Towards Unbiased Evaluation of Detecting Unanswerable Questions in EHRSQL
This addresses a critical issue for medical professionals by improving the trustworthiness of EHR QA system evaluations, though it is incremental as it focuses on dataset refinement rather than novel model development.
The authors identified a data bias in the EHRSQL dataset where unanswerable questions could be detected using N-gram patterns, compromising evaluation reliability, and proposed a debiasing method by adjusting dataset splits, which they demonstrated effectively mitigated this bias on the MIMIC-III dataset.
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.