THE VR SCHOOL
  • Schools & Districts
    • License the academy
    • Free program for your students
    • Request a demo
  • Catalog
    • AI & Data
    • Cybersecurity & Web
    • Programming
    • Health Science
    • All Guilds
    • All courses
  • Families
  • Accreditation
  • Log In
  • Request a demo
  • Home
  • Courses

RLP AI Design B — Case Studies, Prototypes, Patents and Capstone

Curriculum

  • 4 Sections
  • 20 Lessons
  • Lifetime
Expand all sectionsCollapse all sections
  • Unit 1: AI in the Clinic: Case Studies from Past to Present
    5
    • 1.1
      Posters, Press Releases and Proof: How Health Technology Is Sold
      50 mins
    • 1.2
      Storyboarding a Talk: Sticky Notes, Props and the Stage
      100 mins
    • 1.3
      Talking to Machines: A Structured Written Debate on Mental Health Chatbots
      100 mins
    • 1.4
      Radiology, Prediction and Telemedicine: Writing the Talk
      150 mins
    • 1.5
      Portfolio Task — Case Talk: Success, Challenge or Failure
      200 mins
  • Unit 2: Building Small Models: Testing AI as Designers
    5
    • 2.1
      Artists and Machines: From AARON to Trained Collaborators
      50 mins
    • 2.2
      Train, Test, Tally: A Controlled Experiment on a Classifier
      100 mins
    • 2.3
      Ninety Percent Accurate? Base Rates and Honest Accuracy Graphics
      50 mins
    • 2.4
      Sequences and Anomalies: Coding a Heart-Rate Flag
      150 mins
    • 2.5
      Portfolio Task — Tested Prototype Exhibit
      200 mins
  • Unit 3: Engineering Design and the Patent Drawing
    5
    • 3.1
      The Art of the Patent Drawing and the Inventor’s Process
      50 mins
    • 3.2
      Drawing to Patent Standard: Views, Numerals and Line
      100 mins
    • 3.3
      Reading a Real Patent: Claims, Drawings and What Can Be Owned
      100 mins
    • 3.4
      From Sketch to Prototype: Engineering a Biomedical Design
      150 mins
    • 3.5
      Portfolio Task — Engineering Design Portfolio and Hypothetical Patent
      200 mins
  • Unit 4: Capstone: A Biomedical AI Innovation from Concept to Exhibition
    5
    • 4.1
      Showing the Future: Exhibitions, Pitches and the Capstone Launch
      50 mins
    • 4.2
      Paper Prototypes and Think-Aloud Testing
      100 mins
    • 4.3
      Money, Rules and Ethics: A Review Panel on the Record
      100 mins
    • 4.4
      Resume, Portfolio and Recorded Mock Interview
      150 mins
    • 4.5
      Portfolio Task — Capstone Exhibition: A Biomedical AI Innovation
      250 mins

Radiology, Prediction and Telemedicine: Writing the Talk

Unit 1  ·  Studio Production Workshop  ·  Lesson 4 of 20

Radiology, Prediction and Telemedicine: Writing the Talk

MA:Cr2A.IIIRLP-AIDB
Clinicians review an AI-prioritized scan remotely while a student plans a redesign presentation.
By the end of this lesson I can…

explain AI’s documented roles in radiology, predictive analytics, personalised medicine and telemedicine, and script and design a talk that moves from evidence to a redesign proposal.

Instruction

This workshop turns your case research into a script and a set of visuals. First, four areas you may draw on, each with a lesson for your talk.

Radiology. Radiology, the reading of medical images, is where most FDA-listed AI devices sit: the FDA’s list of AI-enabled devices is dominated by radiology tools that detect, measure or prioritize findings on X-ray, CT, MRI and mammography images. Many are designed to assist a radiologist, for example by moving a scan with a suspected bleed to the top of the reading list. Lesson for your talk: say exactly what the tool does (detect, measure, prioritize, or decide) because those are very different roles.

Predictive analytics: the sepsis case. Sepsis is the body’s extreme response to an infection and can be life-threatening. Hospitals want early warnings. A sepsis prediction model built into a widely used electronic health record was deployed in many U.S. hospitals. In 2021, researchers at the University of Michigan (Wong and colleagues, JAMA Internal Medicine) tested it independently on their own patients and found it identified only about a third of the sepsis cases and generated many alerts, a much weaker performance than the developer had reported. Lesson: a prediction that is not independently validated on the population where it is used may not work there, and too many alerts cause alert fatigue, where staff start ignoring them.

Personalised medicine. Also called precision medicine, this tailors prevention and treatment to a person’s genes, environment and lifestyle. The NIH’s All of Us Research Program is gathering health data from a large and diverse group of U.S. volunteers to support such research. AI methods help find patterns in these large datasets, and protein-structure prediction tools such as AlphaFold are used in drug research. Lesson: personalisation depends on who is in the data, which returns us to bias.

Telemedicine. Care delivered at a distance by video, phone or remote monitoring expanded rapidly during the COVID-19 pandemic. AI appears in triage tools, remote-monitoring alerts and transcription of visits. Missouri has a well-known example of the model: in 2015 the health system Mercy opened the Mercy Virtual Care Center in Chesterfield, a suburb of St. Louis, a $54 million building its officials called a “hospital without beds”, staffed by doctors, nurses and technology professionals who care for patients somewhere else. KFF Health News reported in 2021 that Mercy conducted an estimated 660,000 virtual visits during the pandemic, after a couple hundred in 2019. Lesson: technology that increases access for some can exclude people without reliable internet or devices.

Structuring the talk. Use this spine: hook (a concrete moment: a nurse retaking a retinal photo for the third time); the promise (what the system was designed to do); the evidence (what studies and deployment showed, with sources named aloud); the why (your analysis: data, label, deployment, people); the redesign (a specific change and how you would test it); the close (the question you leave with the audience). A redesign is specific: not “make it better” but “validate on each hospital’s own patients for three months before alerts go live, and cap alerts per nurse per shift”.

Script with visual cues. Write in two columns: what you say and what the audience sees. Speak in short sentences. One number per slide, large, with its source. If you choose the podcast format, your companion visual must carry the key evidence the audience cannot see while listening.

📖 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.

  • Radiology, Sepsis, Telemedicine: Numbers to Check and a Model First MinuteFDA, AI-Enabled Medical Devices list (retrieved 2026-09-26; counts are SLS counts of the rows); Wong et al., JAMA Intern Med 2021 (abstract); Telehealth.HHS.gov; CDC Sepsis quick facts; SLS original script model and computations · about 5 min

    Numbers you can compute for radiology, sepsis prediction and telemedicine, and a model first minute of a two-column script with its sources marked.

Vocabulary in context

  • predictive analytics — Using data and models to estimate the likelihood of a future event, such as a patient developing sepsis.
  • alert fatigue — When frequent alerts, many of them false, lead staff to ignore or override them.
  • precision medicine — Tailoring prevention and treatment to a person’s genes, environment and lifestyle.
  • telemedicine — Health care delivered at a distance through video, phone or remote monitoring.

Formative check

Work through these before moving on. They are not graded — they tell you whether the standard below has landed yet.

+50 XP

What did the 2021 independent study of the widely used sepsis model find?

Independent validation showed much weaker performance than reported.
+50 XP

Which is a specific redesign proposal?

A redesign names the change and how it answers the problem.
Myth or Fact?

Most devices on the FDA’s AI-enabled device list are used in radiology.

Radiology tools dominate the list; check the current version for the latest figures.

Paste the first minute of your two-column script (what you say / what the audience sees). Mark each claim with its source.

0 words
Quick self-check

How confident are you that you can explain AI’s roles in radiology, prediction, personalised medicine and telemedicine, and script a talk from evidence to redesign?

Not yetVery confident

Practice

Work these on paper or in your notebook, then open Check your answer. Aim for all of Fluency and Application; try at least one Challenge.

Printable version: this unit’s practice workbook (PDF)

Fluency

Build speed and accuracy with the core skill.

  1. Define predictive analytics, alert fatigue, precision medicine and telemedicine.
    Check your answer
    Answer: Predictive analytics: using data and models to estimate the likelihood of a future event. Alert fatigue: frequent alerts, many false, lead staff to ignore or override them. Precision medicine: tailoring prevention and treatment to a person’s genes, environment and lifestyle. Telemedicine: care delivered at a distance by video, phone or remote monitoring.
    Each comes with a lesson for your talk.
  2. List the six parts of the talk spine in order.
    Check your answer
    Answer: Hook, the promise, the evidence, the why, the redesign, the close.
    The hook is a concrete moment; the close is the question you leave with the audience.
  3. The lesson says to say exactly what a radiology tool does. Name the four roles it lists.
    Check your answer
    Answer: Detect, measure, prioritize, or decide.
    Many radiology AI tools assist a radiologist, for example by moving a scan with a suspected bleed to the top of the reading list (prioritize).

Application

Use the skill in context. Show your reasoning.

  1. The 2021 Michigan study found the sepsis model identified only about a third of sepsis cases. In an invented hospital with 300 sepsis cases, about how many would it identify and miss? What lesson does the case teach?
    Check your answer
    Answer: About 100 identified and 200 missed. A prediction not independently validated on the population where it is used may not work there, and too many alerts cause alert fatigue.
    300 / 3 = 100; 300 – 100 = 200. The study was independent (Wong and colleagues, JAMA Internal Medicine).
  2. KFF Health News reported an estimated 660,000 Mercy virtual visits during the pandemic, after ‘a couple hundred’ in 2019. Taking 200 as the 2019 figure, how many times larger is 660,000? What design rule does the lesson give for showing this number on a slide?
    Check your answer
    Answer: 3,300 times larger. One number per slide, large, with its source.
    660,000 / 200 = 3,300. Because ‘a couple hundred’ is approximate, say ‘more than 3,000 times’ or show both figures as reported.
  3. Turn this vague redesign into a specific one: ‘The telemedicine service should be fairer.’ Use the lesson’s warning about telemedicine.
    Check your answer
    Answer: Model: ‘Offer every video visit as a phone visit too, and track for three months how many patients without reliable internet or devices complete visits, before expanding.’
    A specific redesign names the change and how you would test it. The lesson’s warning: technology that increases access for some can exclude people without reliable internet or devices.
  4. Write the first two rows of a two-column script for a talk on the retinopathy case: what you say and what the audience sees.
    Check your answer
    Answer: Model: Row 1, Say: ‘A nurse retakes a retinal photo for the third time.’ See: a hand-drawn sketch of a dim clinic room and a camera. Row 2, Say: ‘In the lab, the system graded retinal photographs as well as eye specialists (Google, JAMA, 2016).’ See: one large phrase, ‘Lab result’, with the source in small type.
    Short sentences; sources named aloud; one idea or number per slide. The hook is a concrete moment.
  5. Why does the lesson connect personalised medicine back to bias? Name the program it mentions.
    Check your answer
    Answer: Personalisation depends on who is in the data; if groups are missing, the patterns found may not fit them. The NIH’s All of Us Research Program is gathering health data from a large and diverse group of U.S. volunteers.
    This returns to the Semester A bias checkpoint: who is in the training data, and who is missing.

Review

Keep earlier skills sharp.

  1. Review 1.3: Your podcast version of the talk has no slides. What must the companion visual do, according to 1.4, and which 1.2 rehearsal test checks the voice alone?
    Check your answer
    Answer: It must carry the key evidence the audience cannot see while listening. The picture-off test (listening with eyes closed) checks whether the voice carries the story.
    Both the visuals and the voice should work on their own.

Missouri Learning Standards — Media Arts addressed: MA:Cr2A.III, MA:Cr3A.II, MA:Pr4A.II, WHST.11-12.1, WHST.11-12.2, RST.11-12.1

Instructional focus: Scripting and designing an evidence-based media presentation

Reading

Radiology, Sepsis, Telemedicine: Numbers to Check and a Model First Minute

FDA, AI-Enabled Medical Devices list (retrieved 2026-09-26; counts are SLS counts of the rows); Wong et al., JAMA Intern Med 2021 (abstract); Telehealth.HHS.gov; CDC Sepsis quick facts; SLS original script model and computations · Public domain (FDA, HHS, CDC, NIH); facts from published studies cited with attribution; SLS original (computations, script model)

Why this reading. The workshop asks you to write a script that moves from evidence to a redesign, marking each claim with its source. This reading is the evidence for three of the four areas (radiology, prediction, telemedicine), each with numbers you can compute and check, and a model first minute so you can see what a marked script looks like.

A. Radiology and the FDA’s AI-enabled device list

The FDA publishes a list of AI-enabled medical devices authorized for marketing in the United States. The FDA says the list is not comprehensive: devices are included “primarily based on the use of AI-related terms” in their marketing authorization summaries. Retrieved September 26, 2026, the page listed 1,614 devices by SLS’s count of its rows, with final decision dates from 09/29/1995 to 06/29/2026. Counting the “Panel (Lead)” column, which names the FDA review panel:

Panel (Lead)DevicesShare of 1,614
Radiology1,23076.2%
Cardiovascular1549.5%
Neurology734.5%
Anesthesiology301.9%
Gastroenterology-Urology271.7%
Hematology221.4%
All other panels784.8%

Check: 1,230 divided by 1,614 is 0.762. Decisions by year: 2019, 80; 2020, 114; 2021, 129; 2022, 162; 2023, 226; 2024, 235; 2025, 335; 2026 through June 29, 181. Example rows: MammoScreen (5), Therapixel, K260714, Radiology, 06/26/2026; Saige-Dx, DeepHealth, K253825, Radiology, 06/15/2026; IDx-DR, IDx LLC, DEN180001, Ophthalmic, 04/11/2018. Note what the list does not say: the panel is the FDA review panel, not the tool’s job. Your script has to say the job, using one of four verbs: detect, measure, prioritize or decide. The FDA’s own release describes IDx-DR as giving a screening decision without a clinician reading the image, so its verb is decide (within a narrow intended use). A tool that moves a scan to the top of a reading list is prioritize.

B. Prediction: the sepsis model, with a confusion picture

From the Wong et al. abstract (JAMA Intern Med 2021;181:1065-1070): 38,455 hospitalizations; sepsis in 2,552 (7%); the model missed 1,709 (67%); alerts in 6,971 hospitalizations (18%); AUC 0.63. From these, SLS computes (assuming the alert and sepsis counts refer to the same hospitalizations): caught = 2,552 − 1,709 = 843, sensitivity 843 / 2,552 = 33.0%; alerts that were true = 843 out of 6,971, about 12.1%; so about 88% of alerts were not sepsis-onset alerts under this definition. The study ran from December 6, 2018 to October 20, 2019, which is 318 days, so 6,971 alerted hospitalizations is about 22 a day for one health system. That is what the glossary calls alert fatigue: 22 alerting patients a day, of whom about 3 were caught cases (843 / 318 = 2.7).

Redesign the lesson suggests, checked against the numbers. “Validate the model on each hospital’s own patients for three months before alerts go live, and cap alerts per nurse per shift.” The first half answers the finding that the model performed worse than reported when tested on a new hospital’s patients; the second half answers the 22-a-day alert load. A good script also says how you would test the redesign: for example, run the model silently for 90 days, compare its flags with what clinicians did, and go live only if sensitivity and alerts per day meet limits set beforehand.

C. Telemedicine

The U.S. Department of Health and Human Services site telehealth.hhs.gov says that telehealth, “sometimes called telemedicine,” lets patients see a health care provider without going to an office, and lists three modes: talking by phone or video, sending secure messages, and tracking health with technology so information such as blood pressure can be shared with a provider. Its listed benefits include less travel, faster appointments and more choice of provider, and it cautions that telehealth “may not be right for everyone or every health care condition.” For the equity point your script needs, use the retinopathy deployment study from Lesson 1.1: in the Thai clinics it studied, images sometimes took 60 to 90 seconds to upload, and one two-hour internet outage cut screening from 200 patients to 100. Reliable connectivity is a design requirement, not a given. (Sources: HHS Telehealth.HHS.gov, “Why use telehealth?”; Beede et al., CHI 2020.)

D. Personalized medicine

NIH describes its All of Us Research Program as an effort to gather health information from one million or more people living in the United States, with a priority on people and communities historically underrepresented in biomedical research, in order to make prevention and treatment more precise. The lesson to carry into a script: personalization depends on who is in the data. This reading does not host any claim about AlphaFold or other AI tools in drug research; leave such claims out of your script unless you have a source you can name.

A model first minute (SLS original)

What I sayWhat the audience seesSource I will name aloud
Every year, about 1.7 million U.S. adults develop sepsis, and at least 350,000 die in hospital or are discharged to hospice.One large number: 1.7 millionCDC, Sepsis: Basics (quick facts)
So hospitals bought a computer program to warn them early. When one hospital tested it on its own patients, it caught 843 of 2,552 cases.Two rows of dots: 843 filled, 1,709 emptyWong et al., JAMA Intern Med 2021
It also alerted on 18 percent of all patients: about 22 a day.One day’s list of 22 names, graySLS computation from Wong et al.

Word count of the “What I say” column: about 60 words, so roughly 26 seconds at 140 words per minute (60 / 140 x 60 = 25.7). Speaking-rate arithmetic for the whole talk: at 130 to 150 words per minute, a seven-minute talk is 910 to 1,050 words (130 x 7 = 910; 150 x 7 = 1,050).

Sources: FDA, Artificial Intelligence-Enabled Medical Devices (list) (fda.gov); Wong et al., External Validation of a Widely Implemented Proprietary Sepsis Prediction Model in Hospitalized Patients, JAMA Intern Med 2021 (doi.org); HHS Telehealth.HHS.gov, Why use telehealth? (telehealth.hhs.gov); CDC, About Sepsis (quick facts) (cdc.gov); NIH, All of Us Research Program (allofus.nih.gov)

Up NextComplete & Continue →

Leave a Reply Cancel reply

Talking to Machines: A Structured Written Debate on Mental Health Chatbots
Prev
Portfolio Task — Case Talk: Success, Challenge or Failure
Next

© 2026 The VR School | Career content by Wiley | Powered by Sejal Learning Systems.

Join free · Accreditation · Privacy Policy · Terms of Service

Modal title

Main Content