Sensitivity And Specificity

These are characteristics of the test itself, determined by the test's biology/technology and the threshold chosen. They are theoretically independent of disease prevalence in the tested population.

Sensitivity (True Positive Rate)

Sensitivity = TP / (TP + FN) = a / (a + c)

• Proportion of people WITH the disease who test POSITIVE
• "Of all the sick people, how many does the test catch?"
• High sensitivity → few false negatives → good for RULING OUT disease
• SnNOut: A highly Sensitive test, when Negative, rules Out the disease
• Clinical application: Use highly sensitive tests for screening and when missing the disease has severe consequences (e.g., troponin for MI, D-dimer for PE)

Specificity (True Negative Rate)

Specificity = TN / (TN + FP) = d / (b + d)

• Proportion of people WITHOUT the disease who test NEGATIVE
• "Of all the healthy people, how many does the test correctly identify as healthy?"
• High specificity → few false positives → good for RULING IN disease
• SpPIn: A highly Specific test, when Positive, rules In the disease
• Clinical application: Use highly specific tests for confirmation and when a false positive carries significant consequences (e.g., biopsy for cancer diagnosis)

Extrinsic Test Properties (Predictive Values)

These depend on the prevalence of disease in the tested population. The same test gives different predictive values in different clinical settings.

Positive Predictive Value (PPV)

PPV = TP / (TP + FP) = a / (a + b)

• Probability of having the disease given a POSITIVE test result
• "If the test is positive, what is the chance the patient actually has the disease?"
• Increases with higher prevalence, higher specificity
• In low-prevalence settings, even specific tests yield many false positives

Negative Predictive Value (NPV)

NPV = TN / (TN + FN) = d / (c + d)

• Probability of NOT having the disease given a NEGATIVE test result
• "If the test is negative, what is the chance the patient is truly disease-free?"
• Increases with lower prevalence, higher sensitivity
• In high-prevalence settings, even sensitive tests may not sufficiently rule out disease

Prevalence Dependence

This is the most clinically important and most commonly misunderstood concept.

The Prevalence Effect

• Low prevalence (e.g., screening general population): PPV drops dramatically, NPV is very high. Even with sensitivity and specificity both at 99%, if prevalence is 1%, PPV is only 50% — half of all positive results are false positives.
• High prevalence (e.g., symptomatic patients in specialty clinic): PPV is high, NPV drops. The test becomes more useful for confirmation.

Worked Example

A test with 95% sensitivity and 90% specificity applied to:

Prevalence 1% (1,000 people, 10 with disease)

• PPV = 9.5/108.5 = 8.8%
• NPV = 891/891.5 = 99.9%

Prevalence 50% (1,000 people, 500 with disease)

• PPV = 475/525 = 90.5%
• NPV = 450/475 = 94.7%

Same test. Radically different clinical utility depending on who you test.

Clinical Interpretation Framework

Step 1: Estimate Pre-test Probability

Before ordering any test, estimate the probability of disease based on:

• History and physical examination
• Risk factors and demographics
• Clinical prediction rules (Wells score, HEART score, etc.)
• Local disease prevalence

Step 2: Choose the Right Test

• If pre-test probability is low and you want to exclude disease → choose a highly sensitive test (SnNOut)
• If pre-test probability is moderate-high and you want to confirm disease → choose a highly specific test (SpPIn)
• If pre-test probability is very low (<5%) → consider whether testing adds value at all
• If pre-test probability is very high (>95%) → testing may be unnecessary; treat empirically

Step 3: Interpret the Result in Context

• A positive result on a sensitive test in a low-prevalence population → likely false positive → confirm with a more specific test
• A negative result on a specific test in a high-prevalence population → may be false negative → consider further workup

Trade-offs and Threshold Selection

The ROC Curve

• Plots sensitivity (y-axis) vs 1-specificity (x-axis) across all possible thresholds
• Area Under the Curve (AUC) measures overall discriminative ability (0.5 = coin flip, 1.0 = perfect)
• AUC 0.7-0.8 = acceptable, 0.8-0.9 = excellent, >0.9 = outstanding
• The optimal threshold depends on clinical consequences of FP vs FN

Threshold Selection Principles

• When missing disease is catastrophic (cancer, MI, meningitis): Lower the threshold → increase sensitivity at the cost of specificity → accept more false positives to catch all true positives
• When treatment is harmful or expensive (surgery, chemotherapy): Raise the threshold → increase specificity at the cost of sensitivity → accept some missed cases to avoid unnecessary treatment
• Screening vs diagnostic thresholds: Screening tests use lower thresholds (maximize sensitivity); confirmatory tests use higher thresholds (maximize specificity)

Accuracy and Related Measures

• Accuracy = (TP + TN) / N — Can be misleading in imbalanced datasets (high accuracy from simply predicting the common outcome)
• False Positive Rate (FPR) = FP / (FP + TN) = 1 - Specificity
• False Negative Rate (FNR) = FN / (TP + FN) = 1 - Sensitivity
• F1 Score = 2 x (PPV x Sensitivity) / (PPV + Sensitivity) — harmonic mean, useful when classes are imbalanced

Common Pitfalls

1. Confusing sensitivity with PPV: "The test is 99% sensitive" does NOT mean "a positive result is 99% likely to be correct"
2. Ignoring prevalence: Applying screening test results from high-prevalence research populations to low-prevalence clinical settings
3. Spectrum bias: Sensitivity/specificity measured in studies comparing extreme cases (advanced disease vs healthy controls) may not apply to the clinically relevant gray zone
4. Verification bias (workup bias): When only test-positive patients get the reference standard, sensitivity is overestimated and specificity is underestimated
5. Using a single threshold when a range of thresholds exists: Many tests (troponin, D-dimer, PSA) have different clinical implications at different cutpoints
6. Forgetting that sensitivity and specificity can change across subgroups: Age, sex, comorbidities, and disease severity all affect test performance