Specialty centers · Clinical AI & Diagnostic Bias
Free health-science briefs that teach students to read an algorithm the way they would read a lab result. Students audit a published diagnostic model, convert its reported accuracy into predictive values at a real prevalence, and write the one-page clinical caution sheet the paper did not. CLIN-04 and CLIN-01 have single-brief PDFs; every brief is in the bank PDF; the other ten are full on the page. No account. No student data leaves your school.
The PHI rule
No protected health information enters this track — not from a placement, not from a shadowing day, not de-identified. Students work from published metrics, teacher-approved public datasets, and clearly labeled synthetic cases. A student who cannot get data does not invent it: the access barrier is written up as a finding.
How to run one
Worked sample · printable PDF
Pick one published diagnostic or clinical risk model. Can you (1) reproduce or re-derive the headline performance the authors report, (2) convert it into positive and negative predictive values at a prevalence a clinician would actually see, (3) document where subgroup performance is unequal or simply absent, and (4) write the one-page caution sheet the paper did not — without claiming you validated anything?
Featured · printable PDF
Holding the operating point completely fixed, can you show what happens to its positive predictive value as prevalence moves across the settings that test would actually be used in — and then write the explanation a patient or a first-year nursing student would understand on the first read?
Take one model with public scores or a published ROC, rebuild the confusion matrix at two operating points, and defend which point belongs in a screening use and which belongs in a confirmatory use. Whose error did you choose to accept?
Same test, same threshold, two Virginia clinics with genuinely different prevalence. Using public Virginia health data, recalculate predictive values for both settings — and say plainly which clinic should not be using this result to rule a condition in.
On a public clinical dataset, define two defensible labeling rules for the same outcome — a stricter and a looser one — and measure how much your conclusions move when only the definition changes. Does the finding survive the definition?
In a public dataset, identify one variable that plausibly proxies for race, socioeconomic status, or access to care — then show, numerically, how it reaches the predictions. What is the model actually learning?
An imaging model can post excellent numbers by reading something that is not the disease — a marker, a drain, a laterality token, a scanner signature, the fact that sicker patients get portable films. Take one documented shortcut, re-analyze or reproduce the evidence for it, and propose the check that would have caught it before anyone trusted the model.
Where does the score add information, where does it merely add steps, and at what decision threshold does the answer change?
A public clinical dataset exists because patients were treated, not because they volunteered for your project. Write the consent and data-use analysis for training on it as if your health system were the steward deciding whether to release it — and say what you would refuse to release.
Take a model whose limits are already understood — your own CLIN-04 audit, a classmate's, or a published model card — and produce the one page a preceptor would actually post. Then test it on real readers and revise until it survives them. The deliverable is not the analysis; it is the communication artifact.
What gets substituted, what gets synthesized, what stays hypothetical, and what simply does not get done?
Then evaluate a model that meets its headline target and fails your bar, and write the recommendation against deployment. Can you hold a line you drew before you knew what it would cost you?
Six-dimension rubric
Scope unclear? Clinical language off? Tell us here — you do not have to leave an email.