Vision AI for automated ABC data collection: the antecedent problem in ABA behaviour logging

At an autism conference in Bengaluru last week, I expected the corridor conversations to be about AI. They were not. Over two days, the theme I heard most often from therapists was far less glamorous and far more urgent: write it down.

Senior clinicians telling juniors that documentation is not paperwork, it is the raw material of clinical reasoning. Without the log, there is no analysis. Without the analysis, there is no behaviour plan.

I build computer vision systems, and my son is non verbal and autistic. I went home and looked seriously at the evidence behind what those therapists were saying. What I found is a field that agrees almost unanimously on the importance of behaviour data collection, and a body of research that quietly documents how rarely it actually happens.

This piece is written for the people who live inside that problem: BCBAs and behaviour analysts, RBTs and behaviour technicians, special educators, occupational and speech therapists, and the researchers who depend on their records.

Why ABC data collection is the whole game in ABA

The workhorse of behavioural practice is ABC recording: Antecedent, Behaviour, Consequence. You record what happened immediately before a behaviour, the behaviour itself, and what followed. The purpose is not archival. By recording the antecedents that trigger a behaviour and the consequences that follow it, practitioners gain insight into why the behaviour occurs and what might be reinforcing it.

That is the entire logic of behavioural intervention, and it is the evidentiary base of a functional behaviour assessment. You cannot change a behaviour you cannot explain, and you cannot explain it without knowing what preceded it. Clinicians systematically record using standardised methods, and analysts then use that documentation to identify patterns, evaluate whether interventions are working, and adjust treatment plans for the individual child.

The log is not administrative overhead sitting next to the clinical work. The log is the clinical work, in its raw form.

Why ABC data collection fails in practice

Here the literature turns uncomfortable.

The methods are fragile by design. Traditional approaches, including ABC recording, require clinicians to manually observe and document behaviours on paper or in basic digital formats. These methods are prone to human error, particularly in chaotic or group settings, and continuous observation is not feasible for therapists managing several clients at once. The method assumes continuous observation. The working conditions make continuous observation impossible.

Different observers produce different data. This is the problem interobserver agreement is designed to measure, and it is a real constraint in day to day practice. Observer variability, meaning differences in how individual data collectors interpret and record the same behaviour, introduces inconsistency into the record. Inconsistencies compound when collectors apply different methods or do not adhere to standardised protocols, and frequent rotation of responsibilities among technicians makes this worse. A behaviour recorded by one therapist on Monday and another on Thursday may not be the same data point at all.

Delay degrades everything. Manual methods also introduce delays in analysis and make trends difficult to visualise, creating a disconnect from real time decision making. By the time the pattern is visible, the window to act on it has often closed.

And people sincerely overestimate their own compliance. This is the finding that should unsettle anyone who relies on self maintained records. In a study of sleep diary adherence, participants who were unaware that their diary usage was being objectively tracked substantially over reported how consistently they had completed it, and the authors concluded that non adherence with diary protocols poses a genuine challenge for researchers relying on the tool. The failure is not dishonesty. It is ordinary human memory, under load.

Now place that finding in a home where a child is in distress, or in a therapy room with three other children waiting. The expectation that a caregiver or clinician will pause, retrieve a form, and accurately reconstruct an incident is not a reasonable expectation. It is a design flaw.

When research programmes had to fix behaviour logging first

The evidence is not in papers about logging. It is in papers that had to solve logging before they could study anything else.

Janssen’s My JAKE platform, part of the Autism Knowledge Engine, was built on precisely this premise. Its designers noted that autism interventions are commonly evaluated using retrospective caregiver reports, which often require recalling specific behaviours weeks after they occurred, potentially reducing the accuracy of those ratings. The app was created to let caregivers log symptoms continuously and track progress, specifically to mitigate the difficulties of retrospective reporting. A pharmaceutical research programme concluded it could not trust its own outcome measures until it fixed the logging problem underneath them.

The same pattern appears in ProVIA Kids, a smartphone app for caregivers of children with autism and developmental disabilities. It uses algorithm based behaviour analysis to identify the causes of challenging behaviour and offer individualised guidance, and in an eight week pre post study with 18 caregivers it collected caregiver stress through ecological momentary assessment inside the app rather than through recall.

Read those two together and a pattern emerges. When researchers need behavioural data they can trust, they stop asking people to remember and start capturing in the moment. Logging is not treated as a clerical preliminary. It is treated as the methodological bottleneck standing between a research question and a credible answer.

The antecedent problem in ABC recording

There is a deeper issue, and it is the one I keep returning to.

Even a perfectly disciplined human log has a structural blind spot. A caregiver begins recording when they notice the behaviour. But the antecedent, the “A” in ABC, happened before anyone was paying attention. It is therefore always reconstructed backwards, from memory, after the fact, by someone who was mid crisis at the time.

So the single most analytically valuable element of the record, the antecedent that a function hypothesis rests on, is the element collected under the worst possible conditions.

This is where continuous automated capture does something manual methods structurally cannot. A system that was already watching does not need to reconstruct the two minutes before an incident. It has them.

The antecedent is not recalled, it is recorded.

That is not a convenience improvement over paper. It is a different category of data.

What automated behaviour logging makes possible, and what it does not

If behaviours are logged automatically and continuously, across the home and the therapy room, a few things become possible that are difficult today. Patterns across settings become visible, so a behaviour that clusters after a particular transition, at a particular hour, or only in one environment can be seen rather than suspected. Session notes can be reviewed against what actually happened rather than what was remembered. A functional behaviour assessment can draw on weeks of continuous antecedent data rather than a handful of sampled observations. And the clinician’s scarce time shifts from capturing data to interpreting it.

I want to be careful about the limits; overclaiming here would be easy and wrong.

An automated log records observable events. It does not record internal states. When my son was in pain from a toothache, what a camera would have captured was withdrawal and distress, not the cause. The log would have shown a clinician a pattern worth investigating. It would not have supplied the diagnosis.

Automated logging also does not perform the analysis. It produces a cleaner, more complete, time stamped substrate. The interpretation, the hypothesis about function, the intervention design, all of that remains clinical work belonging to trained humans. A better microscope does not replace the pathologist.

And these systems inherit obligations. Continuous capture in the life of a child demands consent, governance, minimisation, and the discipline to log behaviour without profiling a person.

The open question for BCBAs and ABA researchers

Here is what I think is genuinely unexamined, and what I would like to work on with clinical researchers.

We have good evidence that manual behavioural documentation is incomplete and inconsistent. We have research programmes that adopted continuous capture: retrospective reporting was not trustworthy enough. What we do not yet have is a clear answer to the next question:

Not whether the technology detects behaviours accurately, which is a separate and answerable engineering question. Whether a clinician holding a complete, continuous, time stamped record reaches different, better founded conclusions than a clinician holding a partial one reconstructed from memory. Whether the plans built on that record work better for the child.

That study needs clinicians, not engineers. If you are a BCBA, a behaviour analyst, a special educator, or an autism researcher, and that question interests you, I would like to hear from you.

We spend a great deal of energy treating behaviours. We spend remarkably little making sure we accurately recorded what happened before them.

Pratush Charan is Head of Product at Ethan AI and a computer vision researcher, and the parent of a non verbal autistic child. Ethan AI is a video and vision AI system for automated behaviour logging in autism care, built to support ABC data collection and ABA practice on cameras therapy centres and families already own.