hamsaram/GenMedX-Adapter

---

library_name: peft base_model: BioMistral/BioMistral-7B language: - en pipeline_tag: text-generation tags: - medical - healthcare - triage - risk-prediction - biomistral - rag - fine-tuned license: apache-2.0

🏥 GenMedX: Emergency Triage & Risk Prediction Adapter

GenMedX is a specialized medical AI system designed to assist in Emergency Department (ED) triage. It utilizes a fine-tuned BioMistral-7B model combined with Retrieval-Augmented Generation (RAG) to analyze patient complaints and vital signs, predicting risk levels and recommending diagnostic tests.

📊 Model Details

• Model Name: GenMedX-Adapter
• Base Model: BioMistral/BioMistral-7B (A Mistral-based model further pre-trained on PubMed).
• Architecture: - LLM: 4-bit Quantized LLaMA architecture (via bitsandbytes and peft).
  • RAG System: Uses sentence-transformers/all-MiniLM-L6-v2 to retrieve similar historical cases from a vector database.
• Task: Medical Risk Stratification, Clinical Reasoning, and Diagnostic Test Recommendation.
• Fine-Tuning Method: QLoRA (Quantized Low-Rank Adaptation).

🎯 Intended Use

This model is intended for research and academic demonstration purposes only. It is designed to simulate an "AI Medical Assistant" that can:

1. Analyze Vitals: Detect abnormalities (Tachycardia, Hypoxia, Hypertension, etc.).
2. Predict Risk: Classify patients as High, Medium, or Low risk based on ESI (Emergency Severity Index) standards.
3. Recommend Tests: Suggest immediate diagnostic actions (e.g., ECG, Troponin, CT Scan).
4. Retrieve Context: Compare the current patient with similar historical cases using RAG.

⚠️ Limitations & Ethical Warnings

This model is NOT a doctor.

• Clinical Use: Do not use this model for real-world medical decision-making without human supervision. It is a decision-support tool, not a replacement for a clinician.
• Hallucinations: Like all Large Language Models (LLMs), GenMedX can generate plausible-sounding but incorrect information.
• Bias: The model's outputs are dependent on the quality of the training data. If the data contains biases regarding demographics or disease prevalence, the model may reproduce them.
• Privacy: Ensure no PII (Personally Identifiable Information) is fed into the model during inference.

⚙️ Training Data & Procedure

Data

The model was fine-tuned on a preprocessed emergency dataset (train_preprocessed.csv) containing:

• Inputs: Patient Chief Complaints (e.g., "Chest pain radiating to jaw") and Vital Signs (HR, RR, BP, Temp, O2).
• Targets: Risk Labels (High/Low), Acuity Scores, and Clinician Notes.

Training Hyperparameters

• LoRA Rank (r): 64
• LoRA Alpha: 16
• Dropout: 0.1
• Quantization: 4-bit (NF4)
• Optimizer: PagedAdamW
• Learning Rate: 2e-4
• Scheduler: Cosine

🚀 How to Use (Python)

To use this model, you need the peft and transformers libraries.

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
from peft import PeftModel

# 1. Configuration
base_model_id = "BioMistral/BioMistral-7B"
adapter_id = "hamsaram/GenMedX-Adapter"

# 2. Load Base Model (4-bit for efficiency)
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16,
)

model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    quantization_config=bnb_config,
    device_map="auto",
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained(base_model_id)

# 3. Load GenMedX Adapter
model = PeftModel.from_pretrained(model, adapter_id)

# 4. Inference
prompt = "[INST] Patient has chest pain and HR 120. Assess risk. [/INST]"
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")

outputs = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))