AI in Pharma Quiz

Answers and explanations appear as you take the quiz.

1. A model ranks compounds for laboratory testing. What must be specified before interpreting its performance?
   • Only the model family name
   • The number of dashboard charts
   • The final commercial price
   • Intended use, population, endpoint, and decision boundary
2. Why is retrospective performance insufficient by itself for a clinical or regulatory claim?
   • Historical data are never useful
   • The validation may not match prospective context, population, or information timing
   • Retrospective models cannot be evaluated
   • A large test set guarantees approval
3. A biomarker model is evaluated on records from the same sites used for training. What transfer question remains?
   • Whether it generalizes to a new study site or data conditions
   • Whether the training loss reached zero
   • Whether it has a polished interface
   • Whether the feature names are alphabetized
4. Why inspect batch effects and missing measurements?
   • They only affect storage costs
   • Missing data always improve generalization
   • They can distort model inputs or create apparent patterns unrelated to the target
   • Batch labels are equivalent to outcomes
5. What does the guide say FDA/EMA AI principles do?
   • Approve every model that follows a checklist
   • Offer a framework for evidence and accountability, not blanket model approval
   • Replace context-of-use validation
   • Guarantee a regulatory submission succeeds
6. A model is used to prioritize exploratory lab work. How should that result be described?
   • As proof of clinical effectiveness
   • As a validated patient outcome
   • As a regulatory decision by itself
   • As an exploratory hypothesis unless evidence supports a stronger claim
7. When assay data or model versions change, what should the team do?
   • Assume the prior validation remains unchanged
   • Delete the previous protocol
   • Use the documented reassessment trigger and evaluate the new setup
   • Keep the old version in reporting only
8. Why retain protocols, model outputs, and review decisions?
   • They support traceability and later review of the evidence and decisions
   • They guarantee the model is unbiased
   • They eliminate monitoring
   • They are substitutes for data provenance