Towards Unbiased Evaluation of Detecting Unanswerable Questions in EHRSQL

TL;DR

This work identifies a data bias in the EHRSQL benchmark for EHR QA, where unanswerable questions correlate with recurring N-gram patterns, enabling heuristic filtering that inflates evaluation metrics. It introduces a simple debiasing approach by constructing a new validation/test split that confines high-ratio N-gram phrases to the test set, thereby reducing cross-generalization from biased patterns. Through experiments on the MIMIC-III dataset using a text-to-SQL model (T5-base) and uncertainty baselines (entropy, beam score), the authors show that the bias can mislead assessments of model understanding, and that the proposed split yields a more faithful measure of true capability to refrain from answering uncertain queries. The findings advocate for more robust dataset design in healthcare QA benchmarks and motivate extending debiasing methods to additional datasets beyond EHRSQL.

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.

Figures (1)

• Figure 1: The overview of our proposed data split.