10
Medium

Model accuracy

—

Misleadingly high

Null model accuracy

—

Flags nothing at all

True positives

—

Caught cases

False positives

—

Unnecessary referrals

False negatives

—

Missed cases

True negatives

—

Correctly cleared

AUC-ROC

—

Looks good — but misleads

PPV at this threshold

—

The real clinical picture

What AUC doesn't show you — alert breakdown at 90% recall:

True positives
—
False positives
—

PPV@90Recall

—

Precision at fixed sensitivity

Useful alerts

—

True positives per total alerts

True positives (caught)

—

Cases acted on correctly

False positives

—

Avoidable referrals

False negatives

—

Missed at 90% recall

True negatives

—

Correctly cleared