Accuracy is one of the most persuasive words in healthtech because it sounds complete. A percentage follows it, and the reader is invited to imagine a device being right almost all the time. The awkward question is what being right actually means.
Clair Health's founders reported 94.1% accuracy for classifying cycle phase. That is a result attached to a particular task. Its clinical interpretation depends on preserving that task as the number travels into headlines, investor discussions and patient expectations.
The decimal point is the least informative part of the claim. The definition of the endpoint does the real work.
A percentage needs a complete sentence
An interpretable accuracy statement identifies what was predicted, what counted as correct, which observations were included and what reference established the answer. It also explains how the model was evaluated and how uncertainty around the estimate was calculated.
Without those details, accuracy can mean several different things. It might describe the proportion of daily labels that matched a reference, successful classification of complete cycles, identification of a transition or performance after excluding unreadable signals. A clinician cannot safely substitute one meaning for another.
Cycle-phase classification also differs from estimating the concentration of a hormone. A label answers a categorical question. A concentration answers a quantitative one. Even perfect categorisation would leave unanswered how close a reported concentration lies to a suitable reference measurement.
This is the point at which polished communication can outrun the evidence. A task-specific result becomes a general impression of hormonal precision, despite the absence of a demonstrated link between those two claims.
Overall accuracy can hide the difficult moments
Consider a hypothetical evaluation with 1,000 observations. Nine hundred come from relatively straightforward periods and 100 from transitions that are harder to identify. Suppose the system gets all 900 straightforward observations right, but only 40 transition observations right.
Its overall accuracy is 94%. Its accuracy on the difficult observations is 40%.
Those figures are invented to demonstrate the arithmetic, not to estimate Clair's performance. They show why the overall result cannot tell a clinician how the system behaves at the moment that matters for a particular question.
A relevant evaluation would therefore show the distribution of observations and performance across the different classes or periods. It would explain whether some errors are concentrated in a small group of participants, a particular transition or situations with incomplete data.
The clinical importance of an error also depends on its location. Confusing neighbouring categories may have little consequence for one use and substantial consequence for another. A single aggregate percentage treats those mistakes as interchangeable unless the analysis says otherwise.
Sensitivity leaves another half of the story open
Sensitivity describes how many true events a system detects. It does not establish how often an alert corresponds to a true event. That second question depends on false positives and the frequency of the target event in the evaluated setting.
Here is another explicitly hypothetical example. Across 1,000 candidate windows, 100 contain an event. A detector with 90% sensitivity identifies 90 of them. If its specificity is also 90%, it incorrectly flags 90 of the 900 event-free windows. There are 180 alerts, of which half are correct.
Nothing in that calculation concerns Clair's actual false-positive rate. The example explains why an impressive sensitivity cannot stand in for the probability that a particular alert is right.
The STARD reporting framework makes transparent diagnostic-accuracy reporting a central objective. Its underlying discipline is useful here: define the test, reference, participants and analysis so that readers can inspect what the result means.
Timing can matter more than the headline
For an event detector, a successful result depends on the permitted time window. Calling an event correctly within several days is a different achievement from identifying it on the day, and both differ from predicting it before it occurs.
A study should distinguish retrospective recognition from prospective prediction. If an algorithm can use observations recorded after an event, it may identify a pattern that would have been unavailable to a user deciding what to do earlier. That can still be a useful retrospective feature, provided it is described accordingly.
Likewise, the reference schedule limits what can be established. Sparse reference measurements may be sufficient for one classification task and inadequate for evaluating rapid changes. A comparison should explain how observations were aligned and what happened when the reference itself was ambiguous.
The right question is whether the evaluation resembles the situation in which the output will be used. A beautifully precise result for a different timing question may be of limited practical relevance.
Ask what happens when the system cannot answer
A wearable may encounter poor contact, interrupted use or insufficient signal. The evaluation should explain whether those periods were excluded, labelled uncertain or converted into estimates using surrounding observations.
Excluding difficult periods can produce a useful estimate of accuracy when a valid signal is available. It leaves a separate availability question: how often will a user receive an answer of that quality? An app that answers reliably half the time and one that answers almost continuously are different products.
Showing uncertainty can be a strength. An honest gap in a graph may communicate more than an uninterrupted line that disguises where the model had to rely heavily on assumptions. Clinicians need to know which type of output is in front of them.
For prediction models used in care, external-validation guidance in The BMJ addresses performance and clinical usefulness separately. That separation prevents statistical success from becoming an automatic treatment recommendation.
A useful result should survive a useful question
The appropriate response to a strong headline is to specify the decision it might support. What would change if the output were available? Which errors could change that decision? What existing information would it improve upon?
These questions can be practised through the iatroX Tutor, using a hypothetical study to distinguish the measured endpoint from the intended application. Clinicians can also use the iatroX methodology to inspect the platform's approach to evidence when exploring related clinical questions.
Clair's reported result deserves to be discussed in its proper category. It should neither be dismissed because it comes from a company nor promoted into a claim the endpoint did not test. The useful challenge is exact: show how the number connects to the information being sold and the decision the user is expected to make.
Frequently asked questions
Does 94.1% accuracy mean the device gives correct hormone levels 94.1% of the time?
No. The reported percentage concerns cycle-phase classification. Quantitative hormone accuracy requires its own definition, suitable reference measurements and analysis of the size and direction of errors.
Can high overall accuracy coexist with weak performance on important events?
Yes. Common or easier observations can dominate the total. Performance by class, event, participant and relevant subgroup helps reveal weaknesses that an overall percentage can hide.
What is the most useful follow-up to an accuracy claim?
Ask for the complete endpoint definition and the underlying error results. Then assess whether those results address the specific timing, population and decision for which the output is being proposed.