Quiz — Chapter 15: AI in Healthcare
Multiple Choice
1. What was the key flaw identified in the Obermeyer et al. (2019) study of a widely used healthcare algorithm?
a) The algorithm used race as a direct input variable to make predictions b) The algorithm used healthcare costs as a proxy for healthcare needs, which encoded existing racial disparities in spending c) The algorithm was never tested on real patients before deployment d) The algorithm was designed to be biased in order to reduce costs
2. What is the primary difference between FDA "clearance" and FDA "approval" for medical devices?
a) Clearance is for software; approval is for hardware b) Clearance requires clinical trial evidence; approval does not c) Clearance requires showing "substantial equivalence" to an existing device; approval requires clinical trial evidence d) There is no meaningful difference; the terms are interchangeable
3. Automation bias in a clinical setting refers to:
a) The tendency of AI systems to automate tasks that should remain manual b) The tendency of clinicians to over-rely on AI recommendations, even when their own judgment would be correct c) The tendency of hospitals to replace physicians with automated systems d) The tendency of patients to prefer AI diagnosis over human diagnosis
4. Why might an AI skin cancer detection system perform worse for patients with darker skin tones?
a) Darker skin makes cancer detection inherently impossible b) The training data likely contained far fewer images of skin conditions on darker skin c) The AI is programmed to prioritize lighter-skinned patients d) Skin cancer is rarer in patients with darker skin, so the AI ignores it
5. According to the Lancet Digital Health systematic review discussed in Section 15.2, what percentage of diagnostic AI studies were prospective (tested in real clinical settings)?
a) About 50 percent b) About 25 percent c) About 6 percent d) About 75 percent
6. Post-market surveillance refers to:
a) Monitoring AI systems before they receive regulatory clearance b) Systematic monitoring of how AI systems perform after real-world deployment c) Surveillance of patients using AI-enabled devices d) Market research conducted after a product launch
7. Which of the following best describes the "automation bias paradox" in healthcare AI?
a) The more accurate an AI system is on average, the more damage its remaining errors can cause because clinicians become less critical b) The more automated a system is, the more biased it becomes c) Physicians who trust AI are paradoxically less accurate than those who do not use AI at all d) AI systems become less accurate the longer they are deployed
8. The explainability problem in healthcare AI is an ethical concern because:
a) Patients might be frightened by technical explanations b) Physicians do not understand the AI systems they use c) A diagnosis that cannot be explained cannot be meaningfully questioned, undermining patient autonomy d) Explaining AI decisions would violate intellectual property laws
9. In the MedAssist AI case study, which of the following was NOT identified as a problem after real-world deployment?
a) Lower accuracy for patients with darker skin tones b) Physician over-reliance on the system c) Alert fatigue leading clinicians to dismiss warnings d) The system recommending treatments without physician review
10. Which of the following is the best example of AI addressing an administrative (rather than clinical) challenge in healthcare?
a) Detecting tumors in radiology images b) Predicting which patients are at risk of sepsis c) Automating medical coding and billing from clinical notes d) Screening retinal photographs for diabetic retinopathy
Short Answer
11. Explain what is meant by "the gap between lab and clinic" in healthcare AI. Provide at least three specific factors that can cause an AI system to perform differently in a real hospital compared to a controlled study.
12. A hospital administrator says: "Our AI diagnostic system has 94% accuracy, so it is clearly better than having no AI at all." Using concepts from this chapter, explain why this statement may be misleading. What additional information would you need to evaluate the claim?
13. Describe the ethical tension between beneficence and justice in the MedAssist AI case. How might a hospital weigh the benefits to some patients against the potential harms to others?
14. Compare the approaches of the United States (FDA) and the European Union (MDR + AI Act) to regulating healthcare AI. What is one strength and one weakness of each approach?
Applied Scenario
15. A startup has developed an AI system that analyzes voice recordings to detect early signs of Parkinson's disease. The system was trained on voice samples from 5,000 patients at three academic medical centers in the northeastern United States. In the company's study, the system achieved 91% sensitivity and 88% specificity.
a) Apply the three-question test from Section 15.2 to evaluate this claim. b) Identify three equity concerns that should be investigated before deploying this system widely. c) What regulatory pathway would this system likely follow in the United States? What are the limitations of that pathway? d) A rural health clinic in Guatemala is considering deploying this system because they have no neurologists. What additional validation would be needed? What ethical considerations does this raise?