Access Restricted - Academic and Research Use Only

This model is protected under IP Application 2442 filed with the United Arab Emirates University (UAEU). Access requires institutional affiliation and is granted for non-commercial research purposes only. Commercial licensing inquiries should be directed to the UAEU Intellectual Property Unit at IP@uaeu.ac.ae

Log in or Sign Up to review the conditions and access this model content.

🩺 Diagnostic-Reasoning-Q3

Pentabrid V9 — Clinical Reasoning AI

The highest-performing open-source sub-10B medical reasoning model

Benchmark USMLE Prof Med

Base Parameters License IP

Developed by Dr Adnan Agha College of Medicine and Health Sciences | United Arab Emirates University Clinical-Reasoning-Hub


🎯 Overview

Diagnostic-Reasoning-Q3 represents the second generation of the Pentabrid clinical reasoning methodology — a novel approach to medical AI training that mirrors how physicians actually learn and reason. Unlike conventional medical QA models that rely on pattern matching, this system is trained using evidence-based Bayesian clinical reasoning methodology, incorporating likelihood ratios from peer-reviewed sources and structured diagnostic frameworks.

The model achieves 76.4% average across 7 major clinical benchmarks — a +24.2 percentage point gain over the Qwen3-8B base model. These results place it among the strongest medical AI models in the sub-10B parameter class globally.


✨ Key Innovations

Innovation Description
Pentabrid Framework 5-phase self-correcting clinical reasoning methodology
Evidence-Based Reasoning 500+ likelihood ratios extracted from clinical scenarios and textbooks
Diagnostic Frameworks Structured approaches from textbooks to identify diagnosis based on presentation
Clinical Behavior Training 125+ real anonymized clinical cases using clinical reasoning methodology
Weighted Curriculum Learning Prevents catastrophic forgetting through strategic data mixing
Transparent Reasoning Uses <think> tags for explicit step-by-step clinical analysis
Cross-Architecture Validation Pentabrid methodology validated on both Qwen3 and DeepSeek-R1 bases

📈 Headline Results

Benchmark Base (Qwen3-8B) V9 Score Improvement
Professional Medicine 58.9% 89.7% +30.8pp
Medical Genetics 58.6% 88.0% +29.4pp
Clinical Knowledge 61.0% 86.4% +25.4pp
Anatomy 57.6% 79.3% +21.7pp
MedQA (USMLE) 43.9% 66.3% +22.4pp
PubMedQA 48.3% 66.6% +18.3pp
MedMCQA 37.3% 58.6% +21.3pp
Overall Average 52.2% 76.4% +24.2pp

All benchmarks evaluated using lm-evaluation-harness v0.4.11 with zero-shot log-likelihood scoring on official test splits.


🏆 Competitive Positioning

vs Published Medical AI Models (sub-10B class)

Model Parameters MedQA Prof Med Clinical Knowledge
Diagnostic-Reasoning-Q3 (Ours) 8B 66.3% 89.7% 86.4%
MedReason-8B 8B 61.7%
OpenBioLLM-8B 8B 51.0%
Huatuo-o1-8B 8B ~59%
DeepSeek-R1-Distill-8B (base) 8B 52.0%

Diagnostic-Reasoning-Q3 surpasses MedReason-8B on MedQA by +4.6pp and achieves MMLU medical scores competitive with models 4-9x larger.


📊 Training Methodology: The Pentabrid Framework

This model was trained using the Pentabrid clinical reasoning framework — a proprietary methodology that mirrors how physicians actually learn and reason, built on three evidence-based pillars.

Full methodology details are proprietary and protected under IP Application 2442. For licensing, see 💼 Commercial Licensing and Enterprise Partnerships below.


💡 Example: Bayesian Clinical Reasoning in Action

User: A 65-year-old male presents with crushing chest pain radiating to the left arm, diaphoresis, and ST elevation in leads V1-V4. What is the most likely diagnosis?

A. Stable angina B. Acute anterior STEMI C. Pericarditis D. Pulmonary embolism

Model: Analyzing this case using Bayesian clinical reasoning.

Key Findings:

65-year-old male — cardiac risk age Crushing chest pain radiating to left arm — classic ACS pattern Diaphoresis — autonomic response suggesting significant myocardial compromise ST elevation in V1-V4 — anterior leads, LAD territory Pre-test Probability: Male >55 presenting to ED with chest pain: ~30% baseline for ACS

Applying Likelihood Ratios:

Radiation to left arm: LR+ 2.3 → probability ~50% Diaphoresis: LR+ 2.0 → probability ~65% ST elevation in V1-V4: LR+ >10 → probability >95% Eliminating alternatives: A. Stable angina — does not cause ST elevation; resolves with rest C. Pericarditis — shows diffuse ST elevation, not localized to V1-V4 D. Pulmonary embolism — shows right heart strain, not anterior ST elevation

The localized anterior ST elevation with classic symptoms points to LAD territory infarction.

The answer is B. Acute anterior STEMI

Post-test probability: >95%. This patient requires immediate cardiac catheterization for primary PCI. The ST elevation in V1-V4 indicates LAD occlusion with anterior wall myocardial infarction.


🔧 Technical Specifications

Attribute Value
Base Model Qwen/Qwen3-8B
Base Pretraining 36 trillion tokens, 119 languages, explicit STEM reasoning stages
Architecture Transformer with native reasoning support
Parameters 8 Billion
Precision BF16 (full precision)
Context Length 4,096 tokens (training)
Fine-tuning Method QLoRA (rank 128, alpha 256)
Training Hardware NVIDIA H100 80GB
Training Duration 16 hours (main) + 0.3 hours (stabilizer)
Training Data ~92,000 curated examples (composition proprietary)
Evaluation lm-evaluation-harness v0.4.11

🔀 Parallel Architecture: V8 (DeepSeek-R1 Base)

The identical Pentabrid training pipeline applied to DeepSeek-R1-Distill-Llama-8B produces V8 at 64.5% average — confirming the methodology transfers across architectures. The Qwen3-8B base stronger pretraining foundation (36T vs 15T tokens) acts as a multiplier on training effectiveness.


🔒 Intellectual Property and Access

IP Protection Status

This model, its training methodology, and associated datasets are protected intellectual property.

IP Application 2442 Filed with the United Arab Emirates University (UAEU) Research and Sponsored Projects Office Intellectual Property Unit

All rights reserved. Unauthorized commercial use, reproduction of methodology, or derivative works are prohibited under the CC-BY-NC-ND-4.0 license and applicable intellectual property law.

Requirements for Access

Requirement
Verifiable real name
Institutional affiliation (university, hospital, research organization)
Clear research purpose with specific use case
Agreement to non-commercial, research-only use
Acknowledgment of UAEU IP protection

Requests that will be declined

Reason
Anonymous accounts
Vague or missing research justification
Commercial deployment intentions
Clinical decision-making applications

🏛️ Institutional Details

United Arab Emirates University College of Medicine and Health Sciences Al Ain, Abu Dhabi, UAE

IP Contact:


💼 Commercial Licensing and Enterprise Partnerships

We welcome inquiries from healthcare technology companies, pharmaceutical organizations, hospital systems, and EHR vendors interested in deploying clinical reasoning AI at scale.

What We Offer

Offering Description
Technology Licensing License the Pentabrid methodology and trained models for commercial deployment
Custom Fine-Tuning Specialty-specific models (cardiology, oncology, emergency medicine, etc)
Enterprise Integration API deployment, EHR integration, clinical decision support systems
Research Collaboration Joint development with academic medical centers and health tech companies
White-Label Solutions Branded clinical reasoning AI for your platform

Commercial Contact

All commercial licensing, partnership, and acquisition inquiries should be directed to:

UAEU Intellectual Property Unit Research and Sponsored Projects Office United Arab Emirates University

Or contact the developer via UAEU website helpdesk uaeu.ac.ae :

  • Dr Adnan Agha, College of Medicine and Health Sciences, United Arab Emirates University

⚠️ Intended Use and Safety

✅ Intended Use

Use Case Description
Clinical Research Analyzing patterns in de-identified medical data
Medical Education Simulating diagnostic scenarios for training
AI Benchmarking Evaluating against other medical reasoning systems
Methodology Research Studying Bayesian reasoning in AI systems

❌ Limitations & Warnings

⚠️ NOT A MEDICAL DEVICE — This model is NOT a licensed medical professional and must NOT be used to provide medical advice, diagnosis, or treatment to real patients.

Limitation Description
Hallucination Risk Like all LLMs, outputs may contain fabricated information
No Real-Time Knowledge Training data has a cutoff date
Bias May reflect biases present in medical literature
Validation Required All outputs must be verified by qualified professionals
Not FDA/CE Approved Not cleared for clinical or diagnostic use
No Derivative Works Must not be used to create derivative models (ND license term)

📖 Citation

@misc{agha2025diagnosticreasoningq3,
  author       = {Agha, Adnan},
  title        = {Diagnostic-Reasoning-Q3: Pentabrid Clinical Reasoning Model},
  year         = {2025},
  publisher    = {Hugging Face},
  institution  = {United Arab Emirates University, College of Medicine \& Health Sciences},
  howpublished = {\url{https://huggingface.co/Clinical-Reasoning-Hub/Diagnostic-Reasoning-Q3X1}},
  note         = {IP Application \#2442, UAEU. Fine-tuned from Qwen3-8B using 
                  evidence-based Pentabrid clinical reasoning methodology.}
}
Downloads last month
14
Safetensors
Model size
8B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Clinical-Reasoning-Hub/Diagnostic-Reasoning-Q3X1

Base model

Qwen/Qwen3-8B-Base
Finetuned
Qwen/Qwen3-8B
Finetuned
(991)
this model