Lesson 1 of 6
Module 3 · AI in practice
From the first test to the last prescription, AI is part of care in some places and still research in others. This lesson tours the main uses, and what has been learned from them.
Module 2 gave you a way to judge any model: the right measures at the right prevalence, and the rung of evidence they come from. Module 3 puts that to work on the tools hospitals use today. This lesson surveys where AI appears along a patient's journey and how far the evidence for each area has climbed; the lessons after it cover large language models, deployment, safety and regulation, and everyday use of AI assistants.
Reading the journey
The five stages follow a patient from first contact to ongoing care, and the label on the right of each is the kind of data from Health data: the raw material that the AI mostly uses there. Two points are worth noticing.
- Screening and diagnosis differ. Screening tests large numbers of people who feel well, so the disease is rare and, as Module 2 showed, false positives are common. Diagnosis concerns a patient who already has symptoms, so the disease is far more likely and the same tool behaves differently.
- Not all uses are clinical. Administration, such as drafting letters or clinical coding (turning a record into the standard codes used for statistics and payment), may not touch a diagnosis at all, but errors there still reach patients' records.
As you read the sections below, keep the evidence ladder from Evidence: from validation to trials in mind. The areas differ less in how clever their algorithms are than in how far up that ladder their evidence has reached, and in how directly an error would reach a patient.
Imaging: the most mature area
Most regulator-cleared AI tools read medical images, and radiology leads by a wide margin. In 2018 the US Food and Drug Administration authorised the first system allowed to give a diabetic retinopathy screening result without a specialist reading the image. Tools that triage scans, highlight suspected fractures or flag findings on mammograms are now common.
A tool is "regulator-cleared" when a medicines and devices regulator has agreed that it may be sold for a stated use, on the evidence the maker supplied. Imaging led for practical reasons. Images are stored in one standard format, DICOM, in large digital archives. Most come with a written report that can serve as a label. And many imaging tasks are tightly defined: is there a fracture on this wrist X-ray, yes or no. Deep learning, which learns directly from pixels, suits such tasks well.
To triage scans means to sort them by urgency, so that a scan likely to show something serious is read first. The usual mistake is to read "cleared" as "proven to help patients". Clearance is about safety and performance for the intended use; it rarely means a trial has shown better outcomes, the top rung of the ladder in Module 2.
Imaging tools are also used in different roles, and the role changes what the evidence must show. A second reader looks at images after the clinician and flags anything it thinks was missed. A concurrent tool shows its marks while the clinician is reading. A triage tool changes the reading order without giving a finding. A few tools work autonomously, giving a result with no specialist reading the image, as in the retinal screening system above. The further a tool moves towards working alone, the stronger the evidence of its own accuracy needs to be, because there is less human checking to catch its mistakes.
A pitfall with concurrent tools is that people tend to accept a machine's suggestion too readily, a habit called automation bias that Deploying AI in a hospital looks at closely. If the tool marks a shadow, a reader may agree too readily; if it marks nothing, a reader may look less carefully. A study that measures the tool alone does not show this effect. Only a study of clinicians reading with and without the tool can.
Reading the figure. An ECG trace shows the heart's electrical activity over time, one repeating wave per heartbeat. A cardiologist reads the shape and timing of each wave; a deep learning model is shown many thousands of labelled traces and learns which shapes go with which diagnosis, for example an irregular rhythm such as atrial fibrillation. Because every beat repeats, even a short recording gives the model many examples to learn from.
Signals also show how a model can find something a specialist does not look for. Research models have been trained to estimate, from an ordinary ECG, features that are normally measured another way, such as how well the heart's main pumping chamber is working, which usually needs a heart scan. Findings like this are promising, but they still have to climb the ladder: the step that matters is a study showing that acting on the result leads to earlier diagnosis and better care.
Early warning: a cautionary tale
When one widely used sepsis model was independently evaluated in 2021, it performed much worse than its developer had reported, missing many cases while raising frequent alerts. Too many alerts cause alert fatigue: staff start to ignore them, including the real ones.
Sepsis is a life-threatening reaction to infection, in which the body's response starts to damage its own organs. Early treatment matters, so automatic warning tools are attractive. They watch the record for signs such as a rising heart rate, low blood pressure, fever and abnormal blood results, and raise an alert when a score passes a threshold.
Why might an independent evaluation find worse results than the developer? Several reasons from this course can combine. The patients and recording habits may differ from the development data (dataset shift). Sepsis may be defined differently in the two studies, so the labels differ. And the model may be judged on whether it warns before clinicians have already noticed, which is harder than spotting sepsis that is already being treated. Alert fatigue then compounds the problem: if nine in ten alerts turn out to be false, staff learn, reasonably, that alerts are usually wrong.
The label problem from Health data: the raw material is especially sharp here. Sepsis has no single test. Studies define it using different combinations of diagnosis codes, antibiotic orders, blood cultures and organ-failure scores, and each definition picks out a somewhat different group of patients. A model trained to predict one definition and judged against another will look worse than it is, or better. When comparing sepsis tools, check that the definitions match before comparing the numbers.
What would a careful hospital do before switching such a tool on? The steps from Module 2 apply directly: run it silently on local patients, measure sensitivity and positive predictive value at the planned threshold, count the alerts it would create each day, and agree in advance what staff should do when it fires. A tool whose alerts lead to no clear action adds workload without adding benefit.
The evaluation in this section is Wong and colleagues (JAMA Internal Medicine, 2021), who studied the Epic Sepsis Model across about 38,000 hospital stays at Michigan Medicine. The model had an AUC of about 0.63, against 0.76 to 0.83 reported by the developer. At the alert threshold the hospital used, it missed about two-thirds of sepsis cases and alerted on about 18% of all hospitalised patients. When citing it, note that it describes one health system, one version of the model and one sepsis definition. The developer disputed parts of the method. The study shows why independent, local validation matters, not that all sepsis tools fail.
Discovery: proteins and drugs
AlphaFold, a deep learning system that predicts a protein's 3D structure from its sequence, changed structural biology; two of its developers shared the 2024 Nobel Prize in Chemistry, with David Baker, honoured for designing new proteins. AI now helps propose drug targets, screen molecules and design proteins, though every candidate must still prove itself in the lab and in trials.
A protein is a chain of amino acids that folds into a precise three-dimensional shape, and its shape largely decides what it does: which molecules it binds, and where a drug could attach. Working out a structure in the laboratory can take months or years, so a reliable prediction from the sequence alone saves a great deal of work. A drug target is the molecule in the body, often a protein, that a medicine is designed to act on. Most candidate molecules, however they were found, fail on the way to approval, so AI speeds up the early search but does not remove the need for laboratory work and clinical trials.
A second example is virtual screening: instead of testing hundreds of thousands of molecules in the laboratory, a model scores them on a computer for how likely they are to bind a target, and only the most promising few hundred are made and tested. This can save years of work at the start of a project. The pitfall is to judge such a method only on old benchmark datasets, where the answers may already have influenced how the model was built. The real test is whether the molecules it picks work when they are made and measured, and later, whether medicines found this way succeed in trials more often than others, which is not yet clear.
Go deeper: what a predicted structure can and cannot tell you
AlphaFold gives each part of a predicted structure a confidence score from 0 to 100, called pLDDT. High scores usually mean the shape of that region is reliable. Low scores often mark regions that have no fixed shape in the body at all, so a low score is information, not just a failure.
There are also limits that a good score does not remove. The system predicts one static shape, while many proteins change shape as they work. A confident structure does not say how strongly a drug will bind, whether it will reach its target in the body, or whether a single mutation will change the protein's behaviour. A public database, expanded in 2022, holds predictions for more than 200 million proteins, which makes it a starting point for experiments, not a replacement for them.
Precision oncology
Models combine tumour genomics, pathology images and clinical data to predict which patients will respond to a therapy. It is one of the fastest-moving areas of medical AI research.
Precision oncology means choosing cancer treatment from the features of each patient's tumour, not only from where in the body it started. Some targeted drugs work only in tumours that carry a particular genetic change, so finding that change decides whether the drug is offered. AI can help by combining many weak signals, from DNA, from the appearance of the tumour under the microscope and from the patient's history, into one prediction. Most such models are still research tools rather than part of routine care.
One active line of research uses the routine stained slide that every tumour biopsy already produces. Models have been trained to predict, from the slide image alone, some genetic features of the tumour that are normally found by a separate molecular test, such as microsatellite instability, a pattern of DNA repair failure that affects which treatments are likely to work. The appeal is speed and cost. The caution is that the molecular test remains the reference standard. A prediction from an image can help decide who is tested first, but it is not yet a substitute for the test itself, and its performance must be checked at each laboratory, whose staining and scanners differ.
Precision oncology also shows the problems of the omics row in the data table of Module 1: very many measurements for relatively few patients, which invites the overfitting described in How a medical AI model is built. External validation in independent groups of patients matters even more here than in imaging.
Worked example · triage time saved
A radiology department reads 1,000 chest X-rays a day. An AI tool moves the 5% it flags as urgent to the top of the list.
if each would otherwise wait 4 hours on average
→ up to 200 hours of waiting removed per day for those patients
The benefit depends on the flagged scans being the right ones, and on the remaining scans not waiting longer. Both must be checked locally. The numbers are for illustration.
Step by step. 5% of 1,000 is 0.05 × 1,000 = 50 scans flagged each day. If each of those would have waited 4 hours, the total wait for the flagged group was 50 × 4 = 200 hours. Reading them at once removes up to all of it, which is why the result says "up to": in practice even a flagged scan waits a little.
What the answer means. The radiologists' reading time has not changed; the order has. The 950 unflagged scans now each sit behind the flagged ones, so their wait grows a little. The trade is worth making only if the flagged scans really are the urgent ones. If the tool flags many normal scans and misses some serious ones, it moves the wrong patients forward and may delay the ones who most need a quick report.
Key point
Success depends less on the algorithm than on the setting: a clear task, good data, evaluation in real practice, and a workflow in which staff know what to do with the output.
A second example: picture a retinal screening tool that works well in a programme with good cameras, trained photographers and a fast referral route to eye clinics. In a setting where many photos are blurred and referred patients wait months to be seen, the same software will meet many more images it cannot grade, and its positive results will lead to less benefit. The algorithm is identical; the setting is not.
The next lesson turns to the newest group of tools on the administration row of the journey: large language models, which draft and summarise text and which many staff now meet every day.
Further reading · Module 3Wong et al. 2021: a widely used sepsis model that performed much worse when tested independently. · All reading for this course
Check your understanding
1. Which area has the most regulator-cleared AI tools?
2. What is alert fatigue?
3. What does AlphaFold predict?
Up next
Lesson 2
Large language models in medicine