Artificial intelligence is being applied across medicine — from reading X-rays to mapping protein structures — with results that are sometimes comparable to trained specialists and often faster. But the technology is still at an early, uneven stage: some applications are FDA-cleared and deployed across thousands of hospitals; others remain experimental. Here is what AI can and cannot do in healthcare today.
Diagnosing disease: where AI is already working
The area with the most clinical deployment is medical imaging. AI systems can analyze X-rays, CT scans, MRI images, and retinal photographs to flag signs of disease — often at speeds a human radiologist cannot match at scale.
Two examples already in clinical use:
LumineticsCore (formerly IDx-DR) became the first autonomous AI diagnostic device cleared by the US Food and Drug Administration in 2018. It analyzes retinal photographs to detect diabetic retinopathy — a leading cause of blindness — without requiring an ophthalmologist on-site. In clinical trials it achieved 87% sensitivity and 90% specificity.
Viz.ai operates across more than 1,800 hospitals in the US and Europe, analyzing CT scans in real time to detect large vessel occlusion strokes and immediately alerting the neurology team — cutting the time between scan and specialist response.
Overall, the FDA has authorized more than 1,430 AI-enabled medical devices, with roughly three-quarters in radiology. A 2025 meta-analysis found that AI and radiologists performed comparably on several cancer detection tasks — neither consistently outperforming the other, with AI offering speed and scalability advantages.
Drug discovery at machine speed
Developing a new drug typically takes 10–15 years and costs over a billion dollars. AI is compressing parts of that timeline by predicting how molecules behave before a single test is run.
AlphaFold, a protein-structure prediction system developed by Google DeepMind, won the 2024 Nobel Prize in Chemistry. Its database contains predictions for more than 200 million protein structures and has been used by over 3 million researchers across 190 countries, referenced in more than 35,000 published papers. AlphaFold 3, published in Nature in May 2024, showed a 50% improvement over traditional methods in predicting how potential drug compounds bind to their protein targets — a critical step in drug design.
Beyond target identification, AI is also used to design novel molecules, optimize clinical trial recruitment, and flag drug candidates likely to fail early — before expensive human trials begin.
Risks: what can go wrong
Despite the promise, medical AI carries real risks that researchers and regulators are actively debating.
Algorithmic bias. A landmark 2019 study in Science examined nearly 50,000 patient records and found that a widely used US hospital risk-prediction algorithm systematically underestimated the health needs of Black patients. The algorithm used healthcare spending as a proxy for medical need — but because of longstanding inequalities in healthcare access, Black patients with the same severity of illness spent less. The result: more than 50% fewer Black patients were flagged for high-priority care than their actual health warranted. This kind of bias can stem from imbalanced training data, flawed proxy variables, or both — and it can be invisible until someone actively looks for it.
Data privacy. Medical AI requires large amounts of patient data. In the EU, the GDPR and Medical Device Regulation (MDR) set strict rules on how patient data may be used to train or run AI systems. In the US, HIPAA applies. There have been cases where patient data was used for AI development without adequate consent frameworks, raising serious ethical and legal questions.
Human oversight gaps. AI systems can fail silently — producing confident-looking outputs even when the input falls outside their training distribution. The EU AI Act mandates human oversight and transparency requirements for high-risk AI systems, which includes most clinical AI embedded in or functioning as a medical device.
Why it matters for Georgia
Georgia is at an earlier stage in digital health than most EU countries — building foundational infrastructure such as ePrescription interoperability and telemedicine access in rural areas, with EU and WHO support. Direct deployment of clinical AI in Georgian hospitals is not yet documented at scale.
The EU AI Act is, however, directly relevant to any Georgian medical technology company that wants to sell into the EU market. AI systems embedded in medical devices of class IIa and above are classified as high-risk under the Act, requiring full conformity assessments, transparency documentation, and post-market surveillance. Compliance is required by August 2026 (or August 2027 for CE-marked devices undergoing Notified Body review).
In the news
A vivid example of AI’s reach into medicine appeared this week, when Meta revealed that its Brain2Qwerty v2 system can decode brain signals into text without surgery — using non-invasive EEG sensors worn outside the skull. The technology has clear implications for patients who cannot speak or type. See our report on the breakthrough.
FAQ
Is AI replacing doctors?
Not in any near-term sense. Current AI systems are tools that assist clinicians — flagging abnormalities for review, ranking drug candidates, and accelerating administrative work. They work best when a trained professional interprets the output. Regulators in both the EU and the US require human oversight for high-risk medical AI applications.
Can I trust an AI diagnosis?
AI diagnostic tools cleared by the FDA or EU regulators have passed rigorous safety assessments. But no diagnostic tool — human or machine — is infallible. Always discuss AI-generated results with a qualified clinician before acting on them.
Why does AI work well with medical images?
Pattern recognition in images is a task where large neural networks, trained on millions of labeled scans, can surface subtle visual signals that are difficult for the human eye to spot consistently. Performance depends heavily on the quality and diversity of the training data — which is why bias from unrepresentative datasets is such a persistent concern.
Is my medical data private when AI is involved?
It should be, and regulations require it. In the EU, both GDPR and the MDR impose strict requirements on using patient data for AI training or deployment. In practice, auditing real-world compliance across thousands of health systems remains an active challenge.
Sources: FDA AI/ML-enabled medical devices database; AlphaFold impact report, deepmind.google; Obermeyer et al., Science 2019 (doi:10.1126/science.aax2342); EU AI Act, Regulation (EU) 2024/1689; WHO Health Systems in Action: Georgia 2024.