AI in Healthcare
AI in healthcare is the use of machine-learning systems to support clinical care and its administration: reading medical images, flagging risk in patient records, drafting documentation, triaging cases and assisting research. It is a set of narrow tools with very different risk profiles, from software that suggests a billing code to software that influences a diagnosis, and regulators and clinicians treat them accordingly.
Imaging was the first area to produce regulated products. Deep learning suited image classification, and a 2016 JAMA paper by Gulshan and colleagues at Google showed a neural network detecting diabetic retinopathy in retinal photographs with accuracy comparable to ophthalmologists. In 2017 Stanford researchers published CheXNet, a chest X-ray model, and Esteva and colleagues reported in Nature a skin-lesion classifier. In April 2018 the US Food and Drug Administration authorized IDx-DR, from the company IDx, through its De Novo pathway as the first autonomous AI diagnostic system: it analyzes retinal images in a primary-care setting and returns a result on whether more than mild diabetic retinopathy is present, without a clinician interpreting the image. The FDA maintains a public list of AI-enabled medical devices that it has authorized, and radiology accounts for the large majority of entries.
Diagnostic aids and risk models are the second category. Computer-aided detection tools mark suspicious regions on mammograms, CT scans and pathology slides for a human reader to confirm. Risk models in electronic health records estimate sepsis, deterioration or readmission. These are decision support: a clinician remains responsible, and the practical question is whether the tool improves outcomes rather than merely scoring well on a retrospective dataset. External evaluations have sometimes been sobering. A 2021 study in JAMA Internal Medicine by Wong and colleagues found that a widely deployed proprietary sepsis prediction model performed much worse in hospital practice than its developer had reported, with poor sensitivity and many false alerts. IBM's Watson for Oncology, publicised as a treatment advisor, drew criticism in 2018 press reports for unsafe or poorly supported recommendations, a cautionary tale about marketing outrunning validation.
Documentation is the fastest-growing use. Ambient scribe tools record a clinic visit, transcribe it, and draft the clinical note for the physician to review, aimed at the administrative burden that is widely linked to clinician burnout. Large language models also draft patient-message replies and summarize records. The risks are specific: a fluent note can contain an invented detail or omit a symptom, reviewers may skim, and recording a consultation raises consent and data-protection issues. Because such tools do not diagnose, they often fall outside the strictest device regulation, which makes institutional governance and clinician review the main safeguards.
Bias and validation are the central risks. Models learn from the data they are given, and health data reflects unequal access and measurement. Obermeyer and colleagues reported in Science in 2019 that a commercial algorithm used to allocate extra care resources underestimated the needs of Black patients because it used past health spending as a proxy for illness, and spending was lower for equally sick Black patients. Performance can also fall when a model moves to a new hospital, scanner or population, a problem called dataset shift. Sound practice involves external validation on local data, subgroup performance reporting, monitoring after deployment, and clear rules for when humans override the system. Software that changes after approval poses a further regulatory challenge, and the FDA has issued guidance on predetermined change-control plans for such devices.
Misconceptions to avoid. AI does not replace the doctor in most settings; even the autonomous IDx-DR case handles one narrow screening question. High accuracy on a benchmark is not evidence of patient benefit, and a tool that finds more abnormalities may drive overdiagnosis. Related entries cover bias, validation, and the language models behind scribes; this entry concerns where those ideas meet clinical responsibility.
Read AI in healthcare as a collection of narrow, evidence-bound tools: where one has been tested in a defined population against a clear clinical question it can extend care, and where it is trusted beyond that evidence it adds risk.