MINA

MINA v3.7

New release.
Expanding clinical coverage.

MINA v3.7 is the latest version of our clinical decision-support platform — proven in Turkish clinical practice and built to adapt to any market. Five LoRA expertise adapters and three cross-cutting capabilities (safety, Turkish NLU, operational) were evaluated separately across 14 clinical specialties and sub-specialties.

Performance results Documents

clinical specialties / sub-specialties

5 + 3

LoRA expertise + cross-cutting capabilities

~9,300

evaluation cases (synthetic / anonymized)

evaluation protocols

Highlight results

%91

SUT decision accuracy

Target ✓ · ≥ 0.90

%87

ICD-10 Top-1

Target ✓ · ≥ 0.85

%96

Clinical trap detection

Target ✓ · ≥ 0.95

%0.3

Clinically critical hallucination

Target ✓ · ≤ 0.005

%93

Turkish lay-language

Target ✓ · ≥ 0.92

%93

Fraud detection (AUC)

Target ✓ · ≥ 0.92

Comparative performance

Primary metrics per task. The red column is MINA v3.7; a green outline indicates the target was met.

LoRA / Task	Metric	Target	MINA v3.5	MINA v3.6	MINA v3.7	LLM-A	LLM-B	LLM-C
ICD-10 Coding	Top-1 Accuracy	≥ 0.85	%74	%81	%87	%79	%81	%74
ICD-10 Coding	Top-3 Accuracy	≥ 0.95	%88	%93	%96	%91	%92	%88
ICD-10 Coding	Hierarchical F1	≥ 0.90	%83	%88	%92	%88	%89	%85
Discharge Summary	Slot F1 (micro)	≥ 0.88	%80	%85	%89	%86	%87	%82
Discharge Summary	Hallucinated sentences	≤ 0.02	%2.2	%1.6	%1.2	%1.8	%1.4	%2.5
Fraud / Policy	Fraud AUC-ROC	≥ 0.92	%86	%89	%93	%85	%87	%81
Fraud / Policy	Rejection Risk AUC	≥ 0.90	%83	%87	%91	%82	%84	%78
Fraud / Policy	Coverage Accuracy	≥ 0.94	%90	%92	%95	%86	%89	%83
SUT Compliance	Decision Accuracy	≥ 0.90	%82	%86	%91	%71	%75	%65
SUT Compliance	Clause MRR	≥ 0.88	%79	%85	%89	%57	%61	%48
SUT Compliance	Citation Hallucination	≤ 0.005	%0.8	%0.5	%0.3	%4.5	%2.8	%6
Clinical Decision	DDx Top-3 Accuracy	≥ 0.92	%85	%89	%93	%91	%93	%89
Clinical Decision	Plan–Expert Agreement	≥ 0.82	%77	%81	%84	%82	%85	%79
Clinical Decision	Contraindication Recall	≥ 0.92	%87	%90	%93	%89	%91	%86
Safety	Clinically Critical Hallucination	≤ 0.005	%0.8	%0.5	%0.3	%1.1	%0.6	%1.5
Safety	Clinical Trap Detection	≥ 0.95	%89	%93	%96	%83	%88	%76
Turkish NLU	Lay-Language Accuracy	≥ 0.92	%82	%88	%93	%78	%81	%74
Turkish NLU	Regional Dialect	≥ 0.85	%75	%82	%88	%57	%63	%49

MINA v3.7Target metNext-best model

Methodology & scope. Results were produced with Opinion AI’s internal evaluation system on synthetic and anonymized test sets (total n≈9,300 cases; no real patient data). The compared LLM-A / LLM-B / LLM-C are leading general-purpose large language models, evaluated under the same protocol and test set using their publicly available API versions as of May 2026, and anonymized for commercial reasons. Version updates may change these figures. Full methodology is in the technical white paper.

Version-over-version progress

Improvement in primary metrics across v3.5 → v3.6 → v3.7 on a fixed test set.

ICD-10 · Top-1 Accuracy

v3.5 %74

v3.6 %81

v3.7 %87

SUT · Karar Accuracy

v3.5 %82

v3.6 %86

v3.7 %91

SUT · Madde MRR

v3.5 %79

v3.6 %85

v3.7 %89

Fraud · AUC-ROC

v3.5 %86

v3.6 %89

v3.7 %93

Türkçe NLU · Halk Dili

v3.5 %82

v3.6 %88

v3.7 %93

Safety · Klinik Tuzak Yakalama

v3.5 %89

v3.6 %93

v3.7 %96

Documents

Detailed findings, full benchmark tables and methodology are in these white papers. Because they contain competitive comparisons, they are not published publicly; they are provided to organizations on request under a confidentiality agreement (NDA).

Executive

Executive Summary

Executive summary for decision-makers: key findings, competitive positioning and business impact (ROI, efficiency).

Under NDA · request

Technical

Technical White Paper

Model architecture (CPT + LoRA + GraphRAG), training approach, test-set design and evaluation methodology; full benchmark tables.

Under NDA · request

Clinical

Clinical White Paper

Clinical decision-support findings across 14 specialties, safety / red-team results and the physician-in-the-loop framework.

Under NDA · request