1. Sensitivity, Specificity & Predictive Values

(Screening & Diagnostic Accuracy Measures)

When are they used?

Used when evaluating how good a screening or diagnostic test is.

Every patient falls into two realities:

• Disease present or disease absent
• Test positive or test negative

We arrange these in a two-way table:

Sensitivity

“Positive in disease”

If someone has the disease, how often is the test positive?

Sensitivity = A / (A + C)

High sensitivity = low false negatives.

Specificity

“Negative in health”

If someone is healthy, how often is the test negative?

Specificity = D / (B + D)

High specificity = low false positives.

Positive Predictive Value (PPV)

If the test is positive, what is the chance the patient actually has the disease?

PPV = A / (A + B)

PPV depends heavily on prevalence.

Negative Predictive Value (NPV)

If the test is negative, what is the chance the patient is actually free of disease(healthy)?

NPV = D / (C + D)

Likelihood Ratio (LR)

How much a test result shifts probability.(

LR: If the test is positive, how much more likely the patient is to have the disease than not have it.

LR+ = Sensitivity / (1 – Specificity)

Values >10 = strong rule-in

Values <0.1 = strong rule-out

LR− (Negative Likelihood Ratio) ❗

This is the missing piece.

Question it answers:

If the test is negative, how much does it reduce the chance of disease?

Correct formula:

LR− = (1 − Sensitivity) / Specificity

Interpretation:

• <0.1 → strong rule-out
• 0.1–0.2 → moderate rule-out
• 1 → no diagnostic value

Example from the book

Blood test for gastric cancer (n = 100):

Calculations:

• Sensitivity = 20/25 = 0.8 (80%)
• Specificity = 45/75 = 0.6 (60%)
• PPV = 20/50 = 0.4 (40%)
• NPV = 45/50 = 0.9 (90%)
• LR+ = 0.8 / (1 – 0.6) = 2

Interpretation:

• Test is good at ruling out (high NPV)
• Poor at ruling in (low PPV)

2. Level of Agreement & Kappa (κ)

(Inter-observer or Test–retest Agreement)

When is it used?

Used when:

• two observers classify something
• or one test is repeated
• and outcomes fall into categories (especially ordered categories)

Example: CIN1, CIN2, CIN3 classification in cervical smears.

Why not use simple percent agreement?

Because some agreement occurs by chance.

Kappa corrects for this.

Kappa (κ) — Interpretation

Kappa ranges from 0 to 1:

Example from the book

Two hospitals read the same cervical smear slides:

• κ = 0.3 → poor agreement

Meaning: patients could get different results depending on the lab → unreliable classification.

When NOT to use Kappa

If measurements are continuous (e.g., blood glucose), use:

• Intraclass correlation coefficient (ICC) instead.

Summary for Exams

Sensitivity

Positive in disease → A / (A+C)

Specificity

Negative in health → D / (B+D)

PPV

If test positive → chance disease is present

NPV

If test negative → chance disease is absent

LR+

Sensitivity / (1 – Specificity)

Kappa (κ)

Agreement beyond chance

• 0.7 = good
• <0.4 = poor