Two independent accuracy metrics, each answering a completely different question — and a diagnostic device can excel at one while failing the other.
Every diagnostic device — a blood glucose meter, a pulse oximeter alarm threshold, an ELISA assay, an AI-based imaging classifier — eventually gets reduced to a yes/no call against ground truth: does the patient actually have the condition, and did the device say so? From that single 2×2 comparison, two structurally different metrics fall out. Sensitivity answers: of everyone who truly has the condition, what fraction did the device correctly flag? Specificity answers a different question entirely: of everyone who truly does not have the condition, what fraction did the device correctly clear?Both are measured on the same 0–100% scale and both sound like generic "accuracy," which is exactly why they get conflated — but a device can be excellent at one and mediocre at the other, and which one matters more depends entirely on the clinical cost of being wrong in each direction.
The reason a device can be simultaneously excellent at one metric and poor at the other is that the two numbers are never computed from the same patients. Sensitivity is derived entirely from the population that truly has the condition — it literally cannot be affected by how the device behaves on healthy patients, because healthy patients don't appear in its denominator (TP + FN) at all. Specificity is the mirror image, derived entirely from the population that truly does not have the condition (TN + FP), and is mathematically blind to how many true disease cases the device catches or misses. A device could, in principle, catch every single diseased patient (100% sensitivity) while also flagging every single healthy patient as positive (0% specificity) — that would be a device that always says "positive," which trivially never misses a true case but is clinically useless because it never rules anything out either.
It doesn't — and worse, neither sensitivity nor specificity alone tells you what a patient actually wants to know: given my positive test result, what is the actual probability I have the condition? That question is answered by positive predictive value (PPV), which depends not just on sensitivity and specificity but also on disease prevalencein the population being tested. A test with 95% sensitivity and 95% specificity screening a disease with 0.1% true prevalence will still generate far more false positives than true positives in absolute numbers, because the healthy population being screened is so much larger than the diseased one — even a small false-positive rate applied to a huge healthy population outnumbers the true positives found in a tiny diseased population. This is exactly why population-wide screening tests for rare conditions require extremely high specificity, not just high sensitivity, and why a device's clinical usefulness can never be judged from sensitivity or specificity in isolation.
Explains the two core diagnostic-accuracy metrics used to validate any medical device or assay that produces a positive/negative call: sensitivity (the fraction of truly diseased patients correctly flagged) and specificity (the fraction of truly healthy patients correctly cleared) — two independent numbers derived from disjoint patient populations, both of which move in opposite directions when a device's decision threshold is adjusted.
Both terms sound like synonyms for generic "accuracy," and both are expressed as a percentage, which makes it easy to treat "95% sensitive" and "95% specific" as interchangeable claims about how good a test is. They are not measuring the same thing at all. Sensitivity is calculated exclusively from patients who truly have the condition; specificity is calculated exclusively from patients who truly do not. A device can score high on one and low on the other, and knowing only one number tells you nothing about the other.
Sensitivity (also called the true positive rate) = TP / (TP + FN) — of all patients who truly have the condition, what fraction did the device correctly call positive? Specificity (also called the true negative rate) = TN / (TN + FP) — of all patients who truly do not have the condition, what fraction did the device correctly call negative? Both metrics are properties of the test itself and, importantly, are independent of disease prevalence in the population — unlike positive predictive value (PPV) and negative predictive value (NPV), which do depend on prevalence and answer the patient-facing question "given my result, what is my actual probability of having the condition?"
Every diagnostic device with a binary or thresholded output — a lab assay, a continuous glucose monitor's hypoglycemia alarm, an AI-based radiology triage tool — has a decision threshold, and sliding that threshold trades sensitivity against specificity along a curve (formally, the ROC curve, receiver operating characteristic). A screening test intended to rule out a dangerous, easily-treatable condition in a broad population is usually tuned for high sensitivity, accepting more false positives, because missing a true case is far costlier than a false alarm that gets resolved with a confirmatory test. A confirmatory test used after a positive screen, or a test where a false positive triggers an invasive or risky follow-up procedure, is often tuned for high specificity instead, accepting that it may miss some borderline cases in exchange for not sending healthy patients to unnecessary risk. Regulatory submissions (FDA 510(k) and PMA clinical validation sections in particular) require both metrics reported with confidence intervals against a defined reference standard — reporting only one is considered an incomplete performance claim.
Only for a perfectly separable measurement with no overlap between the diseased and healthy population distributions — which is rare in real biological measurements, since biological variability usually causes at least some overlap. In practice, essentially every real diagnostic test involves some tradeoff, visible as the overlap region between the two population curves; the device's job is choosing (or letting the clinician choose) where along that tradeoff to sit.
A receiver operating characteristic (ROC) curve plots sensitivity (true positive rate) on the y-axis against 1 minus specificity (false positive rate) on the x-axis, across every possible decision threshold. It visualizes the entire sensitivity/specificity tradeoff for a given test in one curve, and the area under that curve (AUC) is a single summary number for overall discriminative performance, independent of any one chosen threshold.
Sensitivity and specificity are conditional on already knowing the true disease state — they describe how the test behaves within each group separately, so they don't change if you test a different mix of healthy and diseased patients. PPV and NPV instead answer a question that runs the other direction — given a result, what's the true state? — and that answer depends on how common the disease actually is in the population being tested, because it changes the base rate of true cases relative to false alarms even if the test's own sensitivity and specificity are unchanged.
There is no universal answer, but a common design pattern is a highly sensitive, lower-cost first-line screen (accepting some false positives to avoid missing true cases) followed by a highly specific confirmatory test on anyone who screens positive, to rule out those false alarms before committing to treatment. This two-stage approach is common in infectious disease testing algorithms and is a direct, practical application of the sensitivity/specificity tradeoff.
Try our Biomedical Engineering
More calculators, simulators, and guides for this discipline.
Applied Biomedical Engineering Professional Program
✨ Premium ContentA structured, paid professional training program for this discipline.