MINA.
Türkiye's Clinical
Artificial Intelligence.
MINA is a clinical decision support platform designed and built from scratch for Turkish clinical practice. On top of an open-weight large language model, Opinion AI used Continued Pre-Training (CPT) to internalize the Turkish clinical language; added specialty expertise via LoRA adapters; and connected it to live clinical literature via a GraphRAG architecture.
The birth of an AI model
in Türkiye.
Clinical practice is not universal. The patient a physician sees in Türkiye, the guidelines they follow, the discharge summaries they write, the SGK and reimbursement context they work within — all are a reality unique to this geography. MINA was born as a response to this reality.
We know that translating a foreign general-purpose model into Turkish is not enough. MINA was built on top of an open-weight base model with a three-layer process: CPT internalized the Turkish clinical language; LoRA adapters added specialty expertise; and GraphRAG linked it to live clinical literature.
A three-layer
clinical AI architecture.
MINA was not fine-tuned. The clinical knowledge was instilled into the model by a different method: first the language itself with CPT, then specialty knowledge with LoRA, and finally live information retrieval with GraphRAG.
CPT · Continued Pre-Training
On top of a 70B-parameter open-weight base model, we applied continued pre-training instead of instruction fine-tuning. The model internalized the Turkish clinical language as its "native tongue" through continued pre-training on Turkish medical terminology, clinical guidelines and real-world clinical records. Trained on H100 GPU infrastructure.
LoRA · Specialty Adapters
After CPT, the base weights are frozen, and a dedicated LoRA adapter is trained for each clinical specialty and task (ICD-10 coding, reimbursement compliance, discharge summary analysis, MRI interpretation, neuropsychological assessment, social-insurance pre-auth). The appropriate adapter activates based on context — a single model becomes a multi-specialty system at lower compute cost.
GraphRAG · Live Information Retrieval
To prevent trained knowledge from going stale, we integrated GraphRAG (Graph-based Retrieval Augmented Generation). PubMed clinical literature is indexed in a graph database; medical textbooks in a vector search store; routing between the two sources is managed by a PyTorch-based ML routing layer. Responses are always produced with reference to current sources.
Scientific foundation:
MINA's data sources.
MINA's knowledge base is built from international medical literature, clinical trial registries, open-access health datasets, insurance billing codes and evidence-based clinical guidelines. No output is produced without sources; every answer is traceable to these origins.
Medical Literature & Publications
- PubMed (PMC, PubChem, TPA)
- MedLine (Bulk Data)
- Google Scholar
- Semantic Scholar
- ArXiv Health
- SpringerLink Health
- Elsevier ScienceDirect
- IEEE Xplore Medical
Clinical Trials & Patient Data
- ClinicalTrials.gov
- NIH OpenData
- UK Biobank
Electronic Health Records (EHR)
- MIMIC-IV
- eICU Collaborative Research Database
- PhysioNet
Insurance & Billing Datasets
- CMS.gov (Centers for Medicare & Medicaid Services)
- ICD-10 & CPT coding (via university collaboration)
Clinical Guidelines & Evidence-Based Medicine
- WHO & IRIS (WHO publication repository)
- COVID-19 Clinical Trials Registry
- FDA Drug Database (essential drug information)
- The Lancet (Elsevier · ScienceDirect)
- JAMA
- BMJ Open
Turkish Health Regulations & Standards
- SGK Healthcare Implementation Communiqué (SUT)
- Ministry of Health Targets (BH-13, BH-32, BH-33, BH-34, BH-35)
- ADSH-13, ADSH-14 indicators
- KVKK (Personal Data Protection Law No. 6698)
- TİTCK Drug & Medical Device Database
- e-Nabız and MKYS integration standards
- Official Gazette — health regulations & notices
- Turkish Medical Association clinical guidelines
- Ministry of Health diagnosis & treatment guides
All sources are used in compliance with copyright and usage rights; licensed sources are accessed through institutional / academic agreements.
Turkish health modules:
a product layer for Türkiye only.
On top of the model layer, product modules were developed that match the daily realities of the Turkish healthcare ecosystem. Each module invokes the relevant LoRA adapter, validates with GraphRAG and integrates into the user's workflow.
SGK Pre-Auth & Reimbursement
Social-insurance pre-authorization rules and reimbursement compliance validation.
GDPR-Compliant Data Handling
Clinical data is processed locally, in an isolated environment, within the GDPR framework.
Neuropsychological Assessment
Turkish neuropsychological tests and assessment tools.
MRI Image Analysis
Lesion, anomaly and finding extraction in medical imaging.
Patient Digital Twin
Modeling all clinical interactions in a single live profile.
Discharge Summary & ICD-10
Discharge summary consistency analysis and ICD-10 coding suggestions.
Enterprise-grade infrastructure
for sensitive data.
All training and inference processes run on KVKK- and GDPR-compliant infrastructure. The platform holds international information security, privacy, quality, environment and occupational safety management system certifications.
- KVKK · Turkish Personal Data Protection
- GDPR · European Data Protection
- ISO/IEC 27001 · Information Security
- ISO/IEC 27701 · Privacy Information Management
- ISO 9001 · Quality Management
- ISO 14001 · Environmental Management
- ISO 45001 · Occupational Health & Safety
- Isolated / On-Premise Deployment
Same clinical brain,
in two sectors.
MINA runs on the same clinical knowledge base across Opinion AI's two platforms; its role and authorization specialize to the platform it is in.
Hospital Platform
Clinical decision support for the hospital operator and physician. Write access to the patient digital twin; ICD-10 suggestions, reimbursement compliance, discharge summary analysis, pre-auth preparation, MMIS inventory and audit simulation.
Insurance Platform
Decision support for the insurance specialist in pre-authorization. File-level document reading, reimbursement compliance, policy coverage assessment, similar case comparison, fraud risk screening, and portfolio-level actuarial audit simulation.
MINA is not the decision maker,
it is the decision supporter.
Action with the User
All critical actions (approve, missing document, reject, prescribe) are taken by the end user. MINA offers a reasoned suggestion; verified or rejected with a single click.
Transparent Rationale
For every prediction, source text, confidence score and transparent rationale are provided. Responses are auditable, educational and verifiable.
Hallucination Control
All outputs are tied to reference sources; unverifiable claims are not produced. Answers are grounded in clinical guidelines and scientific databases.
Want to see MINA
in your own processes?
Let us walk through how MINA works in your clinical and operational workflow.
Contact Us
AI-powered decision platform for healthcare. Better-informed decisions, better outcomes.