Table of Contents
Nearly a thousand AI-enabled medical devices have been cleared by the FDA, and roughly three-quarters of them sit in radiology. That lopsided number is a useful clue about how this technology actually succeeds. It works best where the data is already digital, the question is narrow, and a trained human can check the answer before anything irreversible happens.
The first wave landed in radiology
A CT scan is, at bottom, a grid of numbers. Asking a model to outline a lung nodule or flag a suspected bleed on a brain scan is a well-defined task with a verifiable answer. Companies like Aidoc and Viz.ai built businesses on exactly that narrowness. Their software does not replace the read. It sits in the background and pushes a priority alert to a specialist’s phone when something looks urgent.
Where the minutes are actually saved
The gains rarely come from the reading itself. A radiologist takes seconds to glance at a chest X-ray. The time is won in the queue. A suspected large-vessel stroke that might have waited forty minutes for an on-call neuroradiologist gets flagged in two or three. Multiply that across a busy night shift and you understand why stroke networks were among the earliest buyers.
Diagnosis moves from reactive to predictive
In 2018 the FDA cleared IDx-DR, the first device allowed to diagnose a medical condition without a clinician interpreting the image. It reads retinal photographs for diabetic retinopathy and returns one of two answers: refer, or come back in twelve months. The setting mattered more than the algorithm. Those cameras went into primary care clinics where many patients with diabetes had never had a proper eye exam, simply because there was no ophthalmologist to send them to.
Sepsis is the other headline case. The TREWS system at Johns Hopkins, evaluated in a 2022 study, watched patients whose vitals were drifting toward sepsis and reported an 18.7% relative reduction in deaths among those whose alerts fired within three hours. It also generated plenty of false alarms, which is the honest trade-off nearly every predictive tool faces. Where those failures originate is easier to reason about once you understand how AI learning works and why it matters now: models reproduce the patterns of their training data, including the gaps.
Drug discovery on a compressed timeline
AlphaFold’s 2020 release put predicted structures for roughly 200 million proteins into a public database. Structural biologists who once spent years on a single crystallography project now start with a strong guess. That is a research accelerator rather than a cure, but it compresses the earliest and most uncertain stage of the pipeline.
The more concrete test is Insilico Medicine’s INS018_055, a candidate for idiopathic pulmonary fibrosis that reached Phase 2 trials. The company reports that the AI-assisted process took about eighteen months and a fraction of the usual early-stage cost to nominate the molecule. Independent replication of those numbers is still thin, and Phase 2 is where most drugs die regardless of how they were found. Still, the timeline is the story: this is the first serious evidence that generative chemistry can shorten the front end of a decade-long grind.
The unglamorous revolution: paperwork
Ask a clinician what eats their day and few will say diagnosis. A widely cited study in the Annals of Internal Medicine found physicians spending roughly two hours on electronic health record tasks for every hour of direct patient care. That ratio is where ambient documentation tools have found traction.
Products like Nuance DAX Copilot, Abridge and Suki listen to a consultation and draft the note. The clinician edits rather than types. Early hospital deployments report meaningful drops in after-hours documentation, though the drafts still need checking and the room has to be quiet enough for the microphone to work. This is one of the clearest examples of generative AI reshaping an industry not by doing something new, but by removing a task nobody wanted.
Behind the clinic door
Coding, claims and prior authorisation consume enormous administrative budgets. Insurers and providers are both experimenting with systems that read a chart, match it to a policy, and assemble the justification package automatically. This is where agentic AI that can act across multiple systems starts to look less like a demo and more like a cost centre worth attacking, provided a human signs off before anything is submitted.
Monitoring that never clocks off
Consumer wearables pushed the idea of continuous measurement into the mainstream; the Apple Watch atrial fibrillation detection feature has been studied in hundreds of thousands of users. In hospitals the same principle arrives as virtual nursing stations and camera-based fall detection, where a model watches twenty beds overnight and raises an alarm when a patient tries to climb out unassisted. Nobody wants to be filmed in a hospital bed, so consent, retention and who reviews the footage become the real design questions.
What still goes wrong
The failure modes are consistent and worth naming plainly.
- Bias baked into the data. A model trained mostly on one demographic can quietly underperform on another. Pulse oximeters already do this, and algorithms inherit the same blind spots.
- Distribution shift. A scanner is replaced, a coding standard changes, a care pathway is redesigned, and accuracy drifts. Models need monitoring after deployment, not just validation before it.
- Automation bias. Clinicians who see a confident score stop looking as hard. The tool becomes a ceiling rather than a floor.
- Confident nonsense. Language models generating summaries or answers can invent a detail that reads perfectly plausible. In a note it is an annoyance. In a clinical recommendation it is a hazard.
- Liability fog. When an alert is missed or a suggestion is followed and the patient is harmed, the accountability question is still unsettled in most jurisdictions.
How to deploy this without wasting a budget
Hospitals that get value from AI tend to follow the same short list. They pick one painful, measurable workflow instead of a platform-wide rollout. They record a baseline before switching anything on, because “it feels faster” is not evidence. They put a clinician who actually uses the tool in charge of the pilot, since procurement decisions made by committees alone reliably miss the workflow friction that kills adoption.
Vendor diligence matters more than the demo. Ask for validation on a population that resembles yours, ask what happens when the model is uncertain, and ask who is contractually responsible when it is wrong. The market is crowded and not every name on a sales deck will still exist in three years, so a working sense of which AI companies are actually shaping the future is worth the reading time.
Then plan for the boring part. Someone has to own drift monitoring, revalidation and retraining, and that someone needs a budget line. A model that was 94% accurate at go-live is not a finished product; it is a commitment to keep watching. The hospitals treating it that way are the ones with results to show a year later, while everyone else is still writing press releases about a pilot that ended quietly in March.
