Towards Unbiased Evaluation of Detecting Unanswerable Questions in EHRSQL

TL;DR

This paper highlights the data bias in unanswerable questions within the EHRSQL dataset for EHR QA systems and proposes a data split strategy to mitigate this bias, improving evaluation reliability.

Contribution

It identifies a bias caused by N-gram patterns in unanswerable questions and introduces a simple data split adjustment to reduce this bias in EHR QA evaluation.

Findings

Bias due to N-gram patterns in unanswerable questions

Data split adjustment reduces bias impact

Improved evaluation reliability on MIMIC-III

Abstract

Incorporating unanswerable questions into EHR QA systems is crucial for testing the trustworthiness of a system, as providing non-existent responses can mislead doctors in their diagnoses. The EHRSQL dataset stands out as a promising benchmark because it is the only dataset that incorporates unanswerable questions in the EHR QA system alongside practical questions. However, in this work, we identify a data bias in these unanswerable questions; they can often be discerned simply by filtering with specific N-gram patterns. Such biases jeopardize the authenticity and reliability of QA system evaluations. To tackle this problem, we propose a simple debiasing method of adjusting the split between the validation and test sets to neutralize the undue influence of N-gram filtering. By experimenting on the MIMIC-III dataset, we demonstrate both the existing data bias in EHRSQL and the effectiveness of our data split strategy in mitigating this bias.