Artificial intelligence in medicine

A tool that supports the clinician, or one that replaces them?

In its clinical decision support software guidance (January 2026), the FDA distinguishes software that gives recommendations to a healthcare professional about prevention, diagnosis or treatment and lets them independently review the basis of its recommendations so they do not rely on it primarily; this category does not include software that analyses medical images or signals from diagnostic devices.

The guidance warns of "automation bias", meaning the human tendency to over-rely on the suggestion of an automated system.

Among the WHO's six principles for the ethics of AI in health is protecting human autonomy: humans should remain in control of health systems and medical decisions; they also include transparency, accountability, inclusiveness and equity, and continuous assessment of applications during actual use.

Medical imaging

The FDA publishes a list of AI-enabled medical devices authorised for marketing in the United States, and says it is not comprehensive and is updated periodically.

Example: in 2018 the FDA authorised a device that automatically detects, without a clinician's help, more-than-mild diabetic retinopathy in adults with diabetes, in primary care clinics, with immediate referral of those with a positive result to an eye doctor.

In its study of 900 participants at 10 sites, its observed sensitivity was 87.4% and specificity 89.5%; it is designed to detect diabetic retinopathy only, does not screen for glaucoma, and works only with a specific camera and good-quality images.

Decision support and drug development

In decision support: software that displays and analyses patient information and gives recommendations to the healthcare professional, and the FDA requires that the professional be able to review the basis of the recommendation independently.

In drug development: early drug discovery is among the areas of greatest interest for AI, such as predicting the chemical properties and bioactivity of compounds; the FDA received more than 500 submissions containing AI components between 2016 and 2023.

In January 2025 the FDA proposed, in draft guidance, a seven-step framework for assessing the "credibility" of an AI model according to its "context of use", meaning the specific question it is used to answer.

Risk prediction and the limits of external validation

An external validation study (2021) of a commercial sepsis prediction model deployed in hundreds of US hospitals, covering 27,697 patients: performance (area under the curve 0.63) was far worse than the developer reported (0.76–0.83), it missed 67% of sepsis cases, and raised alerts for 18% of all hospitalised patients.

Among the "Good Machine Learning Practice" principles (FDA, Health Canada and the MHRA, 2021): participants and data should represent the intended patients, training data should be separate from test data, the performance of the "human–AI team" should be assessed together, and models should be monitored after deployment.

The WHO warns of biases built into algorithms and of overstating the benefits of AI for health.

Do not use general-purpose AI tools (such as chatbots) to diagnose yourself or change your treatment; discuss any information with your doctor.

Common questions