The question changes when the denominator changes
Sensitivity starts with people who have a condition and asks how many receive a positive result. Positive predictive value starts with positive results and asks how many belong to people with the condition. These sound similar in casual conversation, but they answer different questions.
Specificity starts with people without the condition and asks how many receive a negative result. Negative predictive value starts with negative results and asks how many belong to people without the condition. A statement such as “the test is 90% accurate” obscures these distinctions unless it defines the statistic and population. It may refer to sensitivity, overall classification accuracy, or something else entirely.
A hypothetical example with 1,000 people
Suppose 10 of 1,000 people have a condition. Assume a test has 90% sensitivity and 95% specificity in this population. The expected counts below include fractions because they describe an average under the assumptions, rather than an actual observed group.
| Result | Condition present: 10 people | Condition absent: 990 people |
|---|---|---|
| Positive | 9 true positives | 49.5 false positives |
| Negative | 1 false negative | 940.5 true negatives |
Among the 58.5 expected positive results, only 9 are true positives. The positive predictive value is about 15.4%. That is compatible with both the stated sensitivity and specificity. There are many more people without the condition, so even a relatively small false-positive proportion produces a substantial number of positive results.
Among the expected negative results, about 99.9% are true negatives. This high negative predictive value partly reflects the low starting frequency of the condition. It is not a universal promise about what a negative result means for every person.
Change the starting frequency, and the result changes
Now suppose 100 of 1,000 people have the condition, while sensitivity and specificity remain the same for illustration. There would be 90 expected true positives and 45 false positives. Positive predictive value would be 90 divided by 135, or about 66.7%.
The test's assumed sensitivity and specificity did not improve. The tested population changed. In actual practice, these performance measures may also differ across populations because the mix of disease severity, alternative conditions, and testing circumstances changes. Holding them fixed is useful for understanding the arithmetic, not a guarantee of transfer between settings.
This is why a predictive value from a study cannot simply be copied into a different screening program. A population selected because of symptoms or prior findings may have a different starting probability from an unselected population.
Thresholds involve tradeoffs
Some tests and AI tools produce a score that becomes positive or negative only after a threshold is chosen. Changing that threshold can alter sensitivity and specificity. A threshold that catches more potential cases may also produce more false positives; the exact tradeoff depends on the score distributions.
There is no context-free “best” balance. Missing a condition, prompting an unnecessary follow-up, delaying care, and consuming limited assessment capacity can have different consequences. A study that reports several thresholds is not necessarily indecisive. It may make those tradeoffs more visible than a single headline number.
What counts as the true condition?
The calculations require a way of determining who has the condition. A suitable reference standard and its application are therefore central to the interpretation. If only some participants receive reference assessment, or if ambiguous cases are excluded without explanation, the reported results may not represent the full intended population.
FDA guidance on diagnostic-test studies distinguishes accuracy estimates using a reference standard from agreement measures using a non-reference comparator. It also emphasizes reporting enough detail to interpret estimates and their uncertainty. A precise percentage without a sample size or confidence interval can look more settled than the underlying evidence warrants.
Reading an AI detection claim
The same reasoning applies when a health-related classifier uses machine learning. The model architecture does not change the denominators. A claim about high sensitivity still leaves questions about false positives, predictive values, thresholds, and the population tested.
For an individual, a test result is interpreted alongside the reason for testing and other clinical information. The worked numbers here explain population-level statistics; they do not determine a person's diagnosis. For a research reader, the most revealing detail is often the full table of results. It shows which errors occurred, how often, and among whom, instead of compressing very different outcomes into one reassuring score.
