Medical AI · Foundations · Lesson 1 of 6

The healthcare-AI landscape: what is in use, what works, and what to ask

Reviewed by Dr. Ram Kumar G, MBBS, MD (Internal Medicine), MPH (USA), on 2026-10-18 · Last updated 2026-10-18

Beginner
Clinical focus
About 20 minutes

In short

Most AI in healthcare today is narrow: tools that do one task, such as reading an image, flagging a patient at risk, or drafting a note. By September 2026 the US FDA had authorised more than 1,600 AI-enabled medical devices, most of them in radiology. Large language models have added new uses, especially documentation, but strong performance in development often shrinks in real clinics, so every tool needs evidence from settings like the one it will be used in.

You will learn to

Distinguish AI, machine learning, deep learning and generative AI in plain terms
Name the main areas where AI is used in care, research and operations today
Explain, with real examples, why models that perform well in development can fail in practice
Apply a short set of questions to any clinical AI tool before trusting it

What 'AI' means in healthcare

'Artificial intelligence' is an umbrella term. In healthcare it usually means software that learns patterns from data rather than following rules written by hand. The terms below are nested: deep learning is a kind of machine learning, and most generative AI is built with deep learning.

The distinction matters because each kind fails in different ways. A rule-based alert fails when the rule is wrong; a machine-learning model fails when the data it meets differ from the data it learned from; a generative model can produce fluent text that is simply untrue.

Each row is a subset of AI; deep learning and generative AI are kinds of machine learning.
Term	What it is	Healthcare example
Rule-based decision support	Logic written by people: if X and Y, then alert	A drug-interaction warning in a prescribing system
Machine learning (ML)	Models that learn a mapping from inputs to an outcome from examples	A risk score for hospital readmission trained on past admissions
Deep learning	ML with many-layered neural networks; works well on images, signals and text	Detecting diabetic retinopathy on retinal photographs
Generative AI	Models that produce new text, images or audio; large language models (LLMs) are the best-known type	Drafting a clinic letter from a consultation recording

Where AI is used today

AI tools reach patients through many routes: as regulated medical devices, inside electronic health record (EHR) systems, as administrative software, and in research. The table gives the main areas. The evidence behind tools within each area varies widely, so the examples show what exists, not that it has been proven to help.

Area	Typical tasks	Notes
Medical imaging	Detecting findings on X-ray, CT, MRI, mammography and retinal images; prioritising worklists	The largest group of regulator-authorised AI devices
Screening	Diabetic retinopathy screening in primary care	The first FDA-authorised autonomous AI diagnostic system (2018) screens for diabetic retinopathy without a specialist reading the image
Risk prediction	Sepsis, deterioration, readmission, kidney disease progression	Often built into EHRs; performance frequently drops at new sites
Signals and monitoring	ECG interpretation, arrhythmia detection on wearables, continuous glucose monitoring analytics	Many are consumer or prescription devices with their own regulatory paths
Documentation and communication	Ambient note drafting, summarising records, drafting replies to patient messages	Mostly generative AI; clinicians remain responsible for the final text
Operations	Scheduling, bed management, coding and billing	Often not regulated as medical devices
Research	Trial recruitment, drug discovery, protein structure prediction	Indirect patient impact; evaluated with research standards

What the regulatory record shows

The US FDA publishes a list of AI-enabled medical devices it has authorised. As of September 2026 it contained more than 1,600 devices. Most are in radiology, and most reached the market through the 510(k) pathway, which relies on showing substantial equivalence to an existing device rather than on new clinical trials.

Authorisation means a device met the regulator's requirements for its stated intended use. It does not, on its own, show that the device improves patient outcomes in your setting. Later tracks cover how regulation works in the US, EU, UK and India, and how to read the evidence behind a device.

Dates matter in this field

Device counts, regulatory rules and model capabilities change every few months. Figures in this course are given 'as of' a date and re-checked at each review.

Three lessons from real deployments

Some of the most useful knowledge in healthcare AI comes from tools that did not perform as expected. Three widely cited examples:

Performance drops outside development. An external validation of a widely deployed proprietary sepsis prediction model across about 27,700 patients at one US health system found an AUROC of 0.63, far below the developer's reported figure, and the model missed two-thirds of sepsis cases while generating many alerts (Wong et al., JAMA Internal Medicine, 2021).
The label can encode bias. A commercial algorithm used to select US patients for extra care used healthcare costs as a stand-in for health needs. Because less money was spent on Black patients with the same level of illness, the algorithm systematically underestimated their needs (Obermeyer et al., Science, 2019).
The clinic is part of the system. A field study of a deep-learning retinopathy screening system in Thai primary-care clinics found that image quality, lighting, internet connectivity and clinic workflow, not the algorithm's accuracy alone, determined whether patients benefited (Beede et al., CHI 2020).

The common thread

A model's published accuracy describes the data it was tested on. Your patients, equipment, workflows and outcome definitions may differ, and any of these differences can change how it performs.

Generative AI and large language models

Large language models can answer questions, summarise records and draft documents. In 2023 a medically tuned model was shown to exceed the pass mark on US medical licensing exam-style questions, yet clinicians reviewing its long-form answers still found important gaps compared with physicians (Singhal et al., Nature, 2023). Passing an exam benchmark is not the same as safe performance on real clinical tasks.

LLMs can also state false information confidently ('hallucination'), and their outputs vary with small changes in wording. Their most established uses so far are in documentation and administration, where a clinician reviews every output. The Generative AI, RAG and Agentic AI tracks cover how these systems work and how to make and evaluate them safely.

Protect patient information

Never paste identifiable patient information into a public AI tool. Use only systems your organisation has approved for clinical data, under its data-protection rules.

Questions to ask about any clinical AI tool

Whether you are a clinician asked to use a tool, a manager asked to buy one or a developer building one, the same questions apply. The flowchart is a starting point, not a full evaluation framework; the Evaluation, safety & deployment track goes further.

A first-pass check before relying on a clinical AI tool

Write down the exact intended use: task, patients, setting, user
Has it been tested on data from outside the developer's own setting?
- If No
  1. Treat performance claims as unproven; ask for external validation
- If Yes
  1. Were those patients and settings similar to yours?
    If No or unclear
    Plan a local evaluation before use with patients
    If Yes
    Check regulatory status, oversight plan and monitoring, then pilot

General guidance for learning purposes; local governance and regulatory requirements always apply.

Key takeaways

Most healthcare AI is narrow: one task, one setting, one kind of data.
More than 1,600 AI-enabled devices were FDA-authorised by September 2026, most in radiology.
Development accuracy often falls in new settings; external validation is essential.
Generative AI adds new uses and new risks, especially confident errors; a person stays accountable.

Terms used: Dataset shift, External validation, Hallucination, Large language model (LLM), SaMD (Software as a Medical Device)

FAQ

Is AI replacing doctors?

Current tools perform specific tasks under clinical oversight. They change parts of clinical work, such as reading images or drafting notes, but responsibility for decisions stays with qualified professionals.

What is the difference between AI and machine learning?

AI is the broad field of making software perform tasks that normally need human intelligence. Machine learning is the main approach used today, in which models learn patterns from example data rather than following hand-written rules.

If a tool is FDA-cleared, does that mean it works?

It means the FDA found it met requirements for its stated intended use, often by showing equivalence to an existing device. It does not by itself show better patient outcomes, or that it will perform the same in your population.

Why do AI models perform worse in new hospitals?

Patients, equipment, documentation habits and outcome definitions differ between sites. A model learns the patterns of its training data, including ones that do not carry over. This is called dataset shift.

Can I use a public chatbot to summarise a patient's notes?

Not with identifiable information. Use only tools your organisation has approved for clinical data, and check every output, since language models can produce confident errors.

Sources

Educational content only. This material teaches methods; it is not clinical advice. Code, models and examples here are not validated medical devices and must not be used to make decisions about real patients without proper clinical validation, governance and any required regulatory approval. Examples use synthetic or openly licensed de-identified data only. Never paste identifiable patient information into public AI tools. Mentions of models, tools or vendors are for identification, not endorsement.