
deliver a TED-style talk or illustrated podcast that explains a documented AI healthcare case, evaluates it from several viewpoints and proposes a specific redesign.
Instruction
This is the unit’s public moment. You will record a talk on camera, or produce a podcast episode, and explain a real case of AI in healthcare: what it promised, what the evidence shows and what should change. The approved course names two formats, a TED-style talk and a podcast with a visual; choose the one that suits your case and your strengths.
Choosing your case. Choose from the unit’s documented cases (retinopathy screening from lab to clinic, the sepsis prediction model, Watson for Oncology, a radiology tool from the FDA list, robot-assisted or autonomous surgery, myoelectric prosthetics, a mental health chatbot) or propose another with at least four reliable sources. Classify your case honestly: success, challenge (works but struggles in deployment) or failure. Many cases are mixtures, and saying so is a sign of good analysis.
What the talk must include. What the system was designed to do; what independent evidence and real deployment showed, with sources named; why, analyzed through data, labels, deployment conditions and people; the viewpoints of at least patients, clinicians and developers; and a specific redesign with a way to test it. You may not invent statistics. If you cannot find a number, describe the finding in words and name the source.
Designing the talk as a media work. A TED-style talk is performance design. Plan your staging for the camera: where you stand in the frame, when you hold up your prop, when the screen goes dark so the viewer looks at you. Visuals are sparse and large: one idea, one image, one number. Pace at about 130-150 spoken words per minute, so a seven-minute talk is roughly 900-1,000 words. For a podcast, design sound: an opening sound or silence, a second voice only if a family member agrees to read a short part (otherwise your own voice, with clearly marked sections), clear transitions, and a companion visual (one-page graphic or short slide set) that carries the evidence.
The reflection. In about 700 words, explain your research process and your most important source; the limits of your evidence; the design choices you made for the talk or podcast and why; how your self-critique changed it; and what you would do with more time.
The four stages. Stage 1: research plan with the claim-evidence chart and storyboard photo. Stage 2: a full rehearsal, recorded, which you score yourself on the reading’s self-critique sheet after a day away, on accuracy, clarity, visuals and delivery (a family member or other adult may fill a second copy as an optional viewer); submit the sheet and your revision plan. Stage 3: the final talk or podcast with visuals and reflection. Stage 4: the final recording, followed by a recording of your answers to the course’s published panel questions, about three minutes in all.
📖 Read for this lesson
The reading opens over the lesson; close it to come back to exactly where you were. The checks below draw on it.
Formative check
Work through these before moving on. They are not graded — they tell you whether the standard below has landed yet.
At about 140 words per minute, roughly how long is a seven-minute talk script?
If you cannot find a statistic for your case, you may estimate one as long as it sounds reasonable.
How confident are you that you can design and deliver an evidence-based talk or podcast on an AI healthcare case with a specific redesign?
Unit 1 checkpoint
Do this before you start the task. It covers the whole unit and tells you what to review first.
Which feature is typical of the WPA health posters?
A study by a developer’s own researchers counts as independent validation.
In Thai clinics, many retinal images taken in poor lighting were rejected by the system as .
What happened to IBM’s Watson Health data and analytics business in 2022?
Which surgical system decides and performs steps itself?
The sound-off test checks whether the visuals carry the story without the voice.
The tendency to attribute understanding to a program that produces human-like text is called the effect.
Which design change most improves a mental health chatbot’s safety routes?
Frequent alerts, many of them false, that lead staff to ignore warnings cause:
‘Make the AI better’ is a specific redesign proposal.
In 2015 Mercy opened its Virtual Care Center in Chesterfield, Missouri, which its officials called a hospital without .
Submit: 6-8 minute recorded TED-style talk or illustrated podcast + 700-word reflection + recorded panel answers
Task: Explain one documented AI healthcare case: its design goal, the independent and deployment evidence, why it performed as it did, how patients, clinicians and developers see it, and a specific redesign with a way to test it. Use at least four reliable sources, invent no statistics and name sources aloud or on screen.
Submission workflow
- Stage 1 — Research plan. Case classification, claim-evidence chart with four or more sources, and storyboard photograph.
- Stage 2 — Self-critiqued rehearsal. Recorded full rehearsal, scored by you on the self-critique sheet after a day away for accuracy, clarity, visuals and delivery (an optional family member or other adult may add a second sheet); submit the sheet and revision plan.
- Stage 3 — Final work. Final talk visuals and script, or final podcast audio and companion visual, plus the 700-word reflection.
- Stage 4 — Delivery and questions. The final recorded talk or podcast, followed by your recorded answers to the course’s published panel questions (about three minutes).
How it is scored
| Criterion | What proficient looks like |
|---|---|
| Presentation design (MA:Pr6A.II) | Staging, visuals or sound design and pacing are designed for the audience and context and used effectively. |
| Production craft (MA:Cr3A.II) | Visuals, prop, script and delivery are deliberate and well executed. |
| Evidence and argument (WHST.11-12.1) | Claims are sourced and accurate; developer claims, independent evidence and deployment are distinguished; the redesign follows from the analysis. |
| Evaluation from several viewpoints (MA:Re9A.III) | Rigorous evaluation that weighs patient, clinician and developer perspectives and acknowledges limits. |
| Delivery and response (SL.11-12.4, SL.11-12.5) | Clear line of reasoning, well-integrated media and accurate answers to questions. |
Case Talk: The Full Brief, Criteria, an Exemplar Entry, the Self-Critique Sheet and the Panel Questions
SLS original assignment brief, criteria, exemplar, self-critique sheet and panel questions; figures in the exemplar are from Wong et al., JAMA Intern Med 2021 and CDC (retrieved 2026-09-26) · SLS original
Why this reading. The Portfolio Task names its criteria, stages and recorded answers to panel questions, but a student cannot start without the brief itself: the assignment in full, what each criterion looks like at four levels, a finished exemplar entry to measure your own against, the self-critique sheet you will use in Stage 2 and the published panel questions you answer on a recording in Stage 4. Nothing here depends on anyone but you choosing what to write.
The assignment
- Choose a case. Documented cases with sources hosted in this unit: retinopathy screening from lab to clinic (Reading 1.1), the sepsis prediction model (Readings 1.1 and 1.4), a mental health chatbot (Reading 1.3), radiology tools on the FDA list (Reading 1.4). Any other case needs at least four reliable sources you name in full.
- Classify it honestly as a success, a challenge (works but struggles in deployment) or a failure. Mixed answers are good analysis.
- Record a six- to eight-minute talk with visuals, or an illustrated podcast with a companion visual, that covers: the design goal; the independent and deployment evidence; why it performed as it did (data, labels, deployment conditions, people); the views of patients, clinicians and developers; and a specific redesign with a way to test it.
- Invent no statistics. If you cannot find a number, describe the finding in words and name the source, aloud or on screen.
- Write a reflection of about 700 words and record your answers to the five published panel questions below, about three minutes in all.
Criteria, at four levels (SLS original)
| Criterion | 1 Beginning | 2 Developing | 3 Proficient | 4 Exemplary |
|---|---|---|---|---|
| Presentation design | Text-heavy slides; no plan for where to stand or when | Some large visuals; staging unplanned | One idea per visual; screen goes dark when the speaker makes the key point | Staging, visuals and pacing are all designed for this audience and this case |
| Production craft | Script read aloud; visuals inconsistent | Script mostly spoken; some polish | Script rehearsed; type, color and images consistent | Prop or graphic carries a claim; rehearsal notes show revisions |
| Evidence and argument | Claims without sources | Some sources; developer claims and independent evidence blurred | Every number sourced; developer claim, independent evidence and deployment kept apart | Also says what the evidence cannot show, and the redesign follows from the analysis |
| Viewpoints | One viewpoint | Names patients, clinicians, developers | Explains what each would want and why they may disagree | Weighs the viewpoints and says where the case turns on them |
| Delivery and response | Reads from notes; cannot answer | Clear but inaccurate answers | Clear reasoning; accurate answers | Answers accurately, admits limits, and points to evidence |
Exemplar entry (SLS original; built only from the hosted evidence)
Case. Sepsis prediction model. Classification. Failure of independent validation. Claim-evidence chart, four sources:
| Claim I will make | Evidence | Source |
|---|---|---|
| Sepsis is common and deadly | About 1.7 million U.S. adults a year; at least 350,000 die in hospital or go to hospice | CDC, Sepsis quick facts |
| The model missed most cases at one hospital | It missed 1,709 of 2,552 (67%); sensitivity 33.0% | Wong et al., JAMA Intern Med 2021 |
| It created heavy alert load | Alerts on 6,971 of 38,455 hospitalizations (18%); about 22 a day | Wong et al.; SLS division 6,971 / 318 days |
| Independent testing is not routine before use | The authors say the model was implemented at hundreds of hospitals before adequate evaluation | Wong et al. abstract |
Viewpoints. A patient wants early warning and few false alarms. A nurse wants alerts that deserve attention. A developer wants the model to work across hospitals and needs local data to tune it. Redesign. Silent-mode validation for 90 days on the hospital’s own patients before any alert appears; an agreed limit on alerts per nurse per shift; publish the hospital’s own sensitivity and alerts per day. How I would test it. Compare silent-mode flags against clinician diagnoses; decide before the test what sensitivity and alert rate would count as good enough.
Exemplar timing. Hook 30 s; promise 45 s; evidence 2 min; why 1 min 30 s; viewpoints 1 min; redesign 1 min 15 s; close 30 s = 7 min 30 s. At 140 words per minute this is about 1,050 words (140 x 7.5).
Self-critique sheet for Stage 2
Record your full rehearsal, put it away for a day, then watch it and fill in this sheet. If a family member or other adult is willing to watch, they may fill in a second copy as an optional viewer.
| Question | Score 1-4 | Evidence: what I saw or heard |
|---|---|---|
| Accuracy: did every number have a named source, and could I trace one? | ||
| Clarity: could I state the case in one sentence afterward? | ||
| Visuals: did each visual carry one idea, readable on a phone screen? | ||
| Delivery: pace within 130-150 words per minute; looking at the camera; the key claim slowed down? |
Then answer in words: the single claim I could not verify; the one change that would help most; one thing not to change.
The published panel questions (Stage 4)
After your final recording, record your answers to these five questions, in order, in about three minutes in all. Name evidence or say plainly what is unknown.
- What is the single strongest piece of evidence in your talk, and who produced it?
- What does your evidence not show?
- Which viewpoint, patient, clinician or developer, would most disagree with your redesign, and why?
- How would you know whether your redesign worked?
- What new evidence would change how you classified the case?
Preparing for the panel questions
Write, in advance, the three hardest questions you expect beyond the five above (for the sepsis exemplar: “Isn’t 33 percent still better than nothing?”; “Who pays for silent-mode testing?”; “How do you know the hospital’s diagnosis codes are right?”), and a two-sentence answer to each that names evidence or admits what is unknown.
Reflection outline, about 700 words
| Section | Words |
|---|---|
| Research process and my most important source | 150 |
| Limits of my evidence | 150 |
| Design choices for the talk or podcast, and why | 150 |
| How my self-critique changed my work | 150 |
| What I would do with more time | 100 |
150 + 150 + 150 + 150 + 100 = 700.