AI in healthcare has moved well past pilot projects in several specific areas, while remaining genuinely early elsewhere. Knowing which is which matters -- both for evaluating vendor claims and for understanding where the interesting engineering problems (and trade-offs) actually are.
Where clinical AI is already deployed
Medical imaging
This is the most mature area by far. Chest X-ray systems detect pneumonia, fractures and tuberculosis at 95%+ accuracy; mammography systems catch early breast cancer at up to 99% accuracy; dermatoscopy models recognise melanoma and other skin cancers; ophthalmology systems (notably Google DeepMind's) detect over 50 eye conditions at expert-comparable accuracy; and pathology models assist with histological slide analysis.
Personalised medicine
Genomic and pharmacogenomic analysis increasingly informs individualised treatment protocols -- identifying risk factors from genetic data, matching medications to a patient's genetic profile, and predicting how a specific patient is likely to respond to a specific treatment.
Drug discovery
AI is compressing drug discovery timelines from a historical 10-15 years down toward 3-5 years in some pipelines, by identifying candidate molecules faster, optimising clinical trial design, and finding new applications for existing approved drugs.
Disease prediction
Models can flag elevated risk of cardiovascular events, developing type 2 diabetes, or sepsis in intensive care -- sometimes months or years before symptoms appear, and in the sepsis case, critically, hours before a human clinician would otherwise notice.
💡 Key idea: the specialties with the most mature AI deployment -- imaging, especially -- share a common structure: a large labelled image dataset and a well-defined, bounded classification task. Less-structured clinical judgement remains much harder to automate well.
The less-discussed problem: prioritisation
Diagnosis gets most of the headlines, but a quieter and arguably more operationally important application is triage: deciding who gets seen first in a busy department. An AI-priority queue can meaningfully reduce wait times for genuinely urgent patients compared with strict first-come-first-served -- but only if the underlying risk model actually discriminates well between urgent and non-urgent cases. A noisy, poorly-validated triage model can end up barely better than random ordering, while still carrying all the operational complexity and liability of a "smart" system.
What to look for before trusting a deployment
- Was the model validated prospectively, on a population like yours, not just retrospectively on historical data?
- Does it have regulatory clearance appropriate to its risk level, and is that clearance for the specific use case being deployed?
- Is there human override authority built into the workflow, not just a recommendation buried in an interface?
- Is accuracy monitored continuously in production, with a plan for what happens when it drifts?
🧪 Try it yourself: the AI Triage Queue Lab simulation lets you experiment with everything described above directly in your browser.