What it actually is, stripped of the hype
You are reading a vascular surgery abstract. It says the authors "used artificial intelligence to predict AAA rupture risk." That sentence tells you almost nothing.
Did they use a logistic regression with a fancy label? A random forest? A deep neural network trained on CT images? The words "artificial intelligence" have become so overloaded that they obscure more than they reveal.
AI is not a method. It is a category -- a big one. Saying "we used AI" is like saying "we used statistics." The first question is always: what kind?
To make sense of any AI paper, you need two things: what kind of tool they used, and how that tool learned. Think of it like surgery -- you need to know both the procedure (bypass vs. stent) and the approach (open vs. endovascular).
This is not a straight line -- it is a tree with branches. AI is the trunk. It splits into different approaches, and those split again. Tap any box to learn more.
An EMR drug-interaction alert? AI. A self-driving car? AI. ChatGPT? AI. The term alone tells you nothing useful.
Most clinical decision support in your EMR. Technically AI, but a human programmed every rule.
You give the computer variables and outcomes; it finds the best way to connect them. You gain prediction accuracy but lose interpretability.
The workhorse of clinical ML. Spreadsheet of patient variables + outcome column = these tools.
Can take a raw CT scan or op note and learn features itself. Tradeoff: needs much more data, harder to explain.
DL on CT scans, ultrasound, or wound photos = almost certainly a CNN.
Extract info from notes, summarize literature. They do not run statistics.
You don't need to memorize these. Deep learning is a whole family.
AI is a tree, not a ladder. When someone says "we used machine learning," there is still a branch to identify (traditional ML or deep learning?). When they say "deep learning," there is still a branch (images? text? time-series?). Each branch has different data requirements, different strengths, and different failure modes.
This is a separate question from the tool hierarchy above. Any ML or deep learning model uses one of these three approaches. You will see these terms in every ML paper.
You give the computer both the inputs (patient data) and the correct answers (outcomes). It learns to predict the answer for new patients.
Analogy: Teaching a resident by showing them 1,000 cases where you already know the diagnosis. "Here are the features, here is what happened -- now predict the next one."
You give the computer the inputs but no answers. It finds patterns and groupings on its own. Useful for discovering patient subtypes or phenotypes you did not know existed.
Analogy: Handing a resident 1,000 patient charts with no diagnoses and asking, "Do you see any natural groupings?" That is clustering.
The computer learns by trial and error, getting rewarded for good decisions and penalized for bad ones. Shows up in treatment optimization and dosing algorithms.
Analogy: A resident learning to titrate a heparin drip -- not from a textbook, but by adjusting the dose, checking the result, and adjusting again over hundreds of patients.
When you read an AI paper, ask two questions: Where does it sit in the hierarchy? (basic ML on a spreadsheet, or deep learning on images/text?) and How did it learn? (supervised with labeled outcomes, or unsupervised looking for hidden patterns?). Most vascular surgery papers: basic ML, supervised, structured registry data.
The same technique gets called different things depending on who is talking. Here is the Rosetta Stone.
Red Flag: When a paper says "we used AI" or "we used machine learning" without specifying the exact algorithm, that is a reporting failure. You would not accept "we used statistics" without knowing if they ran a t-test or a Cox model. Apply the same standard.
When you read an abstract or hear a presentation that mentions AI or ML, the first question to ask is: "What specific algorithm did you use, and what type of data did it process?"
The answer tells you whether you are dealing with a random forest on structured data (familiar territory, just an extension of what you know), or a deep learning model on images (different validation requirements, different failure modes).
A realistic landscape of where AI is being used in your specialty, not where someone hopes it will be in 10 years.
ML models (usually random forest, XGBoost) trained on registry data (NSQIP, VQI) to predict outcomes: mortality, amputation, graft failure, readmission.
Typical AI layer: Machine Learning. Data: structured variables from EHR/registries.
Deep learning on CT angiography (AAA sizing, plaque characterization), duplex ultrasound (stenosis grading), wound photographs (classification, area measurement).
Typical AI layer: Deep Learning (CNNs). Data: medical images.
Extracting structured data from operative notes, radiology reports, clinic notes. Identifying complications, procedures, and anatomy from free text.
Typical AI layer: Deep Learning (transformers) or traditional NLP. Data: clinical text.
"AI that tells you what to do." Real-time surgical guidance, automated treatment planning, LLM-driven clinical reasoning.
Almost entirely aspirational. Regulatory, liability, and validation barriers are enormous. Do not confuse conference demos with clinical deployment.
If someone presents an "AI tool" for clinical use and cannot tell you which layer it sits in, what data it was trained on, and how it was validated -- that is not ready for your patients.
Before you use any of these tools with patient data: Feeding clinical data into AI models -- including commercial LLMs like ChatGPT or Claude -- requires a HIPAA-compliant environment and IRB approval. Most commercial AI tools are not HIPAA-compliant by default. De-identified data still requires IRB review. Your institution's data governance and compliance office should be involved before any patient data touches any AI system. This is not optional.
You have a clinical question. Which type of AI would you reach for to answer it?
Module 1 - Lesson 1 complete
The next time you have a clinical question involving data, ask: what type of data do I have, and which branch of the tree fits? That single step separates you from 90% of the audience at any AI talk.