Most veterinary AI answers a single question at a single moment: given this patient and this presentation today, what are the likely differentials? That is genuinely useful — but it is also a snapshot. A great deal of clinical insight does not live in any one snapshot. It lives in the trajectory: is this animal recovering from surgery on schedule, is a rehabilitation programme actually working, is a senior patient drifting away from their own normal?
TrackerAI now answers those questions too. Our new longitudinal intelligence layer turns a stream of observations — from wearables, rehabilitation devices, or manual clinical entry — into baselines, trends, and drift detection, and then has the model explain what those trends mean in plain clinical language. Here is how it works, and why we built it the way we did.
The limit of the snapshot
Consider a dog six weeks after cruciate surgery. A single mobility score of 68/100 tells you very little on its own. Is that good or bad? The answer depends entirely on where the animal started and where it has been trending. 68 after a baseline of 55 is excellent progress. 68 after a peak of 84 is a regression worth investigating. The number is identical; the clinical meaning is opposite. Point-in-time analysis cannot tell these two cases apart. Longitudinal analysis can.
Baselines, trends, and drift
For every patient, TrackerAI establishes a per-metric baseline — anchored either to an explicit baseline-capture session or to the patient's earliest readings — and then, for each new observation, computes how the current value compares: the absolute change, the percentage change, and a spread-based significance test. Each metric is interpreted in the direction that matters clinically (higher mobility is good; higher pain is not), and a sustained deviation across several readings is flagged as drift rather than treated as a one-off noisy reading.
The result is a structured picture: which metrics are improving, which are declining, which have crossed a threshold that warrants attention — and an overall assessment of whether the patient is improving, stable, declining, or mixed.
Why the maths is deliberately not done by the model
This is the most important design decision in the entire feature. The baselines, deltas, and drift flags are computed deterministically, in code — never by the language model. The model's job is to interpret those numbers, not to produce them.
We call this Human-in-the-Reasoning, as distinct from Human-in-the-Loop. The difference is not cosmetic. A model that both invents the numbers and narrates them is a black box: the practitioner has nothing independent to check it against. By separating a transparent, reproducible computation layer from an explanation layer, we give the veterinarian the underlying evidence and the reasoning, so they can see exactly why a conclusion was reached, agree with it, or challenge it. The clinician remains the final line of accountability — with the information they need to actually exercise it.
Practically, this also means the model can never hallucinate a trend that the data does not support. The numbers it is given are the numbers it must reason about.
Getting data in
Observations reach TrackerAI through a single, versioned ingestion contract. Each observation carries the patient reference, a timestamp, a session type (rehabilitation, daily monitoring, post-operative check, and so on), and a flexible list of metrics — each with a value, a unit, and an optional confidence. Crucially, the metric list is open-ended: a partner can send mobility scores, gait symmetry, weight-bearing distribution, activity, or sleep without us needing to change the contract for every new signal.
Every observation also carries its provenance — the device, firmware, and processing-pipeline versions that produced it — and an idempotency key so the same reading is never double-counted. That provenance is not bureaucratic overhead; it is what makes the resulting record auditable.
From trend to action
When a metric drifts beyond threshold, or a significant decline appears, TrackerAI raises an alert with a severity and a plain-language rationale tied to the specific metric. Alerts generated by the deterministic rules are authoritative; the model can add context and recommended actions, but it cannot quietly downgrade a flag the data has earned. Each analysis is written to an append-only record — inputs, baseline window, model version, reasoning, and the numbers behind it — so a clinic always has a defensible history of what was assessed and why.
Why auditability is the point, not a feature
Veterinary medicine is moving into a period where the governance of clinical AI matters as much as its accuracy. Regulators are increasingly clear that AI in a healthcare setting cannot create a responsibility gap: practitioners need transparent, interpretable outputs to exercise genuine oversight rather than rubber-stamp a recommendation. A system that generates structured, longitudinal, traceable data is not just a monitoring tool — it is accountability infrastructure. We built the longitudinal layer to be that from the ground up: deterministic where it can be, explained where it should be, and recorded throughout.
Where it helps
- Post-operative recovery: confirm an animal is on its expected curve — and catch the one that quietly is not.
- Rehabilitation: show owners and clinicians whether a programme is working, with evidence rather than impressions.
- Chronic and orthopaedic disease: distinguish a bad day from genuine deterioration.
- Senior wellness: detect slow drift away from an individual animal's own normal, before it becomes a crisis.
Availability
Longitudinal monitoring is live in the TrackerAI platform now, alongside our existing completion, reasoning, and behavioural models. It is available through the web app and the REST API, and it is designed to consume data from wearable and rehabilitation devices as those integrations come online. If you are building hardware that generates structured movement, recovery, or behavioural data and want a transparent clinical-reasoning layer on top of it, we would like to talk: hello@trackerai.ai.