Medical AI: from promise to practice

AI models can analyse data continuously, extract coherent information from multiple sources, and handle time-consuming administrative tasks. These are precisely the capabilities that medicine needs — and AI is now making its way from research labs to patients' bedsides. But how do you build AI that is reliable, sovereign, and confidential, without letting it replace the human altogether?

Long a fixture of futuristic thinking, AI has now entered actual medical practice. Tools capable of interpreting images, tracking physiological parameters in real time, or summarising clinical records are gradually being integrated into clinical workflows. In the United Kingdom, the National Health Service launched in 2025 the largest breast cancer screening clinical trial ever conducted. Known as "Edith", it tests several AI systems on mammograms from nearly 700,000 patients. Another telling indicator: the exponential number of AI-based medical devices approved by the globally influential  Food and Drug Administration (FDA) — roughly 1,450 devices in total as of March 2026. "We've moved from promises to reality," sums up Douglas Teodoro, professor in the Department of Radiology and Medical Informatics at UNIGE Faculty of Medicine.

AI is gaining ground so rapidly because medicine has become a data-saturated field. Imaging, vital signs, biological analyses, patient histories, and clinical reports form a vast, fragmented body of information. That is where AI finds its role, complementing human capabilities. Two projects underway at UNIGE are exploring this potential — one in polypharmacy, the other in cardiovascular monitoring. Both share a commitment to a European AI model that respects data privacy.

Augmented medicine

"From a data science perspective, medicine is an extraordinarily complex field. We have text, images, sound, and multiple structured data," explains Douglas Teodoro. "And as humans, we struggle to integrate all of that simultaneously." Current AI models are precisely designed to process these sources simultaneously. "In concrete terms, this means systems can combine different types of medical information to provide a unified view." The potential — for diagnosis, treatment, and above all prevention — is enormous.

That is exactly what Douglas Teodoro and his UNIGE and HUG colleague Elena Tessitore are seeking to demonstrate. Their Argentic project aims to continuously analyse physiological data in cardiology and rehabilitation. Heart rate, blood pressure, and blood oxygenation are to be interpreted in real time to detect the earliest signs of deterioration. "You can't read an ECG for 24 hours straight. A machine can," Teodoro explains. "AI brings that dimension — achieving a form of partial autonomy. It's like an autopilot: it assists, but doesn't replace." In other words, it handles repetitive tasks while leaving final decisions to medical staff.

The second project, AIM-SAFE — developed in collaboration with Caroline Samer of UNIGE and HUG — focuses on the safety of drug treatments. In polypharmacy situations, interactions between active molecules themselves, and between those molecules and the body, are numerous and difficult to anticipate. Systems for identifying drug incompatibilities do exist, but they tend to generate too many alerts. "Clinicians end up ignoring them. We call it alert fatigue," notes the researcher. This work involves integrating clinical context to produce more relevant, situation-specific alerts. The goal is twofold: reduce risk and cut unnecessary prescriptions.

Preventing leaks

To develop these AI tools, one cornerstone is the need to "train" the software using data — which immediately raises fundamental questions of data confidentiality. This is not only an ethical issue and a legal prerequisite; it is also central to public acceptance of AI in healthcare.
To address this, the solutions developed in Teodoro's projects rest on the principle of data non-circulation. "We send the algorithm — because at its core, AI is and remains an algorithm — to the data, not the other way round," he explains. In practice, each hospital involved in the projects trains the AI on its own data and retains that information in-house. AI models are trained locally, and only the outputs of that learning are shared. This approach, known as "federated learning", guarantees data sovereignty to health institutions.

Behind the algorithms

Beyond federated learning, how do these systems actually work? They rely on a combination of several technological building blocks. First, the AI systems are built on so-called "multi-agent" architectures. "Agents are modules dedicated to specific tasks," explains Teodoro. Some handle text only, others images, others still deal exclusively with physiological data. These agents are then implemented using small language models, or SLMs. Similar in nature to the large language models underpinning ChatGPT or Claude, their smaller footprint allows them to be deployed within hospital environments, avoiding data sharing and hardware constraints. "SLMs adapt readily to complex medical tasks, but also to different languages — which is particularly significant in the European context," the researcher notes.

Another key tool used in these projects is the digital twin: models capable of simulating how a given patient might evolve. "You can, for example, virtually test what would happen with or without a medical intervention," illustrates Teodoro. These simulations make it possible to anticipate complications and respond accordingly.

Sovereignty on two fronts

A further critical issue is sovereignty. Most models and infrastructures are currently developed in the United States or China. "Today, we can't control the entire chain — we're obliged to use what exists, particularly in terms of computing infrastructure, such as GPUs graphic processors," acknowledges Teodoro. "On the other hand, we can maintain sovereignty over the data and the models we develop building on this scaffolding." Because the existing systems are far from neutral. "They embed a particular worldview," he says — implying they can be steered, even instrumentalised, according to the visions and interests of those who design them. Biases that could influence diagnoses and treatments, and therefore people's health.

None of this means AI should replace clinicians. "AI steps in where humans reach their limits — in analysing massive volumes of data or sustaining continuous monitoring," insists Teodoro. "It doesn't replace the physicians, but it redefines their role in certain contexts. Less bogged down in documentation and information retrieval, they can focus on interpretation and on the relationship with their patients."

Media