Disengagement can be predicted from behavioural pattern breaks — lengthening interaction latency, shortening dwell, rising tab-switch duration — measured against each learner's own baseline. Published work reports 73.3% prediction accuracy with 40% of the activity still remaining.
What actually predicts it
Not a single signal — a pattern break against the learner's own established behaviour.
| Feature | What to measure |
|---|---|
| Interaction latency | Rising against this learner's own median |
| Dwell per section | Falling below their own typical |
| Tab-switch duration | Excursions getting longer, not just more frequent |
| Scroll velocity | Accelerating — skimming for the end |
| Focus loss | Window losing focus without a tab change |
| Replay followed by advance | Tried again, gave up, moved on |
The last one is the most informative and the most commonly missed. A learner who replays a segment and then advances without their comprehension improving has just decided to stop trying. That decision usually precedes departure by several minutes.
Why per-learner baselines are non-negotiable
Absolute thresholds do not work here, and it is worth being blunt about why.
Some people work with a dozen tabs open and switch constantly while fully engaged. Others open one window and never leave it. A fixed rule — "three tab-switches means disengaged" — permanently flags the first group and never catches the second.
The model has to measure deviation from each individual's own pattern. This needs a calibration window: the first few minutes of a session, or a rolling average across previous sessions. Without it, you are mostly measuring browsing habits.
Population-threshold models systematically penalise people whose normal working style differs from the majority — which correlates with role, seniority, neurodivergence and whether someone is working from a laptop between meetings. Per-learner baselining is not just more accurate; it is the difference between a defensible system and a discriminatory one.
How much warning you get
Boote, Agarwal and Mostow achieved 73.3% prediction accuracy with 40% of the activity still remaining.
Two ways to read that.
Optimistically: in a 40-minute module that is roughly 16 minutes of warning. Ample time to do something.
Realistically: more than a quarter of the flags are wrong. That is fine for adapting content and completely unacceptable for anything that touches a person's record.
The accuracy you need is a function of what a false positive costs. Match the intervention to the confidence, not the other way round.
What to do with a prediction
Ranked by how well they actually work.
1. Fix the content (best). If the model flags the same section for most learners, you have not found forty distracted people. You have found one bad section. This is aggregate, needs no individual identification, and produces a permanent fix rather than a per-session rescue.
2. Adapt the content in-session. Branch to a shorter path, surface a worked example, offer a summary. Invisible to the learner, no false-positive cost, no monitoring conversation.
3. Offer a break point. "This is a good place to stop and come back" is genuinely useful and does not read as surveillance. Resumption beats forcing a bad session to completion.
4. Notify an instructor (cohort settings only). Defensible in education with a tutor relationship. Rarely appropriate in corporate L&D.
5. Prompt the learner directly. Frequently counterproductive. An "are you still there?" modal to someone already irritated accelerates the exit and makes the monitoring visible.
6. Notify a manager. Do not. This converts a content-improvement tool into a performance-management one, guarantees gaming, and will not survive a proportionality assessment.
The governance you need first
Before building any of this:
- A DPIA. Systematic monitoring of employees or students with automated prediction is high-risk processing. Not optional.
- A stated purpose, and a prohibition on other uses. Write down that engagement data will not be used for performance management, and put it in policy and in the vendor contract.
- Transparency. Tell people it exists and what it does. Discovered monitoring is far more damaging than disclosed monitoring.
- No automated consequences for individuals. Article 22 territory, and a bad idea regardless.
- An accuracy statement. If you are flagging people, you should be able to say how often the flag is wrong.
The recommendation
Build the model. Use it on the content.
The instinct with an early-warning system is to point it at people, because that is what "early warning" implies. But the finding that pays for the system is almost always aggregate: this section loses most learners at this point. That insight fixes the problem permanently for everyone, needs no personal data, and requires no difficult conversation with your works council.
Individual-level prediction is the harder build, the harder sell, and the smaller return.
Frequently asked questions
How much warning does a prediction model give you?
Should I alert the learner directly?
Should I alert their manager?
What accuracy is good enough to act on?
See where your content loses people
Book a walkthrough and we will show you the engagement data on your own content.