fafasuper/KE-Coder

Knowledge-Enhanced Generative Framework for Low-Resource Clinical Coding # KE-Coder ## Data-driven Comorbidity Network as Symbolic Constraints: A Knowledge-Enhanced Generative Framework for Low-Resource Clinical Coding ## Overview This repository contains the official implementation of KE-Coder, a knowledge-enhanced generative framework for automated ICD coding. The framework integrates medical definition constraints ($$K_{def}$$), retrieval-augmented case knowledge ($$K_{rag}$$), and cross-modal structural consistency validation to improve the accuracy and reliability of clinical coding under low-resource conditions. The model explicitly targets the multi-label classification of 8 core thoracic diseases and leverages the predicted structural comorbidity matrices to align with real-world clinical concurrent patterns. ## Project Structure ├── data/ # Raw and intermediate processed data (Not tracked by Git) ├── dataset_build/ # Data cleaning, entity matching, and quintuple generation ├── knowledge_base/ # Medical dictionary and definition building modules ├── prompts/ # Prompt templates (System, Baseline, KE-Coder, Few-shot) ├── train/ # Fine-tuning scripts for LLMs (Vanilla & KE-Coder) ├── inference/ # Inference engine for multi-label disease prediction ├── evaluation/ # Metrics calculation (AUC, F1, PCC, MSE) and comorbidity analysis ├── scripts/ # Bash scripts for automating the pipeline ├── results/ # Output directory for predictions and evaluation reports ├── README.md # Project documentation └── requirements.txt # Python environment dependencies ### Dataset Preparation Public Datasets Required :All datasets used in this work are publicly available but require official credentialed access and approval via PhysioNet. The underlying datasets include: MIMIC-III, MIMIC-IV, MIMIC-CXR Due to data privacy and compliance, raw data cannot be directly provided in this repository. You …