Code for the paper "Mitigating Multi-Label Imbalance in ICD Coding: A Comprehensive Evaluation of Loss Functions and Threshold Optimization."

Code for the paper “Mitigating Multi-Label Imbalance in ICD Coding: A Comprehensive Evaluation of Loss Functions and Threshold Optimization.” This repository contains the official implementation for long-tailed multi-label ICD-10 coding on MIMIC-IV, including a range of imbalance-aware loss functions (L0–L4 taxonomy such as BCE, Focal, CB, DB, RAL, and COMIC), global/per-group/per-label threshold optimization, and end-to-end training and evaluation pipelines. It also provides preprocessing and label-grouping utilities together with configuration files to ensure reproducibility, enabling transparent comparison of loss-level and decision-level calibration strategies for imbalanced medical coding.