← Can routine measurements screen for heart disease?

Check your understanding

Six questions on the reasoning in this study, not on memorising its numbers. Every answer comes with an explanation.

You can take this quiz and see every explanation without an account. Sign in only if you want your progress saved.

Question 1 of 6

The majority-class baseline reaches 53.9% accuracy but 0.000 recall. What does that tell you?

Accuracy can look respectable while the model finds no positive cases at all The baseline is a reasonable screening tool The dataset is too imbalanced to model Recall and accuracy always agree

Question 2 of 6

The features ca and thal correlate with the target but were dropped from the served model. Why?

They are results of invasive tests, unavailable when a screening decision is made They correlated above 0.90 with the target, so the leakage audit rejected them They had too many missing values to impute They reduced the model's accuracy

Question 3 of 6

Logistic regression scored 0.886 recall and the random forest 0.857, with 95% intervals of [0.778, 0.974] and [0.733, 0.968]. Which model is better?

Neither — the intervals overlap, so the data cannot distinguish them Logistic regression, because 0.886 is higher than 0.857 Random forest, because it handles non-linearity Logistic regression, because its confidence interval is narrower

Question 4 of 6

The model outputs a calibrated probability of 0.79. What does that number mean?

Among referred cardiac patients in this sample with similar measurements, roughly 79% had disease This patient has a 79% chance of developing heart disease The model is 79% confident in its prediction 79% of people with these measurements in the general population have disease

Question 5 of 6

Why does the study apply Platt scaling even though the calibrated Brier score (0.117) is no better than the uncalibrated one (0.111)?

Balanced class weighting de-calibrates raw output, so correcting it is theoretically right even when the gain is not measurable at n=76 Platt scaling always improves probability estimates It was required to make the model beat the baseline Calibration is what makes the ROC AUC valid

Question 6 of 6

Recall is 0.714 for patients aged 56 and under, but 0.952 for those over 56. Why does the study flag this so prominently?

The model is least reliable exactly where a screening tool would need to be most reliable It proves the model is biased against older patients It means age should be removed from the feature set It shows the training data had too few older patients