Artificial intelligence in medicine: uses and limits
- Concept
AI can read a retina or flag a scan in seconds, and it can also invent a reference with total confidence. Knowing which is which is now a clinical skill.
What it is
Artificial intelligence in medicine covers two main families. Supervised models learn from labelled examples to do one narrow task, such as grading a retinal photograph, and output a probability. Large language models learn from huge amounts of text to predict the next words, so they write fluently on almost anything but can produce plausible falsehoods. Both must be judged like any diagnostic test: on validation in your population, calibration, fairness and effect on patients. The clinician remains responsible for the decision.
Key facts
- Internal validation tests the model on held-back data from the same source; external validation tests it on new hospitals or populations, and performance usually falls.
- Calibration: whether predicted risks match observed rates; a model can rank patients well and still overestimate everyone's risk.
- Bias and fairness: groups under-represented in the training data may get worse performance; report results by sex, age, ethnicity and setting.
- Shortcut learning: models can learn confounders, such as the hospital or device, instead of the disease.
- Language models can invent facts, doses and references; check every clinical claim against a primary source.
- Do not enter identifiable patient data into tools that are not authorised by your institution; health data are sensitive personal data.
Worked example
- A retinal screening model reports sensitivity 87 % and specificity 91 % in a trial where 24 % had referable disease.
- In 1 000 people at 24 %: 240 diseased → 209 true positives; 760 healthy → 68 false positives. PPV = 209 / 277 ≈ 75 %.
- In a primary care setting with 5 % disease: 50 diseased → 44 true positives; 950 healthy → 86 false positives.
- PPV = 44 / 130 ≈ 34 %: two in three referrals are false alarms, though the model has not changed.
- Calibration check: the model assigns 20 % risk to a group in which 10 % have the disease, overestimating twofold.
- Reading: report external performance in your own prevalence and check calibration before trusting a score.
Two families of medical AI
| Feature | Supervised model | Large language model |
|---|---|---|
| Learns from | Labelled examples for one task | Very large amounts of general text |
| Output | A probability or class | Fluent text on almost any question |
| Typical use | Retinal screening, imaging triage, risk scores | Drafting, summarising, answering questions |
| Typical failure | Dataset shift, poor calibration, subgroup bias | Hallucination: confident, plausible falsehoods |
| How it is checked | Sensitivity, specificity, calibration in external data | Checking each claim against a primary source |
In practice
- Before using a tool, ask: what task, which population was it validated in, what are its errors, and who is accountable?
- Use a language model's output as a draft, never as a source: verify facts, doses and references, and remove identifiers first.
- Keep your own judgement switched on: when the tool disagrees with the clinical picture, re-examine rather than defer.
- Tell patients when an AI tool contributed to their care, and record the decision as your own.
Common pitfalls
- Trusting accuracy from the developer's data without external validation.
- Reading sensitivity and specificity without recalculating the predictive value at local prevalence.
- Ignoring calibration because the area under the curve is high.
- Accepting a fluent answer as a verified one.
Common traps
- High accuracy in the paper means high accuracy here: performance usually falls on external data, and predictive values change with local prevalence
- A fluent answer from a language model is a verified answer: these models predict plausible text and can invent facts and references with full confidence
- If the model discriminates well, its risk numbers are right: discrimination and calibration differ; a model can rank patients correctly and still overestimate everyone's risk
- The AI decided, so the AI is responsible: the clinician and the institution remain accountable for decisions made with the tool
- Algorithms are neutral: they inherit the biases of their training data and may perform worse in under-represented groups
Test yourself
Think of the answer, then tap the question to see it.
A screening model with 87 % sensitivity and 91 % specificity is used where 5 % have the disease. About what share of positives are true?
About a third: per 1 000, 44 true positives and 86 false positives, so PPV ≈ 34 %.
A language model gives you a confident answer with three references. What do you do before using it?
Check each claim and reference against a primary source; language models can invent plausible facts and citations.
A sepsis score trained in another country performs poorly in your hospital. What is the likely problem and the fix?
Dataset shift: different patients, practice or data. Validate and recalibrate it on local data before use, and monitor it afterwards.
Why is automation bias dangerous even with a good model?
Clinicians tend to accept the machine's suggestion when it is wrong, overriding their own correct judgement; the rare errors then pass unchecked.
Elsewhere in DurgiPedia
- Ley 1581 de 2012: personal data in health
- Telehealth and telemedicine in Colombia: Law 1419 of 2010 and Resolution 1644 of 2026
- Telephone and video consultations: consulting safely at a distance
Connections
Educational content on epidemiological and statistical methods; the examples use made-up numbers to teach the calculation. Not a clinical or legal guide: current Colombian rules are in the official sources.
Sources: Ethics and governance of artificial intelligence for health (World Health Organization, 2021); Ethics and governance of artificial intelligence for health: Guidance on large multi-modal models (World Health Organization, 2025); Key challenges for delivering clinical impact with artificial intelligence (BMC Medicine (PubMed Central), 2019); Pivotal trial of an autonomous AI-based diagnostic system for detection of diabetic retinopathy in primary care offices (npj Digital Medicine (PubMed Central), 2018).
Spotted a mistake?We are always improving for the community. Tell us what is wrong and our team will review it.Report a mistake
Representative model, for educational purposes only.
