AIQ
AIQ
Part of the AI in Healthcare lesson guide. Teaching a different grade? 🌈 Explorer (5–7) · 🔧 Builder (8–10) · 💻 Hacker (11–14)
Architect mode has no mascot and treats students as near-adults preparing for the real world, including possible careers in this exact field. Open at that level.
Say: "AI is diagnosing diseases, discovering drugs, and predicting protein structures. Let's examine how AI is transforming healthcare — and the risks involved."
Ask directly: "If you were a hospital administrator deciding whether to deploy a new AI diagnostic tool, what would you need to see before you approved it — and who would you hold accountable if it made a mistake?" Don't answer it yet; let it sit as the question the lesson builds toward. This age band should leave with a real opinion on regulation and accountability, not just a list of cool applications.
If your class includes students considering medicine, computer science, biomedical engineering, or health policy as career paths, it's worth naming that explicitly here: the intersection this lesson covers — AI, healthcare, and regulation — is a real and growing field with career paths in it right now, from clinical informatics to regulatory science to ML-focused biotech research. This isn't hypothetical future content; it's a live area they could be working in within the next decade.
Have students work through the three scenes independently, then run the debrief as a structured discussion — at this level, focus less on "what does this tool do" and more on "what would it take to trust this tool in practice."
Scene 1 — Diagnosis. The lesson's examples: an X-ray model detecting lung cancer at an accuracy the lesson describes as sometimes matching radiologists; a retina-scan model (built by Google) screening for diabetic blindness, intended for regions with limited access to ophthalmologists; and a melanoma classifier trained on 130,000+ labeled images performing comparably to dermatologists. Discuss: sensitivity/specificity tradeoffs are the real engineering question behind "accuracy" — a model tuned to catch every cancer case will also flag more false positives, and clinical deployment has to choose where on that tradeoff curve to sit, a decision with real cost and anxiety implications for patients. This is also where FDA/regulatory clearance becomes relevant: these tools are classified as medical devices, requiring validation studies before clinical use, not simply an app-store release.
It's also worth naming a subtlety that separates a strong classroom claim from a rigorous one: "accuracy" alone is a limited way to describe a diagnostic tool's performance, because it doesn't distinguish between the different ways a model can be wrong. A model that never misses a real cancer case but also flags many healthy scans as suspicious behaves very differently in practice from one that misses more real cases but rarely raises a false alarm — the first creates unnecessary patient anxiety and follow-up procedures, the second risks missed diagnoses. Clinical validation studies report both directions of error specifically because a single accuracy percentage can hide which failure mode a tool leans toward, and that failure mode matters enormously for how the tool should be used in practice.
A strong follow-up question for a class with statistics background: "If a disease is rare — say it affects 1 in 1,000 people — why might a test that is '99% accurate' still produce more false alarms than true positives?" This is a real, well-known statistical trap (related to Bayes' theorem and base rates) that applies directly to medical screening tools, and it's worth walking through with actual numbers if your class has the math background: even a 99%-accurate test applied to a rare condition can flag far more healthy people than sick ones in absolute terms, simply because there are so many more healthy people being tested.
Scene 2 — Treatment. Covers AI-assisted antibiotic discovery (100 million compounds screened computationally in a widely reported 2020 result), AI-assisted design of tumor-DNA-personalized cancer treatment, and AI-assisted (not autonomous) surgical robotics. Worth naming directly for this age group: AlphaFold, DeepMind's protein-structure-prediction system, is one of the most significant results in this space — it predicted structures for over 200 million proteins, addressing a roughly 50-year open problem in structural biology (the "protein folding problem": predicting a protein's 3D shape from its amino-acid sequence). That result accelerates early-stage drug discovery by giving researchers structural information they previously had to determine experimentally, often over months or years, for a fraction of the compounds now covered. It does not itself discover or approve a drug — it removes one major bottleneck earlier in the pipeline.
It's worth being precise about what "predicting a protein structure" means, since students may have heard the term without a clear definition: a protein is a long chain of amino acids, and that chain folds into a specific 3D shape that determines what the protein does in the body. Before this technology, determining that shape experimentally — through techniques like X-ray crystallography — could take a research lab months or years per protein. A computational model that predicts the folded shape from the amino-acid sequence alone turns a slow, expensive experimental problem into a fast computational one for a huge fraction of proteins, which is why it mattered so much to structural biology and, downstream, to drug design — a drug often works by binding to a specific shape on a target protein, so knowing that shape accurately is foundational to designing one.
Scene 3 — Patient Care. Covers AI-based arrhythmia detection in wearables, an AI chatbot (Woebot) offering CBT-style techniques, and AI risk-prediction models trained on large health-record datasets. This is the natural place to raise two real, ongoing concerns: algorithmic bias (a model's performance is bounded by how representative its training data is — a model trained predominantly on one demographic can underperform on others, a documented issue in published medical-AI research) and patient privacy (training and running these models requires large volumes of sensitive health data, raising questions about consent, data ownership, and secure handling — see the quiz's federated-learning question for one real technical response to that problem).
Treat this as a seminar-style discussion rather than a question-and-answer drill — these questions are meant to surface genuine disagreement and are close to the kind of case-study discussion used in actual health-policy and bioethics courses. It's fine, and often more valuable, if the class doesn't reach consensus.
This quiz assumes the class has engaged with the regulatory, scientific, and ethical threads raised in the activity above — each question maps directly onto one of the four scene discussions. If a student can answer correctly without being able to explain the underlying concept (what a validation study actually checks, what protein folding actually is, why homogeneous training data is a problem, how federated learning technically works), treat that as a cue to revisit the relevant scene discussion rather than moving on.
Timing note: the in-app content plus the scene discussions above, run at the depth this age group can engage with, typically takes 25–30 minutes; the policy-memo extension is substantial enough to assign as homework or a second class period rather than compressing it into the same sitting. If you're short on time, the memo can be shortened to a single required paragraph covering just the accountability question, which is the hardest and most valuable part of the assignment.
Close by returning to the opening question: "So — what would you need to see before approving an AI diagnostic tool for your hospital? At minimum: regulatory clearance backed by real validation studies, transparency about what population the training data represents, and a clear answer to who is accountable when it's wrong. That last one doesn't have a clean legal answer yet in most places — which is exactly why this field needs people who understand both the technology and the ethics, not just one or the other."
Extension activity (25–30 minutes, suitable for homework or a full period): Have students write a one-page policy memo, as if advising a Philippine hospital or the Department of Health, recommending conditions under which an AI diagnostic tool should be approved for clinical use. It must address at minimum: what evidence of accuracy should be required, how the tool's performance should be checked across different patient groups, and who bears responsibility for an error. This connects the lesson directly to real regulatory questions Philippine and international health systems are actively working through.
If assigning the memo as homework, a useful rubric anchor is to ask students to reference at least one specific concept from this lesson in their reasoning — validation studies, sensitivity/specificity tradeoffs, training-data representativeness, or federated learning — rather than writing generic AI-caution language that could apply to any technology. The goal is for students to demonstrate they can apply the lesson's specific mechanisms to a policy question, not just express a general opinion that "AI should be regulated carefully."
For a class with more time or a strong interest in the topic, a good companion activity is a short structured debate: assign one group to argue that hospitals should be required to disclose to patients whenever an AI tool contributed to their diagnosis or treatment plan, and another group to argue the disclosure requirement is unnecessary friction that doesn't change patient outcomes. This surfaces informed-consent questions that sit right alongside the accuracy and bias questions already covered, and it's a natural bridge if your school later covers AI ethics or health policy in a separate unit.