AI in health

arqiv✳

AI in health

Can AI help a doctor spot what the eye misses?

AI is already reading X-rays and eye scans, predicting the shapes of proteins and searching for new antibiotics. It works best as a second set of eyes, and its mistakes can matter a lot, which makes health a great place to learn how to judge AI.

Try it yourself ↓Classroom activity →
A scan with an AI-highlighted area and a confidence score for a doctor to review AI FLAG“Take a closerlook here”confidence: 0.82 THENa doctor decides
In most hospitals, AI doesn’t diagnose on its own. It flags areas for a doctor to check, like a very fast second reader.

What AI is doing right now

Four ways AI is helping health

The U.S. Food and Drug Administration has now authorized well over 1,000 medical devices that use AI, most of them for medical imaging such as X-rays, CT and MRI.

01

Reading mammograms

In Sweden and much of Europe, two radiologists read every screening mammogram. AI can sort scans by risk and act as one of the readers.

Real exampleIn Sweden’s MASAI trial of more than 100,000 women, AI-supported screening found more cancers and cut radiologists’ reading workload by about 44%. The Lancet (2026)

02

Eye exams without a specialist

Diabetes can damage the eye’s blood vessels. Catching it early prevents blindness, but many people never see an eye doctor.

Real exampleIn 2018, the FDA authorized the first AI system that can detect diabetic eye disease without a specialist reading the image. npj Digital Medicine

03

Predicting protein shapes

Proteins do almost every job in your body, and their shape decides their job. Figuring out one shape used to take years.

Real exampleGoogle DeepMind’s AlphaFold predicted structures for about 200 million proteins. Its creators shared the 2024 Nobel Prize in Chemistry. Nobel Prize

04

Searching for new antibiotics

Bacteria are evolving to resist our medicines. AI can screen millions of molecules to find ones worth testing in the lab.

Real exampleMIT researchers used AI to discover halicin, a new antibiotic candidate that killed many drug-resistant bacteria in lab tests (Cell, 2020). MIT News

Try it yourself

Where should the AI draw the line?

100 people get screened. 10 are actually sick. Move the slider to decide how sure the AI must be before it flags someone for a doctor to check.

100 people screened, colored by what the AI decided
9sick people caught
1sick people missed
4healthy people flagged (false alarms)

A trade-off: 1 sick person missed and 4 false alarms. Real hospitals choose this line carefully, and a doctor reviews every flag.

● caught ● missed ● false alarm ● correctly cleared

The pattern behind every AI tool

Data → pattern → prediction

Every “AI in…” page follows the same five steps. Learn them once and you can understand AI in any field.

1 · Data

Thousands of labeled scans

Images where doctors already know the answer: sick or healthy, and where the problem is.

2 · Pattern

Shapes and textures

Disease often shows up as tiny changes in shape, brightness or texture.

3 · AI

A score for each image

The model gives each new scan a score for how likely something is wrong, and highlights where.

4 · Check

Doctors and clinical trials

Doctors review flags, and large trials test whether patients actually do better.

5 · Use

Faster, wider screening

Busy clinics can screen more people and get them to specialists sooner.

Think like a scientist

What AI can’t do here

Knowing the limits is part of understanding the tool.

It can be unfair

If an AI was trained mostly on one group of people, it may work worse for others, such as different ages, skin tones or body types.

It only knows its own job

A model trained on chest X-rays from one kind of machine may stumble on images from another hospital’s machine.

False alarms have real costs

Every false alarm means worry, extra tests and sometimes procedures for a healthy person.

Privacy matters

Health data is personal. Using it to train AI requires strong protections and permission.

Where people stay in charge: in most cases, doctors make the diagnosis and the treatment decision, talk with the patient, and stay responsible. The FDA reviews AI that counts as a medical device, such as tools that read scans. Some other AI software used in hospitals doesn’t need FDA review, so hospitals have to check it themselves.

Careers

Jobs where people work with this AI

AI is creating new roles in health, and changing old ones.

Radiologist

A doctor who reads X-rays, CT scans and MRIs.

Working with AIUses AI to sort urgent scans first and double-check for things that are easy to miss.

Clinical data scientist

Builds and tests models using health records and images.

Working with AIChecks that AI works fairly for every group of patients.

Biomedical engineer

Designs medical devices, from scanners to wearable sensors.

Working with AIBuilds AI into devices like heart monitors and glucose sensors.

Health AI ethics & safety specialist

Makes sure AI tools are safe, private and fair before and after use.

Working with AIAudits AI decisions and tracks mistakes in real hospitals.

Classroom activity

Where should the line be?

45 minutesGrades 7–12One computer with a webcam per groupFree

Students use the screening simulator to explore trade-offs, then train their own simple image classifier and discover how it can be biased.

Teacher tip: bring two kinds of objects for each group to sort (for example, 2–3 ripe and 2–3 green bananas) and a sheet of white paper to use as a plain background.

  1. Explore (10 min): In groups, use the simulator above. Find the setting your group thinks is best, and write down why.
  2. Debate (10 min): Would you choose a different line for a cancer screening than for a strep throat test? Why?
  3. Build (15 min): Use Google’s free Teachable Machine to train an image model on two classes (for example, ripe vs. green bananas). Use the webcam to capture about 30 images of each class, all against the same plain background.
  4. Test for bias (10 min): Move the objects to a different background (a patterned desk, a colored sheet) or change the lighting, and test again. Did it get worse? Connect this to why medical AI must be trained on many kinds of patients.

Discussion questions

  • Who should decide where the line is: doctors, patients, hospitals or the company that made the AI?
  • Is it fair to use an AI that works better for some people than others, if it still helps everyone a little?
  • Would you want to know if AI read your scan?

Fresh from the lab

Latest research in health

New studies are added every week and explained in plain language. Not every study uses AI, so ask yourself where AI could help. See all research →

New stories in this area are on the way. Browse all research →

Sources for this page

Put curiosity to work

Read it. Try it. Question it.

Explore new AI research, or take a paper-first investigation into your classroom.

Classroom starter pack →Research explained →