The clinical team and the biostatistician write the study protocol and the statistical analysis plan. Together they fix the primary and secondary endpoints, the acceptance criteria, the sample size and its power calculation, the inclusion and exclusion criteria, and the case report forms readers will fill in. The protocol also defines how the reference standard is built and how reader disagreements are resolved. The draft usually goes to FDA as Pre-Sub questions before any data is read.
On the map: Protocol and statistical analysis plan. Endpoints, sample size, case report forms. Adjudication design.
Rows are disciplines, columns are stages 1 to 8. Select a square to open that task.
Back to the full map
Build the sample size from engineering's pilot error, and run endpoints and acceptance criteria past regulatory before the pre-sub. FDA counts only acceptance criteria set in advance.
Pick endpoints that match the output the label promises, such as agreement with the reader average for a measurement or sensitivity and specificity for a detection. Set limits from the predicate, the literature and pilot data, and keep secondary endpoints few, since every missed target has to be explained in the report.
Run a power calculation for each endpoint from an assumed error and spread, then take the largest number and add margin. Simulation is useful when the pass rule uses confidence intervals. The power number is a floor, since subgroup coverage often needs more cases than significance does.
Use an objective clinical anchor if one exists. Otherwise use three independent, blinded readers and combine them by a rule written in advance: median for measurements, majority for categories, STAPLE for masks. Escalate to an adjudicator only when a pre-set disagreement threshold is exceeded.
Ask about the parts that would be costly to redo: the reference standard, the separation of training and test data, and the sample size. Describe the algorithm clearly enough that FDA can judge how much proof the study needs.
See the open positions and apply.