🔒 Access Restricted: Clinical Research Only

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

This is a clinical medical AI model developed for research purposes. Due to the sensitive nature of medical AI, access requires manual review. Please provide your credentials and research justification below.

🩺 Diagnostic-Medicine-R1

A Bayesian Clinical Reasoning AI Model

Developed by the Clinical Reasoning Lab

🎯 Overview

Diagnostic-Medicine-R1 represents 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 demonstrates that small language models (8B parameters) can achieve competitive medical reasoning performance (if trained like medical curriculum of students) through methodological innovation rather than scale.

✨ Key Innovations

Innovation	Description
Evidence-Based Reasoning	> 500 likelihood ratios extracted 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

📊 Training Methodology: The Three Pillars

Our training methodology is built on three core components that mirror how expert clinicians develop their diagnostic skills:

1️⃣ Numbers: Evidence-Based Likelihood Ratios

The model learns to quantify diagnostic certainty using likelihood ratios (LRs) from peer-reviewed sources:

LR+ > 10: Strong rule-IN (+45% probability shift)
LR+ 5-10: Moderate rule-in (+30% probability shift)
LR- < 0.1: Strong rule-OUT (-45% probability shift)

2️⃣ Logic: Diagnostic Frameworks

Structured approaches to common clinical presentations including:

Pivot points (findings that dramatically change probability)
Must-not-miss diagnoses
Systematic differential generation

3️⃣ Behavior: Clinical Decision Patterns

Tacit knowledge of experienced clinicians including:

Golden Rules (e.g., "Always test for pregnancy before imaging")
Cannot-Miss Diagnoses
Structured clinical reasoning worksheets

📈 Performance Metrics

Metric	Before Training	After Training
MedMCQA Accuracy	22%	65-75%
Training Loss	4.86	1.0-1.5
Token Accuracy	27%	66-68%

Training Data Composition

🔧 Model Details

Attribute	Value
Base Model	DeepSeek-R1-Distill-Llama-8B
Architecture	Transformers with Chain-of-Thought (CoT) Reasoning
Parameters	8 Billion
Context Length	4,096 tokens
Training Method	QLoRA (4-bit quantization)
LoRA Rank	64
Developer	Clinical Reasoning Lab
License	CC-BY-NC-ND-4.0

Attribution

This model is a fine-tune of DeepSeek-R1-Distill-Llama-8B, which is distilled from Llama-3.1-8B.

Built with Llama
Original DeepSeek-R1 Copyright © 2023 DeepSeek

💡 Example Output

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?

Model: <think>
Let me analyze this case using Bayesian clinical reasoning.

**Key Findings:**
1. 65-year-old male (cardiac risk age)
2. Crushing chest pain radiating to left arm (classic ACS presentation)
3. Diaphoresis (autonomic response to MI)
4. ST elevation in V1-V4 (anterior leads)

**Pre-test Probability:** 
Male >55 with chest pain in ED setting: ~30% baseline for ACS

**Applying Likelihood Ratios:**
- Radiation to left arm: LR+ 2.3 → probability increases to ~50%
- Diaphoresis: LR+ 2.0 → probability increases to ~65%
- ST elevation in V1-V4: LR+ >10 → probability increases to >95%

**Localization:** V1-V4 = anterior wall = LAD territory
</think>

**Diagnosis: Acute Anterior STEMI**

Post-test probability: >95%

This patient requires immediate cardiac catheterization. The ST elevation 
in leads V1-V4 indicates LAD occlusion with anterior wall infarction.

🔒 Access Policy

Due to the sensitive nature of medical AI, access to this model is gated and manually reviewed.

Requirements for Access:

✅ Verifiable real name
✅ Institutional affiliation (university, hospital, research organization)
✅ Clear research purpose with specific use case
✅ Agreement to non-commercial, research-only use

Requests that will be declined:

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

⚠️ Intended Use & Limitations

✅ 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 Approved	Not cleared for clinical use

📄 Documentation

For comprehensive methodology, training data architecture, and technical specifications:

📥 Download Complete AI Development Methodology (DOCX)

Complete AI Development Methodology (DOCX) available on request only

Document Contents:

Executive Summary
Project Objectives & Target Benchmarks
Theoretical Framework (Numbers, Logic, Behavior)
Training Data Architecture (6-tier system)
Training Methodology & Catastrophic Forgetting Solution
Training Phases & Results
Technical Implementation Details
Key Innovations
Conclusion & Future Work
References

📚 References

Medical texbooks (finely distilled knowledge as per medical curriculum of undergraduate medical students)
DeepSeek-AI. DeepSeek-R1: Incentivizing Reasoning Capability in LLMs. arXiv:2401.02954, 2024.
M42-Health. MEDIC: A Comprehensive Framework for Evaluating LLMs in Clinical Applications. arXiv:2409.07314, 2024.
AI4LIFE-GROUP. MedSafetyBench: Evaluating and Developing Safety of Medical AI. NeurIPS 2024.

💼 Commercial Licensing

For commercial licensing inquiries, enterprise deployment, or partnership opportunities, contact: [[email protected]]

The publicly available model is for research only. Commercial licenses with support, customization, and deployment rights are available.


---
## 📖 Citation

If you use this model in your research, please cite:

```bibtex
@misc{diagnostic-medicine-r1-2025,
  author       = {Clinical Reasoning Lab},
  title        = {Diagnostic-Medicine-R1: A Bayesian Clinical Reasoning Model},
  year         = {2025},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Clinical-Reasoning-Hub/Diagnostic-Medicine-R1}},
  note         = {Fine-tuned from DeepSeek-R1-Distill-Llama-8B using evidence-based 
                  Bayesian clinical reasoning methodology}
}

Developed with ❤️ by the Clinical Reasoning Lab

For research and educational purposes only