AmareshHebbar/icd10-coder-qwen25-7b-merged

ICD-10 Medical Coder — Qwen2.5-7B

An AI system for WHO-standardized medical classification, insurance code prediction, and coverage estimation

---

What Is This?

ICD-10-Coder is the first model in a long-term initiative — AxisMapper — to build an AI-native insurance intelligence layer for the Indian and global healthcare ecosystem.

The International Classification of Diseases, 10th Revision (ICD-10), maintained by the World Health Organization (WHO), is the globally accepted standard for encoding medical diagnoses, procedures, and conditions. Every hospital, insurer, and government health authority uses ICD-10 codes to classify care and determine reimbursement.

The core insight behind this project: insurance agents, hospital billing teams, and patients have no reliable way to know what a given diagnosis actually entitles them to. Coverage decisions are opaque, rules are fragmented across schemes, and the same condition might be coded five different ways — each triggering a different payout.

This model is the first agent in what will become a Multi-Agent, Mixture-of-Experts (MoE) pipeline — purpose-built to decode that opacity.

---

The Bigger Vision: AxisMapper

"One fine-tuned model per insurance scheme. A shared routing layer. Zero ambiguity for the patient."

India's health insurance landscape spans:

• Ayushman Bharat / PM-JAY — world's largest government-funded health insurance scheme
• Star Health — India's largest standalone health insurer
• ESIC / CGHS — central government employee schemes
• State-level programs — varying eligibility, tariff, and admission rules
• NGO-backed schemes — community-level coverage with entirely different logic

Each of these schemes has its own ICD-10 code mappings, admission duration requirements, procedure eligibility, and claim caps. There is no unified interface to query them all.

AxisMapper's roadmap:

Phase 1 (Now)  → WHO ICD-10 base model (this model)
                 Universal code prediction + coverage logic

Phase 2        → Fine-tune per scheme (StarHealth, PM-JAY, ESIC, etc.)
                 Each model specialises in one insurer's rule set

Phase 3        → MoE Router
                 Given a patient + insurer, route to the right specialist model

Phase 4        → Multi-Agent Pipeline
                 Agent 1: Diagnosis → ICD-10 code
                 Agent 2: Code → Coverage estimate (policy-aware)
                 Agent 3: Coverage + Admission rules → Final claim amount
                 Agent 4: Web search → Real-time tariff / market validation

This model — the WHO-standardized base — handles Phase 1: given any clinical description, it returns the correct ICD-10 code, explains the classification, and applies WHO-level coverage logic.

---

Model Details

Property	Value
Base Model	unsloth/qwen2.5-7b-instruct
Architecture	Qwen2 (decoder-only transformer)
Parameters	~8B
Precision	BF16
Fine-tuning Method	LoRA via Unsloth + HuggingFace TRL
Training Hardware	NVIDIA RTX A5000 (24GB VRAM)
Training Duration	~2 hours
Training Speed	2× faster than standard HF training (via Unsloth)
Experiment Tracking	Weights & Biases (W&B)
Max Sequence Length	2048 tokens
License	Apache 2.0

---

Training Infrastructure

This model was trained using the Unsloth optimization library, which achieves 2× training speed and ~60% VRAM reduction compared to standard HuggingFace fine-tuning — without any loss in model quality.

Training stack:

• unsloth — optimized LoRA fine-tuning engine
• trl (HuggingFace) — SFTTrainer for instruction fine-tuning
• transformers — model loading, tokenization, inference
• wandb — real-time loss curves, learning rate scheduling, gradient tracking

All training runs are logged and reproducible via Weights & Biases. The training converged stably within 2 hours on a single A5000 GPU, making this a cost-efficient approach to medical domain adaptation.

---

What This Model Does

Given a clinical description or patient scenario, this model will:

1. Assign the correct ICD-10 code(s) — primary diagnosis, secondary conditions, procedure codes
2. Explain the WHO classification logic — why this code, what the category means, adjacent codes
3. Estimate WHO-level insurance coverage — standard reimbursement brackets, admission duration requirements, procedure eligibility
4. Flag restrictions — minimum admission days, co-morbidity requirements, pre-authorisation triggers
5. Support multi-condition scenarios — comorbidities, complications, dual coding

Example input:

Patient admitted for acute appendicitis with peritonitis. 
Underwent emergency appendectomy. Admitted for 3 days.
What ICD-10 codes apply and what is the expected insurance coverage?

Example output (truncated):

Primary Code: K35.2 — Acute appendicitis with generalised peritonitis
Procedure Code: 0DTJ4ZZ — Resection of appendix, percutaneous endoscopic approach

WHO Classification: Diseases of the digestive system (K00–K93)
Chapter XI, Block K35-K38 (Diseases of appendix)

Coverage Logic:
- WHO standard: Surgical admission, inpatient required
- Minimum admission: 1–3 days (surgery-dependent)
- Reimbursement class: Major surgery
- Pre-auth: Required for elective; emergency bypass available
- Approximate WHO-tier bracket: ₹35,000–₹75,000 (India tier-2 hospital)

---

Quickstart

Using Transformers (Pipeline)

from transformers import pipeline

pipe = pipeline("text-generation", model="AmareshHebbar/icd10-coder-qwen25-7b-merged")

query = """
Patient presents with Type 2 diabetes mellitus with chronic kidney disease stage 3.
What ICD-10 codes apply? What are the WHO-level insurance implications?
What are the admission requirements for this to be covered?
"""

result = pipe([{"role": "user", "content": query}], max_new_tokens=512)
print(result[0]["generated_text"][-1]["content"])

Using Unsloth (Recommended for inference speed)

from unsloth import FastModel

model, tokenizer = FastModel.from_pretrained(
    model_name="AmareshHebbar/icd10-coder-qwen25-7b-merged",
    max_seq_length=2048,
    load_in_4bit=True,  # Optional: 4-bit for lower VRAM
)

messages = [
    {"role": "system", "content": "You are an expert ICD-10 medical coder with deep knowledge of WHO insurance classification standards."},
    {"role": "user", "content": "Patient: acute MI, stented. 2-day admission. Code and coverage?"}
]

inputs = tokenizer.apply_chat_template(
    messages, tokenize=True, add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.1)
print(tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True))

Using vLLM (Production / High Throughput)

pip install vllm
vllm serve "AmareshHebbar/icd10-coder-qwen25-7b-merged" --max-model-len 2048

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="none")

response = client.chat.completions.create(
    model="AmareshHebbar/icd10-coder-qwen25-7b-merged",
    messages=[
        {"role": "system", "content": "You are an expert ICD-10 coder and insurance analyst."},
        {"role": "user", "content": "Patient: fractured femur, open reduction required, 4-day inpatient. ICD-10 codes and insurance coverage?"}
    ],
    max_tokens=512,
    temperature=0.1,
)
print(response.choices[0].message.content)

Using Ollama (Local / Offline)

# Export to GGUF first (via llama.cpp or Unsloth export)
ollama create icd10-coder -f ./Modelfile
ollama run icd10-coder "Patient: appendicitis, emergency surgery. Code and coverage?"

---

🔌 Integrations Supported

Backend	Status	Use Case
HuggingFace Transformers	✅	Research, prototyping
Unsloth FastModel	✅	Fast inference, fine-tuning
vLLM	✅	Production API, high throughput
SGLang	✅	Structured generation
Ollama	✅	Local / offline deployment
Claude API (Anthropic)	🔌 Planned	Hybrid: ICD-10 code → Claude for coverage analysis
Gemini API (Google)	🔌 Planned	Multi-LLM comparison layer
Web Search (Tavily/Serper)	🔌 Planned	Real-time tariff + hospital rate lookup

---

ICD-10 Coverage

This model has been fine-tuned across all major ICD-10-CM chapters:

Chapter	Description
I (A00–B99)	Infectious and parasitic diseases
II (C00–D49)	Neoplasms
III (D50–D89)	Blood and immune disorders
IV (E00–E89)	Endocrine, nutritional, metabolic
V (F01–F99)	Mental and behavioural disorders
IX (I00–I99)	Circulatory system diseases
X (J00–J99)	Respiratory diseases
XI (K00–K95)	Digestive system diseases
XIII (M00–M99)	Musculoskeletal diseases
XIV (N00–N99)	Genitourinary diseases
XIX (S00–T88)	Injuries, poisonings
XXI (Z00–Z99)	Health status, contact with services

---

Limitations & Intended Use

• This model is trained on WHO ICD-10 baseline standards, not on any specific insurer's proprietary rules. Coverage estimates are indicative, not legally binding.
• Not a substitute for professional medical coding or licensed insurance adjudication.
• Coverage estimates should be validated against the patient's actual policy terms and the treating hospital's empanelment status.
• Future scheme-specific models (Ayushman Bharat, Star Health, etc.) will provide more precise, policy-aware outputs.

---

Links

• GitHub (AxisMapper): https://github.com/amareshhebbar/AxisMapper
• Developed by: AmareshHebbar
• Base model: unsloth/qwen2.5-7b-instruct

---

Citation

@misc{hebbar2025icd10coder,
  title={ICD-10 Coder: A Fine-tuned Qwen2.5-7B for Medical Classification and Insurance Coverage Estimation},
  author={Amaresh Hebbar},
  year={2025},
  publisher={HuggingFace},
  url={https://huggingface.co/AmareshHebbar/icd10-coder-qwen25-7b-merged},
  note={Part of the AxisMapper project: https://github.com/amareshhebbar/AxisMapper}
}

---

_{Built with Unsloth · Trained on A5000 · Tracked with W&B · Part of AxisMapper}