A Novel Method to Metigate Demographic and Expert Bias in ICD Coding with Causal Inference

TL;DR

This work provides a novel causality-based interpretation in ICD Coding that models make predictions by three distinct pathways and shows that DECI outperforms state-of-the-art models, offering a significant advancement in accurate and unbiased ICD coding.

Abstract

ICD(International Classification of Diseases) coding involves assigning ICD codes to patients visit based on their medical notes. Considering ICD coding as a multi-label text classification task, researchers have developed sophisticated methods. Despite progress, these models often suffer from label imbalance and may develop spurious correlations with demographic factors. Additionally, while human coders assign ICD codes, the inclusion of irrelevant information from unrelated experts introduces biases. To combat these issues, we propose a novel method to mitigate Demographic and Expert biases in ICD coding through Causal Inference (DECI). We provide a novel causality-based interpretation in ICD Coding that models make predictions by three distinct pathways. And based counterfactual reasoning, DECI mitigate demographic and expert biases. Experimental results show that DECI outperforms state-of-the-art models, offering a significant advancement in accurate and unbiased ICD coding.

Figures (3)

• Figure 1: (a) An example illustrates that model overlooks the medical note and makes predictions rely on demographic factors. (b) Example of irrelevant expert in other department assign false ICD codes.
• Figure 2: (a) The causal graph for ICD Coding. $D$, $T$, $K$, $L$, $E_r$ and $E_\textit{ir}$ represent demographic factors, the medical note, aggregated knowledge, label, relevant expert and irrelevant expert respectively. DECI emphasizes the aggregated knowledge causal reasoning with relevant expert(blue), simultaneously eliminating effects from demographic factors (green) and irrelevant expert(red). (b) The framework of proposed DECI method. The degree of MoE shading assigned to individual experts denotes the varying weights associated with them.
• Figure 3: A representative instance where our proposed method DECI outputs correct veracity prediction while baselines MSMN make mistakes. The words highlighted in green represent terms found about advanced age in medical note.