MINA.
Clinical AI that speaks your
health system's language.
MINA is a clinical decision-support platform that masters a health system's own language, coding and regulations — proven across 14 specialties in Turkey's complex regulatory market, and built to adapt to any market. On top of an open-weight large language model, Opinion AI used Continued Pre-Training (CPT) to internalize a market's clinical language; added specialty expertise via LoRA adapters; and connected it to live clinical literature via a GraphRAG architecture.
Explore MINA v3.7's evaluation results and version-over-version progress across ICD-10 coding, SUT compliance, fraud detection, discharge summaries and clinical decision tasks.
Go to the MINA v3.7 section
Proven where it's hardest:
Türkiye.
Clinical practice is not universal. Every market has its own language, guidelines, discharge formats and reimbursement rules. We proved this first where it is hardest — Türkiye's clinical language and complex SGK/SUT system — because a platform that masters that can adapt to any market.
We know that bending a generic global model onto a local reality is not enough. MINA was built on top of an open-weight base model with a three-layer process: CPT internalized a market's 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 an 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 high-performance 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). Biomedical literature is indexed in a graph database; medical textbooks in a vector search store; routing between the two sources is managed by an 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
- International peer-reviewed biomedical article databases
- Open-access academic literature indexes
- Preprint archives
- Medical reference and textbook corpora
Clinical Trials & Patient Data
- Open clinical-trial registry platforms
- Publicly available national health datasets
- Licensed / approved population-health cohorts
Electronic Health Records (EHR)
- Open-access, de-identified clinical research datasets
- Intensive-care and hospital research databases (anonymized)
Insurance & Billing Datasets
- Publicly available healthcare payment / billing datasets
- International medical coding standards (licensed, via university collaboration)
Clinical Guidelines & Evidence-Based Medicine
- Publication repositories of international health authorities
- Public regulatory drug databases
- Peer-reviewed clinical journals and evidence-based medicine sources
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.
Local-market modules:
built per market — Türkiye first.
On top of the model layer, product modules are built to match the daily realities of a healthcare market — starting with Türkiye (SGK/SUT) and extending to each new market. 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.