Artificial intelligence is no longer science fiction in Estonia's healthcare system but an everyday tool, moving from labs to operating rooms and emergency departments. While chatbots and image generators grab headlines, a quieter but far more consequential revolution is unfolding inside hospitals. At North Estonia Medical Center (PERH) and Tartu University Hospital, doctors in critical situations are now assisted by algorithms that spot what the human eye cannot and complete hours-long analyses in minutes.

The need is most acute when every second counts. In stroke care, the rule is stark: time is brain. From the moment an ambulance delivers a patient with suspected stroke, a race against the clock begins. In the case of an ischemic stroke, millions of brain cells die each minute a vessel remains blocked.

This is where AI steps in. According to Martin Reim, a radiologist and lecturer at Tartu University Hospital, rapid diagnosis is key in stroke treatment, determining whether a patient walks out of the hospital or remains bedridden. Applications used at both hospitals analyze CT scans almost instantly, mapping areas of the brain with critically reduced blood flow.

"The application assesses how much of the brain can still be saved with different treatment methods and how much tissue is already permanently damaged," Reim said. That gives doctors fast, objective information to decide whether to begin clot-dissolving therapy or send the patient to surgery.

Kristo Erikson, chief medical officer at North Estonia Medical Center, said AI has significantly increased the number of patients whose brain damage can be reduced. Where treatment decisions once depended largely on how much time had passed since the stroke began, doctors can now rely on precise imaging data. If AI shows that tissue is still alive, there is still hope.

No shortage of work

Developments at PERH and Tartu University Hospital are part of a rapidly expanding global push. While fewer than 400 AI-powered medical devices had received approval from the U.S. Food and Drug Administration (FDA) in 2020, the figure surpassed 1,250 last year.

According to TalTech researcher Elli Valla and Associate Professor Maie Bachmann, radiology clearly dominates among the hundreds of approved devices. Imaging data is standardized, extensive and well labeled, making it easier to train and validate AI models.

Though radiology is AI's flagship field, its footprint is widening. Estonian doctors point to capsule endoscopy as a vivid example: after a patient swallows a small camera, AI selects the few suspicious frames from thousands of images of the digestive tract, sparing physicians hours of video review.

The greatest time savings may come in screening programs. Erikson and Reim note that in several countries AI already serves as a first reader in breast and lung cancer screening, automatically filtering out clearly healthy cases so that only potentially problematic scans reach a physician for further review.

In intensive care, AI acts as a kind of digital watchdog, helping fine-tune drug interactions and kidney-related dosing while alerting doctors immediately to potential errors. North Estonia Medical Center and Tartu University Hospital are also jointly developing digital pathology solutions in which AI assists in detecting diseased cells in tissue samples.

In neurology, AI-based tools help specialists assess electroencephalogram (EEG) readings, speeding up routine work, though fully autonomous epilepsy diagnoses remain out of reach.

Global trends and leading hospitals worldwide suggest even broader applications. More mature areas include ophthalmology where AI screens retinal images for signs of diabetic retinopathy and cardiology where it analyzes ECG-based heart rhythm disorders.

Machines can detect what the human eye misses

Beyond speed, AI can at times deliver a level of precision that surpasses the human eye. Martin Reim points to interventional radiology where doctors treating cancer must block the blood vessels feeding a tumor to effectively starve it.

"We have on several occasions identified the correct target vessels that might have gone unnoticed by the human eye using conventional methods," Reim said. AI can clearly distinguish a tumor from surrounding tissue and pinpoint the tiny arteries that need to be sealed, helping spare healthy tissue from the effects of systemic medication.

Similar gains in accuracy are seen in radiation therapy. At North Estonia Medical Center, software can automatically detect and outline nearly 300 anatomical structures in just minutes. "With autocontouring, no one on the treatment team can imagine working without it anymore. It saves an average of 50 percent of the time needed to prepare a treatment plan and in some cases up to 90 percent," said Kristo Erikson.

That allows doctors to focus more on complex cases and patient communication rather than drawing lines on a screen. Still, both Erikson and Reim stress that AI is a support tool, not a decision-maker. Final responsibility and control always remain with the physician.

Reality check

If all this makes AI sound like a silver bullet for hospital bottlenecks, the reality is more complicated. Estonian doctors admit that new technology can at times mean spending more time behind a computer — often to manually correct the machine's mistakes.

While AI detects patterns faster than humans, it lacks a broader understanding of a patient's health and clinical intuition. As a result, it can sometimes slow work rather than speed it up. Tartu University Hospital radiologist Martin Reim said introducing new tools often brings more clicks and additional steps.

False positives are a particular headache. In such cases, AI identifies a disease where none exists. According to Kristo Erikson, chief medical officer at North Estonia Medical Center, this can make a radiologist's work more time-consuming. "They have to deal with the consequences of unnecessary follow-up tests ordered based on an AI diagnosis," he said. Instead of focusing on patients, doctors may find themselves reviewing the files of healthy people incorrectly flagged by a machine.

Reim also pointed to a more serious risk: the rapid pace enabled by AI could lead to treatment decisions based on incorrect information. For example, if AI highlights a possible fracture on an emergency room X-ray with a colored box — when in fact it is a normal anatomical structure — problems can arise.

If the analyzed image is sent to the national Health Image Bank before a radiologist confirms it, an emergency physician might see the colored box and put the patient in a cast unnecessarily. "Because of the AI analysis, a patient may receive a cast or an additional radiological exam involving higher radiation that was not needed," Reim warned.

That is why most radiology AI solutions carry the disclaimer "not for diagnostic use." The clause underscores that AI is not a physician but more like a smart yet inexperienced trainee whose work must always be checked by a specialist.

Oversight is complicated, however, by what researchers call the "black box" problem. In straightforward cases such as fractures, a doctor can verify whether the machine's marking is correct. But with more complex models, the decision-making process can be opaque to human understanding.

This raises a fundamental question: how well must a physician understand the technology guiding their decisions? As Maie Bachmann and Elli Valla have asked, is it enough to act like an air traffic controller — monitoring and managing the system — or must doctors also understand the "engine" to retain meaningful responsibility toward patients?

Although tools such as heat maps can show where AI focused its attention on an image, they offer only a partial glimpse into the reasoning process. In machine learning, there is an unavoidable trade-off between accuracy and explainability: the most powerful models are often the hardest to interpret. For that reason, doctors consistently view algorithms as decision-support tools, not autonomous diagnosticians.

E-health paradox

While the "black box" problem challenges AI worldwide, training algorithms also runs up against data limitations. The issue is not unique to Estonia, but in a country known for its e-state, the contrast between a polished digital facade and the actual quality of health data stands out. "Estonia's e-health system is rightly praised — we have a unique digital infrastructure. But for AI, it's not enough that data exists; what matters is the format," Elli Valla said.

In reality, much of Estonia's medical data consists of free-text entries, PDFs or scanned documents. Machine learning, however, requires structured and standardized datasets. "Right now, each research project often requires months of manual work just to make the data usable," the researchers noted.

That means that even when data exists, AI cannot readily use it. Kristo Erikson likewise said data quality is often poor and fragmented, limiting solutions to narrow fields.

To address fragmentation and privacy constraints, researchers are developing new approaches such as synthetic data. TalTech scientists explained that generative models can create synthetic medical images — for example, X-rays that are virtually indistinguishable from real fractures.

"This makes it possible to train new diagnostic models to detect rare diseases without putting sensitive patient data at risk," Valla said. Federated learning is also advancing rapidly. In that model, algorithms are trained across multiple hospitals while sensitive patient data remains within each institution. Even amid data bottlenecks, such methods allow research to move forward.

The price of innovation and gauntlet of tenders

Bringing new AI solutions into hospitals is far from simple. Kristo Erikson cited radiation therapy as an example: automatic contouring alone costs hospitals about €17 per patient. While not astronomical at a system-wide level, he acknowledged that AI does not necessarily deliver immediate financial savings for hospitals.

One reason is market fragmentation. "AI solutions are currently expensive because many small, specialized tools are being developed by different companies and startups," Erikson explained. Instead of buying one comprehensive system, hospitals must purchase, test and integrate numerous niche applications.

Business models vary as well. Some vendors charge a one-time purchase fee; others bill per analyzed patient or image. Reim also pointed to platform-based models where hospitals join a larger ecosystem and deploy different applications as needed. Although a one-time purchase may seem attractive, hospitals tend to prefer license-based models. "With a single purchase, the product is generally not further developed or you have to pay again for a new version," Erikson said.

Often, however, bureaucracy and IT integration pose even greater obstacles than cost. Erikson noted that Estonia's public procurement rules can slow adoption in the fast-moving AI sector, making it complicated for public institutions to acquire new tools. Integrating solutions into existing hospital IT systems is also time-consuming. "We need a system that allows fast-track testing of different applications so we can learn how to integrate them and assess whether they deliver the expected benefit," Reim said.

Beyond domestic hurdles, European Union regulations also set firm boundaries. TalTech researchers stressed that strict certification is an essential safeguard for patient safety, but in its current form it can also hamper innovation. "Software can harm a patient if it fails and someone must ensure it is safe," Bachmann and Valla noted. At the same time, the dual regulatory burden of EU medical device and AI rules can create steep barriers to market entry for smaller developers and researchers.

That also helps explain where procurement funds ultimately go. Contrary to common fears, medical AI is not dominated solely by technology giants such as Google or Amazon. According to Bachmann and Valla, the market is still led by traditional medical technology heavyweights such as GE HealthCare, Siemens and Philips. Hospitals tend to favor established partners whose equipment they already use, leaving startups and large tech firms with a smaller share of the market.

Personal approach

Until now, AI in medicine has largely focused on detecting disease and analyzing images. But the rise of generative AI — now familiar as a consumer technology — has opened an entirely new direction. The next major leap may come not from a new diagnosis, but from cutting bureaucracy.

"Generative AI has shifted the focus from analyzing data to creating new content," Elli Valla said. Doctors are no longer looking to AI solely for a second opinion on an X-ray, but for practical help managing daily paperwork. Generative tools offer real-time speech recognition and translation, automatically converting a doctor-patient conversation into a structured medical record.

Researchers noted a striking trend: healthcare workers have begun adopting generative AI tools on their own faster than hospitals can roll out official solutions. Clinicians are already using platforms such as ChatGPT, Gemini, Claude and Perplexity, as well as the more specialized OpenEvidence, to quickly find scientific consensus and work through differential diagnoses rather than simply googling symptoms.

That creates a new challenge: if staff are already using such tools independently, how can data security and proper practices be ensured? Hospitals' logical first step has been to move toward Microsoft Copilot where security safeguards and agreements are already in place. At the same time, researchers stress the need to improve AI literacy among staff and establish clear usage guidelines.

The future will not remain confined to hospital walls. TalTech researchers predict that diagnostics will increasingly shift from hospitals into people's homes — not only to detect disease, but to monitor overall health. It is still too early, they say, to expect a program that definitively declares, "You have disease X." By the time such a diagnosis is made, intervention may already be late. Instead, researchers and developers are focusing more on long-term monitoring and the detection of early risk patterns.

In the future, smartwatches and home devices that monitor brain signals could track individuals' health indicators. As Maie Bachmann explained, such a personalized approach is essential because each person's brain bioelectrical signals and bodily functions are unique. What is entirely normal for one patient may be a warning sign for another, making universal medical norms difficult to define.

That is where home smart devices can play a crucial role — not by comparing people to a population average, but by detecting deviations from an individual's personal baseline, turning AI into an effective early-warning tool. "Without prevention, healthcare is not sustainable," Bachmann emphasized.

In summary

Despite the dizzying pace of technological development, the rise of smartwatches and the mathematical precision of algorithms, experts share a common message: the machine does not take responsibility away from the doctor.

Although AI is already an indispensable, life-saving assistant in medicine, it frequently produces false alarms and must be constantly monitored. According to Martin Reim, physicians above all must retain a critical mindset. "Not everything that comes out of a machine is pure gold. We need to understand where errors can arise and test different systems," he said.

Still, he remains optimistic about the future. "AI will hopefully make radiologists' work faster and more accurate and in a few years we may be at a point where we can treat AI as a second pair of eyes that does a substantial share of the preparatory work," Reim said.

If all goes well, then, the patient in the hospital of the future will not be treated by an autonomous robot, but by a doctor who — thanks to the machine — once again has the time to be human.

--

Follow ERR News on Facebook and X and never miss an update!