Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LTR-ICD: A Learning-to-Rank Approach for Automatic ICD Coding

Mohammad Mansoori, Amira Soliman, Farzaneh Etminani

TL;DR

This paper addresses the need to preserve ICD code order in automatic coding. It introduces LTR-ICD, a two-module, shared-encoder framework that jointly learns to classify ICD codes and generate an ordered code sequence, optimized via a ranking mechanism. Evaluated on MIMIC-III, the model achieves superior ranking of high-priority codes and better macro and micro F1 metrics compared with state-of-the-art classifiers, demonstrating practical improvements for clinical coding and billing workflows. Limitations include reliance on ICD-9 and a single-hospital dataset, with future work expected to extend to ICD-10/ICD-11, multilingual data, and enhanced interpretability.

Abstract

Clinical notes contain unstructured text provided by clinicians during patient encounters. These notes are usually accompanied by a sequence of diagnostic codes following the International Classification of Diseases (ICD). Correctly assigning and ordering ICD codes are essential for medical diagnosis and reimbursement. However, automating this task remains challenging. State-of-the-art methods treated this problem as a classification task, leading to ignoring the order of ICD codes that is essential for different purposes. In this work, as a first attempt, we approach this task from a retrieval system perspective to consider the order of codes, thus formulating this problem as a classification and ranking task. Our results and analysis show that the proposed framework has a superior ability to identify high-priority codes compared to other methods. For instance, our model accuracy in correctly ranking primary diagnosis codes is 47%, compared to 20% for the state-of-the-art classifier. Additionally, in terms of classification metrics, the proposed model achieves a micro- and macro-F1 scores of 0.6065 and 0.2904, respectively, surpassing the previous best model with scores of 0.597 and 0.2660.

LTR-ICD: A Learning-to-Rank Approach for Automatic ICD Coding

TL;DR

This paper addresses the need to preserve ICD code order in automatic coding. It introduces LTR-ICD, a two-module, shared-encoder framework that jointly learns to classify ICD codes and generate an ordered code sequence, optimized via a ranking mechanism. Evaluated on MIMIC-III, the model achieves superior ranking of high-priority codes and better macro and micro F1 metrics compared with state-of-the-art classifiers, demonstrating practical improvements for clinical coding and billing workflows. Limitations include reliance on ICD-9 and a single-hospital dataset, with future work expected to extend to ICD-10/ICD-11, multilingual data, and enhanced interpretability.

Abstract

Clinical notes contain unstructured text provided by clinicians during patient encounters. These notes are usually accompanied by a sequence of diagnostic codes following the International Classification of Diseases (ICD). Correctly assigning and ordering ICD codes are essential for medical diagnosis and reimbursement. However, automating this task remains challenging. State-of-the-art methods treated this problem as a classification task, leading to ignoring the order of ICD codes that is essential for different purposes. In this work, as a first attempt, we approach this task from a retrieval system perspective to consider the order of codes, thus formulating this problem as a classification and ranking task. Our results and analysis show that the proposed framework has a superior ability to identify high-priority codes compared to other methods. For instance, our model accuracy in correctly ranking primary diagnosis codes is 47%, compared to 20% for the state-of-the-art classifier. Additionally, in terms of classification metrics, the proposed model achieves a micro- and macro-F1 scores of 0.6065 and 0.2904, respectively, surpassing the previous best model with scores of 0.597 and 0.2660.
Paper Structure (18 sections, 9 equations, 4 figures, 4 tables, 1 algorithm)

This paper contains 18 sections, 9 equations, 4 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Works
  3. Methodology
  4. Problem Definition
  5. Pre-trained Language Model
  6. Ranking Algorithm
  7. Experimental Settings
  8. Dataset
  9. Metrics
  10. Implementation details
  11. Training
  12. Inference
  13. Experimental Results
  14. Performance on Evaluation Metrics
  15. Ranking Capabilities of Models: A Cumulative Gain Approach
  16. ...and 3 more sections

Figures (4)

  • Figure 1: Overview of the proposed LTR-ICD framework for medical code prediction and ranking. The model consists of three main components: a classifier module, a generative module, and a ranking algorithm.
  • Figure 2: Ranking algorithm combines classifier and generative outputs to produce final ICD predictions
  • Figure 3: Comparing the performance of the LTR-ICD and PLM-ICD models in terms of cumulative gain across multiple ranking positions.
  • Figure 4: Comparing the impact of diagnosis and procedure code ordering on the performance of the LTR-ICD model, measured by cumulative gain at K using test data.