ReCLLaMA: A Reasoning-Centered LLM Agent for Medical Diagnosis

A modular neuro-symbolic framework that combines Large Language Models, Biomedical Knowledge Graphs, and Symbolic Reasoning for interpretable medical diagnosis.

---

Overview

ReCLLaMA is a lightweight and interpretable clinical reasoning system that integrates:

• Knowledge Extraction from free-text patient narratives using LLMs
• Knowledge Alignment from symptoms/procedures to ICD-9 and biomedical entities
• Knowledge Reasoning via multi-hop symbolic inference over biomedical knowledge graphs

Unlike black-box diagnosis models, ReCLLaMA provides:

• Transparent reasoning paths
• Confidence-aware predictions
• Structured ICD-9 outputs
• Modular and efficient deployment

---

Demo

Interactive Medical Diagnosis Demo

Watch Full Demo Video

If GitHub does not preview .mov, download locally and open it.

---

Project Structure

ReCLLaMA/
├── knowledge_extraction/          # Symptom / entity extraction modules
├── knowledge_alignment/          # ICD-9 mapping + embedding alignment
├── reasoner/                     # Multi-hop KG symbolic reasoning engine
├── Recllama_main.py              # Main Streamlit demo
├── Recllama_ablation.py          # Ablation experiments
├── TruthValue.py                 # Confidence estimation
├── result_evaluation.ipynb       # Evaluation notebook
├── rf_model.pkl                  # Random Forest alignment model
├── node2vec_embeddings.txt       # KG node embeddings
├── ICD9_symptom_mapping.csv      # Symptom-to-ICD9 mapping
├── D_ICD_DIAGNOSES.csv           # ICD9 descriptions
├── patient_descriptions_diagnoses.csv
├── questions.csv
├── demo.mov                      # Demo video

---

Installation

Install dependencies:

pip install pandas numpy scikit-learn torch transformers streamlit

---

Quick Start

Run the interactive demo:

streamlit run Recllama_main.py

Or run directly:

python Recllama_main.py