CK040 - Diagnosis

Question 1

What is the odds at the ‘no treat-treat threshold’ ?

Answer 1

Harm / Benefit

Question 2

What is the probability at the ‘no treat-treat threshold’ ?

Answer 2

Harm / (Harm + Benefit)

Question 3

What is the ‘net harm’ ?

Answer 3

Treatment if no disease (so ‘difference in outcome with/without treatment if disease is NOT present’)

Question 4

What is the ‘net benefit’ ?

Answer 4

Treatment if disease (so ‘difference in outcome with/without treatment if disease is present’)

Question 5

What is the ‘sensitivity’ ?

Answer 5

True Positive Rate
(so the probability of getting a positive test result, GIVEN disease is present)

Question 6

What is the ‘specificity’ ?

Answer 6

True Negative Rate
(so the probability of getting a negative test result, GIVEN disease is NOT present)

Question 7

What is the ‘PPV’ ?

Answer 7

Probability of having the disease, GIVEN a positive test result

Question 8

What is the ‘NPV’ ?

Answer 8

Probability of having NO disease, GIVEN a negative test result

Question 9

If you want to make sure you do not miss any diagnosis, what should be high?

Answer 9

Sensitivity

Question 10

What is the Bayes’s formula for tests?

Answer 10

Posterior odds = Prior odds x Likelihood Ratio

Question 11

What are the formulas for the ‘Likelihood Ratio’ ?

Answer 11

• p(R | D+) / p(R | D-), where R is a certain test result
• TPR / FPR for a positive test result
• FNR / TNR for a negative test result

Question 12

What does a ‘LR > 1’ mean?

Answer 12

That you are more likely to find this test result in diseased

Question 13

What does a ‘LR < 1’ mean?

Answer 13

That you are more likely to find this test result in non-diseased

Question 14

What happens to the thresholds when the specificity gets bigger?

Answer 14

‘No treat-test threshold’ gets smaller

Question 15

What happens to the thresholds when the sensitivity gets bigger?

Answer 15

‘Test-treat threshold’ gets smaller

Question 16

What is on the axis of a ‘ROC curve’ ?

Answer 16

• x-axis –> FPR (1 - specificity)
• y-axis –> TPR (sensitivity)

Question 17

What is the KEY step in constructing a ‘ROC curve’ ?

Answer 17

Ordering the test results according to increasing Likelihood Ratio

Question 18

What is the assumption of ‘conditional independence’ ?

Answer 18

Both for presence & absence of disease, the probability of combination of results is the same as the product of the results independenctly

Question 19

When combining tests, what happens to the ‘sensitivity’ and ‘specificity’ when we us ‘OR’ ?

Answer 19

• Sensitivity goes up
• Specificity goes down

Question 20

When combining tests, what happens to the ‘sensitivity’ and ‘specificity’ when we us ‘AND’ ?

Answer 20

• Sensitivity goes down
• Specificity goes up

Question 21

What is ‘calibration’ ?

Answer 21

Is our model predicting what we think it should?

Question 22

How to check for ‘calibration’ ?

Answer 22

Plot the predicted probabilities (x-axis) against the observed probabilities (y-axis). This line should be linear, and go through the origin.

Question 23

What is ‘discrimination’ ?

Answer 23

How good does the prediction model discriminate diseased from non-diseased?

Question 24

How to check for ‘discrimination’ ?

Answer 24

Use the AUC of the ROC-curve

Question 25

What happens on the ROC curve when there is a high ‘prior odds’?

Answer 25

You move to the top right (because the LR decreases)

Question 26

What happens on the ROC curve when there is a ‘large loss due to FP’?

Answer 26

You move to the bottom left (because LR increases)

Question 27

What happens on the ROC curve when there is a ‘large loss due to missed diagnosis’?

Answer 27

You move to the top right (because LR decreases)

Question 28

What is ‘verification bias’ ?

Answer 28

Selective use of the reference test conditional (based) on the index test result

Question 29

What is ‘test-review bias’ ?

Answer 29

When the interpretation of the results of a test is influenced by knowledge of prior tests

Question 30

What is ‘uninterpretability bias’ ?

Answer 30

When tests that yield uninterpretable or inconclusive results are excluded from the analysis