Free Tool · Plan before you evaluate

How many cases per skin tone do you need before a performance claim means anything? Find out in 10 seconds.

A sensitivity figure for a skin-tone band with 12 positive cases has a confidence interval about ±20 points wide. No method can narrow it. This calculator shows the precision ceiling for each band, and the case counts needed for the claim you want to make.

🔒 Runs entirely in your browser. Nothing is sent anywhere.

1. Positive cases per skin-tone band in your evaluation set

Positive cases = confirmed cases of the condition your model detects (for example, biopsy-confirmed malignant lesions), not total images.

Fitzpatrick I–II (lightest)

Fitzpatrick III–IV

Fitzpatrick V–VI (darkest)

Assumed sensitivity

Images per patient (approx.)

Band	Positive cases	Precision ceiling (95%)	Can the audit make a claim?

Ceiling = 1.96 × √(p(1−p)·DEFF / n), in percentage points: the narrowest 95% interval any sensitivity estimate near p can have with n positives. Our audits make no per-band claim below 30 positives and report the ceiling instead.

2. How many positive cases per band do you need?

To state sensitivity within ±

To detect a gap between bands of

Reference-band sensitivity

Power

—

positives per band to state sensitivity within ±5 points

—

positives in each of two bands to detect a 10-point gap (97.5% two-sided, the level our audit uses for its two primary endpoints)

Gap detection uses the standard two-proportion sample-size formula with equal allocation. Repeated images per patient raise every number by the design effect.

Your darkest band is under the ceiling? That is the finding, and it is still worth reporting properly.

An independent audit states what the data can and cannot support, band by band, with a reproduction package your reviewers can run.

See engagement tiers