Modeling Diagnostic Label Correlation for Automatic ICD Coding

Abstract

Given the clinical notes written in electronic health records (EHRs), it is challenging to predict the diagnostic codes which is formulated as a multi-label classification task. The large set of labels, the hierarchical dependency, and the imbalanced data make this prediction task extremely hard. Most existing work built a binary prediction for each label independently, ignoring the dependencies between labels. To address this problem, we propose a two-stage framework to improve automatic ICD coding by capturing the label correlation. Specifically, we train a label set distribution estimator to rescore the probability of each label set candidate generated by a base predictor. This paper is the first attempt at learning the label set distribution as a reranking module for medical code prediction. In the experiments, our proposed framework is able to improve upon best-performing predictors on the benchmark MIMIC datasets. 1

Hierarchical Text Classification and Its Foundations: A Review of Current Research

“…Thus, thanks to the increased availability of TC datasets that integrate hierarchical structure in their labels, as well as a general interest in industrial applications that utilize TC, a large number of new methods for HTC have been proposed in recent years. Indeed, HTC has many practical applications beyond classic TC, such as International Classification of Diseases (ICD) medical coding [5,6], legal document concept labeling [7], patent labeling [8], IT ticket classification [9], and more.…”

“…Secondly, errors in the upper levels of the hierarchy are inherently worse (e.g., misclassifying "football" as "rugby" is comparatively better than misclassifying "sport" as "food"). These considerations also make sense when considering real-world applications of HTC, such as ICD coding [5,6] and legal document concept labeling [7]. A better mistake entails that most ancestor nodes in the prediction path were correct, meaning that most of the macro categorizations of the sample were accurate.…”

“…Tsai et al [6] tackle the task of automatic ICD coding in medical settings, which requires multilabel classification of clinical nodes in hierarchically dependent diagnostic codes. The authors propose to pair a base predictor (responsible for the generation of top-k most probable label sets) with a re-ranking step.…”

Text2Tree: Aligning Text Representation to the Label Tree Hierarchy for Imbalanced Medical Classification

“…Medical text classification is widely recognized as an urgent yet challenging problem due to its extremely imbalanced data distribution, large variety of rare labels (Johnson et al, 2016;Ziletti et al, 2022), and complicated label relationship (Tsai et al, 2021;Vu et al, 2021). Various downstream clinical tasks have been derived from this problem, including ICD coding (Mullenbach et al, 2018;Yuan et al, 2022;Yang et al, 2022) and automated diagnosis (Chen et al, 2020b), showcasing its potential values in modern clinical practice with machine learning approaches (Berner, 2007).…”