What building clinical AI in Türkiye taught us

When people imagine building a medical AI, they usually picture the model first: the architecture, the parameter count, the benchmark scores. In practice, the model is rarely the hard part. The hard part is the environment the model has to work in: a specific clinical language, a particular set of codes and a detailed rulebook that decides, clause by clause, whether a treatment is payable. We chose to start in one of the most demanding of these environments, and what we learned there turned out to be more useful than any single accuracy figure.

Türkiye is, by most measures, a difficult place to start. Clinicians write in Turkish, but their notes are threaded with Latin terminology and local shorthand; the same complaint can appear as everyday speech in one file and as formal medical language in the next. On top of this come the SGK and its SUT framework, the rulebook that sets out line by line what social insurance will reimburse. A human operator needs years to internalize it. A general-purpose model trained mostly on English web text has had little exposure to it. We chose this environment because of its difficulty, since a system built for complex conditions is easier to adapt to simpler ones than the other way round.

The approach has three layers. The first is continued pre-training. Rather than teaching the model individual facts or instructions, this step continues its training on Turkish medical terminology, clinical guidelines and anonymized clinical records until it reads that language like a native speaker. The second layer adds specialty knowledge through small, swappable adapters, for example one for cardiology, one for coding and one for reimbursement, so a single model can act like several specialists without the cost of training several models. The third layer connects the model to current literature through retrieval, so each answer rests on a source that can be checked rather than on the model's memory.

In our internal evaluation on synthetic, anonymized cases, the model reached high accuracy on ICD-10 coding and on SUT reimbursement decisions, and it very rarely invented or misattributed a regulatory clause. This last point matters because, in reimbursement, a confidently wrong citation can lead to a rejected claim or an audit finding. The trend over time was also informative. On a fixed test set, ICD-10 accuracy rose steadily from release to release, which shows that the results did not come from a single run. Task-level evaluation results for MINA v3.7 are on the MINA v3.7 page.

These results also matter beyond Türkiye. The things that make Türkiye difficult, such as a non-English clinical language, a complex payer system and country-specific codes, are not unique to Türkiye. Every health system has its own version of each. The method we used to handle them is not tied to any one country either: when the same three layers are applied to a new market's language, rules and records, the model can be localized again. In this sense, the results in Türkiye indicate that the method can be applied in other markets.

It is important to be clear about what these results show. They come from internal evaluation on synthetic data, with no real patient records, and the systems we compared against are leading general-purpose models, kept anonymous and run under the same conditions. This makes them a strong signal of capability on realistic tasks, but they do not replace prospective, real-world validation, which is now underway. A benchmark result is a reason to run a pilot, and a pilot is the basis for deciding on deployment. We prefer to state this plainly rather than overstate the results.

The main lesson is that in clinical AI, local competence cannot be added at the end; it has to be built in from the start. Starting in a demanding environment produced better results and also gave us a method that can be applied in other markets.