asked to wear it are all variables that move tolerance up or down.

The cost barrier is real and not subtle. Research-grade wearables, paid platform subscriptions, integration work, and staff training all sit on top of the standard cost of an ABA service hour. Practices serving Medicaid populations, where margins are tightest and authorization processes the most constrained, are the ones least able to absorb the spend on technology that does not yet have a billing code attached. Open-source alternatives exist, but they have less validation and less institutional support than the commercial platforms, and they shift the validation burden onto the practice.

Regulatory pathway is the last open question. The Empatica Health Monitoring Platform now carries FDA 510(k) clearance for continuous physiological monitoring during sleep and for seizure detection through EpiMonitor. There is no equivalent clearance, yet, for behavioral prediction in autism, and the FDA has not signaled what the path to one would look like. Until a regulatory category exists, behavioral-prediction wearables will remain research tools, not approved medical devices, and clinics will use them inside research protocols rather than as billable clinical services.

What This Technology Is Not

It is worth being clear what the published research does and does not show. None of the studies discussed here demonstrate that wearable biosensors prevent aggressive behavior, only that they can predict it under specific conditions. None demonstrate that wearable-driven interventions outperform standard ABA practice in a controlled trial. None resolve the ethical question of consent for continuous monitoring in a population that often cannot give it in the way adult research subjects do. And none have produced a tool that a BCBA can deploy in a community-based ABA clinic tomorrow, billed through Medicaid, with documented patient and family acceptance.

The 2026 JABA paper is a technical report on three therapists. The 2023 JAMA paper is one rigorous prediction study at four primary care psychiatric inpatient hospitals. KeepCalm is a school-based educational team tool with promising co-design data and a separate 226-intervention internal validation. The scoping review is a map of a small and recently formed terrain. None of that is nothing. None of it is yet a product line.

A practice-floor view of biometric data is what wearable-augmented ABA looks like in concept. None of the published research has yet validated the workflow at the level a BCBA would need to bill it.

What ABA Practices Should Watch

For BCBAs and practice owners, the immediate operational priority is not buying wearables. It is following the regulatory and reimbursement signals. The questions worth tracking through 2026 and into 2027 include whether the FDA defines a clearance pathway for behavioral-prediction wearables, whether any payer pilots a coverage policy for biometric-augmented ABA, whether the Goodwin lab’s next replication scales the JAMA result outside the inpatient setting, and whether the KeepCalm team or another group produces a randomized trial showing improved outcomes from a wearable-driven intervention loop.

Practices that want to engage early have several low-cost options. Reading the primary literature directly, rather than the press coverage, is the most important. Following the JABA, Behavior Analysis in Practice, and Research in Autism Spectrum Disorders publication queues for new wearable studies. Engaging local university partners running pilots. And, when the time comes, asking direct questions about consent, discontinuation, and data ownership, the same questions any other clinical technology would face.

The bottleneck is not hardware. It is the regulatory pathway, the consent framework, the cost-benefit math, and the missing trials linking prediction to better outcomes. The 2026 JABA technical report from Munroe-Meyer is the next data point on the staff side. Goodwin’s lab has signaled further replication and intervention work is in the pipeline. The next scoping review, or the first randomized trial, will tell ABA whether biometric monitoring becomes a clinic tool or stays a research tool.

AT A GLANCE

SOURCES & REFERENCES