Turkish medical AI models and how they are built

How a Turkish medical model differs from a general-purpose model

General-purpose chat tools such as ChatGPT or Gemini are designed for a wide range of topics; they are not designed around a particular country's clinical practice or an institution's data and workflows. Turkish clinical text also makes up only a limited share of the training data of models of this kind. Turkish clinical text, meanwhile, has features that a model needs to learn directly:

• The language is agglutinative. The difference between "saptandı" (detected) and "saptanmadı" (not detected) is a single suffix; a model that misses a negation suffix reverses the finding.
• The terminology is mixed. Clinical notes combine Turkish, Latin and English terms, and abbreviations vary by specialty and even by ward.
• Patients speak differently from clinicians. Knowing the clinical meaning of everyday phrases such as "şekerim var" (literally "I have sugar", meaning diabetes) or "tansiyonum çıktı" ("my blood pressure went up") is a precondition for structuring the history correctly.
• Coding practice and reimbursement rules are local. In Türkiye, diagnoses are coded with ICD-10; the rules for public reimbursement are set out in the SUT, the Healthcare Implementation Communiqué published by SGK, the Social Security Institution; and pre-authorization and billing run through Medula, SGK's system.

Translation does not close these gaps; translating a text into English and passing it to a general model can lose a negation suffix, an abbreviation or a local rule along the way. The same applies to any language and health system whose clinical writing and reimbursement rules are its own.

Base model, adaptation and knowledge retrieval layers

The model is built in layers, which makes it easier to test and audit each one on its own.

• Continued pre-training (CPT) means continuing to train an existing base model on the texts of a new language or domain. At this stage the model learns clinical Turkish itself rather than a task: the terms, abbreviations and clinical writing style. Adding some general text to the corpus helps the model retain its earlier language ability.
• LoRA adapters (low-rank adaptation) are small, separately trained layers added without changing the base model's weights. Each adapter carries the expertise of a specialty, an institution or a task such as ICD-10 coding, discharge summary structuring or SUT checks; adding a new task does not require retraining the whole model.
• GraphRAG is retrieval-augmented generation (RAG) carried out over a knowledge graph. Concepts such as diseases, findings, drugs, procedures and rules, and the relationships between them, are held in a graph, and each relationship is linked to its source. A model's training ends on a certain date; through this layer it reaches current guidelines and literature without being retrained.
• Agentic orchestration means the model breaks a task into steps and uses tools in sequence, such as searching the knowledge graph, reading records or running a rule check. In a responsible setup, decisions such as approval or denial stay in an auditable rule engine, and AI explains the finding and its rationale.

Steps to develop a Turkish medical model

The following sequence can be followed when developing a model or evaluating a ready-made solution:

1. Define from the start what the model will do; ICD-10 suggestions, discharge summary structuring, SUT checks and clinical question answering each require different data and different tests. Also put in writing that the model will support decisions rather than make them.
2. Choose an open-weight base model that can run inside the institution and whose license fits the intended use; check how it splits Turkish text into tokens and whether it can handle long documents.
3. Prepare the corpus by bringing together Turkish medical terminology, clinical guidelines, textbooks and regulatory texts, and check the usage rights of every source. If text derived from patient records is to be used, complete anonymization and a legal assessment before training.
4. Put the base model through continued pre-training and measure whether it retains its general language ability.
5. Train and version separate LoRA adapters for specialties, institutions and tasks, so that you can update behavior in one task without affecting the others.
6. Build the clinical knowledge graph and make sure the model's answers show information from documents separately from its own inferences.
7. Define the rules that affect decisions in a rule engine; AI should only explain the finding and its rationale.
8. Validate the model with a fixed test set and review by physicians, then try it in the institution's own workflow through a limited-scope POC.

Clinical validation, hallucination and citing sources

Validation is the most critical stage for a medical model. Hallucination is when a model produces information that looks fluent and convincing but rests on no source or is wrong; an invented dose, a guideline recommendation that does not exist or a wrong code can directly affect patient safety and payment decisions. A larger model does not remove this risk by itself; it has to be measured and contained through the design of the system:

• The test set is kept fixed and separate from the training data; cases can be synthetic or anonymized instead of real patient data and are assessed separately for each specialty.
• Cases with clinical traps, such as a contraindication, an easily missed finding or a misleading history, are included, and whether the model catches them is measured separately.
• When the model cites a guideline or a SUT rule, it is checked that the citation actually exists and supports what is being said.
• Comparisons with general-purpose models use the same protocol and test set, and physicians review the results.

In use, too, the model shows the rationale and source of every recommendation and separates what the document says from its own inference. The decision always rests with the physician or the relevant specialist; AI does not replace the physician.

What Opinion AI does in this process

Opinion AI is a clinical decision support platform for hospitals and health insurers, built on MINA, a clinical AI adapted to Turkish. MINA was developed with a three-layer architecture. Its open-weight base model was adapted to Turkish and the clinical language of 14 specialties through continued pre-training (CPT) and specialized by specialty, institution and task with LoRA adapters; knowledge retrieval runs through agentic orchestration over a clinical knowledge graph (GraphRAG). The 30M+ sources MINA draws on include PubMed, The Lancet, JAMA, BMJ, UK Biobank, MIMIC-IV and ClinicalTrials.

The Physician Ethics and Advisory Board, with physicians from 15 specialties, regularly reviews the model's clinical accuracy and ethical boundaries. In MINA's latest release evaluation, model results were 87% for ICD-10 coding, 91% for SUT decision accuracy and 96% for clinical trap detection (MINA v3.7 · 8 protocols · ~9,300 cases). Details are on the MINA v3.7 page.

The same core runs in TIS, added on top of the hospital's HBYS (hospital information system), for ICD-10 suggestions and discharge summary analysis, and in SIT-I on the insurer's side for pre-authorization and reimbursement checks. The architecture can also be adapted to other languages and health systems; the POCs we have started with a number of hospital groups in Northern Europe build on this design.

People make the decision; MINA shows the rationale and source of its recommendation, and every output can be audited. The data stays inside the institution; Opinion AI does not collect personal data, the institution is the data controller, and the data the model sees passes through KVKK-compliant masking.

Frequently asked questions

What is a Turkish medical AI model?

A Turkish medical AI model is a language model adapted to understand Turkish clinical text, local medical terminology and the rules of the health system in Türkiye. It reads documents such as discharge summaries, consultation notes and reports, knows ICD-10 coding practice and local reimbursement rules such as SUT, and presents its recommendations with their sources. It does not make decisions; it provides decision support to the physician and the relevant specialist.

What is the difference between continued pre-training (CPT) and a LoRA adapter?

Continued pre-training means continuing to train an existing base model on the texts of a new language or domain, and it teaches the model clinical Turkish itself. LoRA adapters are small layers added without changing the base model's weights, and they give the model expertise for a particular specialty, institution or task. In short, language knowledge sits in the base model and expertise sits in the adapters.

What is GraphRAG and why is it used in medical AI?

GraphRAG is a retrieval method in which the model finds relevant information in a knowledge graph before generating an answer and grounds the answer in that information. Clinical concepts, the relationships between them and the sources that support them are held in a graph structure. This lets the model reach current knowledge without retraining and show which source each recommendation rests on.

How can the risk of hallucination be reduced in a medical AI model?

The risk of hallucination is reduced by grounding answers in verifiable sources, showing what the document says separately from what the model infers, and keeping the rules that affect decisions in an auditable rule engine. The model should be measured regularly on a fixed test set, on cases containing clinical traps and for citation accuracy, with physicians reviewing the results. Because the risk never disappears entirely, the decision always rests with the physician or the relevant specialist.

Does developing a Turkish medical model require patient data to leave the institution?

No. Most of the language adaptation can be done with texts that contain no personal data, such as terminology, guidelines and literature; if text derived from patient records is used, anonymization and a legal assessment should be completed beforehand. Health data is special category personal data under KVKK, Türkiye's personal data protection law, and under the GDPR, so the most direct way to protect it is to run the model on the institution's own servers or in an isolated cloud environment. MINA is set up this way; the institution is the data controller, and Opinion AI does not collect personal data.

To assess with us how such a model works on your own documents, use the POC Request form. You can also read the guides on how general-purpose chatbots differ from clinical decision support systems and the role of agentic AI in clinical and administrative healthcare work.