VibeRounds This course is built in the spirit of VibeRounds — Socratic learning (AI that questions rather than answers) and Guided Discovery, part of the wider Clinical Cognition Operating System.
Evidence-Based Medicine · Course for Techies

Lesson 5: Appraising Diagnostic Test Studies

Part 5 — a test isn't "accurate" in the abstract. Its usefulness depends on who you're testing and what you already suspected.

Lessons 3 and 4 were about therapy questions — does this treatment work? This lesson shifts to a different kind of question, one you already met in Lesson 2's question-type table: does this test actually tell me what I think it tells me? A positive result on a bad test can do more harm than the disease it's chasing, and a negative result on a bad test can offer false reassurance. Getting comfortable with test accuracy statistics is how you tell the difference between a genuinely useful test and a confident-sounding number.

Why this matters for techies: this is a classification-model problem you already understand from a different angle. A diagnostic test is a binary classifier with a disease as ground truth. Sensitivity and specificity are recall on the positive and negative classes. And just like a fraud-detection model trained on a 50/50 dataset behaves very differently in production where fraud is rare, a test's real-world performance shifts with how common the disease is in the population you point it at — that's the whole second half of this lesson.

The 2×2 Table: Where Everything Comes From

Every diagnostic accuracy statistic falls out of one table. You take a group of patients, run the test on all of them, and compare the test result against a gold standard — the best available way of knowing whether they truly have the disease (biopsy, culture, long-term follow-up, expert panel, etc.):

The Diagnostic 2×2 Table

Gold Standard
Disease Present Disease Absent
Test Result Positive True Positivea False Positiveb
Negative False Negativec True Negatived
Total a + c b + d

Green cells (a, d) are where the test got it right. Rust cells (b, c) are the two kinds of wrong — and they carry very different consequences.

A false positive (b) means telling a healthy person they're sick — anxiety, further invasive testing, sometimes unnecessary treatment. A false negative (c) means telling a sick person they're fine — missed or delayed treatment. Which error matters more depends entirely on the disease and the test's role, and it's worth asking explicitly before you look at any numbers.

Sensitivity and Specificity

These two numbers describe the test itself, independent of how common the disease is in whoever you happen to test:

Sensitivity = a ÷ (a + c) — of everyone who truly has the disease, what fraction does the test catch?

Specificity = d ÷ (b + d) — of everyone who truly does not have the disease, what fraction does the test correctly clear?
Memory aids that actually stick:

Sensitivity and specificity are properties of the test (and the population it was studied in), and in principle they don't change with disease prevalence — but what they let you conclude about an individual patient changes a great deal, which is exactly the trap the next section is about.

Predictive Values: What the Patient Actually Wants to Know

A patient with a positive result doesn't want to know "of people with the disease, what fraction test positive" — they want to know "given that I tested positive, what's the chance I actually have it?" That's a different question, answered by predictive values:

Positive Predictive Value (PPV) = a ÷ (a + b) — of everyone who tested positive, what fraction truly has the disease?

Negative Predictive Value (NPV) = d ÷ (c + d) — of everyone who tested negative, what fraction is truly disease-free?

Here's the catch: PPV and NPV depend heavily on prevalence — how common the disease is in the population being tested — even though sensitivity and specificity don't. This is the single most common source of diagnostic-test confusion, and it deserves a worked example.

Worked Example — Same Test, Two Populations

Take a test with 90% sensitivity and 90% specificity — a genuinely good test by most standards. Run it on 1,000 people twice, in two different settings.

Setting A — a screening population, 1% disease prevalence (10 truly have the disease, 990 don't):

PPV = 9 ÷ (9 + 99) = 9 ÷ 108 ≈ 8%. Of everyone who tests positive here, roughly 92% are false alarms — despite using a 90%/90% test.

Setting B — a specialist clinic, 50% disease prevalence (500 truly have the disease, 500 don't):

PPV = 450 ÷ (450 + 50) = 90%. Exact same test, exact same sensitivity and specificity — wildly different real-world meaning of a positive result.

Why this matters: this is why population-wide screening with a test validated in a high-prevalence specialist clinic can flood a low-prevalence population with false positives. The test didn't change. The context did.

Likelihood Ratios: A Prevalence-Independent Shortcut

Because PPV and NPV are prevalence-dependent, they don't transfer well between populations. Likelihood ratios (LRs) solve this: they're built from sensitivity and specificity, so they don't shift with prevalence, and they tell you how much a result should move your suspicion.

Positive Likelihood Ratio (LR+) = Sensitivity ÷ (1 − Specificity)

Negative Likelihood Ratio (LR−) = (1 − Sensitivity) ÷ Specificity
LR+ valueEffect on suspicion
> 10Large, often conclusive increase
5–10Moderate increase
2–5Small increase
1–2Rarely important
LR− valueEffect on suspicion
0.5–1Rarely important
0.2–0.5Small decrease
0.1–0.2Moderate decrease
< 0.1Large, often conclusive decrease

The formal way to use an LR is with pre-test odds × LR = post-test odds (via Bayes' theorem), which needs converting probability to odds and back. In practice, most clinicians use a shortcut — the Fagan nomogram — a printed or online chart that lets you draw a straight line from pre-test probability through the LR to read off post-test probability directly, no arithmetic required.

For non-medical readers: this is just Bayesian updating — the same math behind a spam filter recalculating "probability this email is spam" after seeing one more suspicious word. Pre-test probability is your prior; the likelihood ratio is how strongly the new evidence should shift it; post-test probability is your posterior.

ROC Curves and Choosing a Cutoff

Sensitivity and specificity so far assumed the test gives a clean positive/negative answer. Many real tests actually output a continuous value — a lab number, a score — and someone chose a cutoff to call it positive or negative. Move that cutoff and both sensitivity and specificity change, in opposite directions.

Illustrative ROC Curve

False Positive Rate (1 − Specificity) Sensitivity (True Positive Rate) chance (AUC = 0.5) this test (AUC ≈ 0.85) chosen cutoff 0 1 1

Each point on the curve is a different cutoff. Move toward the top-left corner (high sensitivity, low false-positive rate) and you're at a stricter or looser cutoff depending on which way you moved — there's no free lunch: gaining sensitivity by loosening a cutoff always costs specificity, and vice versa. The diagonal line is what you'd get from a test no better than a coin flip. The area under the curve (AUC) summarizes overall discriminative ability in one number: 0.5 is useless, 1.0 is a perfect test, and most decent clinical tests sit somewhere around 0.7–0.9.

Where the cutoff actually gets set is a judgment call, not a statistical one — it depends on the relative cost of false positives versus false negatives discussed earlier. A cancer screening test might accept more false positives to avoid missing cases; a test that triggers major surgery might do the opposite.

Appraising a Diagnostic Accuracy Study

The same three-question shape from Lesson 3 applies here, adapted to test accuracy:

  1. Are the results valid?
    • Was the test compared against an appropriate, independent gold standard — not itself, not a weaker reference?
    • Was the comparison blinded — did whoever read the gold standard not know the test result, and vice versa?
    • Did the study include a realistic spectrum of patients — both obvious and borderline cases — not just clear-cut sick people versus healthy volunteers (a common way accuracy gets artificially inflated)?
    • Was the test's cutoff decided before seeing the data, not tuned afterward to make the numbers look good?
  2. What are the results? Sensitivity, specificity, likelihood ratios, and their confidence intervals — ideally with the raw 2×2 numbers reported so you can recompute anything.
  3. Will this help my patient? Is your patient's pre-test probability anywhere close to the study population's prevalence? Is the test available, safe, and affordable in your setting? And critically — will the result actually change what you do next? A test that can't change management, whatever it shows, usually shouldn't be ordered.
Pro tip: before ordering or interpreting any test, try to state your pre-test probability in words first — "I think there's maybe a 20% chance this patient has a PE" — even roughly. Without that starting estimate, a likelihood ratio has nothing to multiply against, and a lone positive or negative result is much easier to over- or under-interpret.

Homework for Lesson 5

  1. Find a diagnostic accuracy study for a test relevant to your patient from Lesson 1 (or any clinical question you're curious about). Extract the raw 2×2 numbers if reported, and calculate sensitivity, specificity, PPV, and NPV yourself.
  2. Recalculate PPV and NPV for a prevalence different from the study's — for example, at 5% and at 50% — and note how much they move even though sensitivity and specificity stay fixed.
  3. Calculate LR+ and LR− for the test, and state in one sentence how much a positive result should shift your suspicion for a patient you estimate started at roughly 20% pre-test probability.
  4. One paragraph: is this test's spectrum of patients (as described in the study) similar enough to your patient that its numbers actually apply?

Resources for This Lesson

CEBM — Diagnostic test calculators (sensitivity, specificity, LRs) CEBM Oxford — Critical Appraisal Tools (incl. diagnostic studies) CASP UK — Diagnostic Test Study Checklist PubMed — search medical research papers