Stability and robustness of explainable artificial intelligence methods in clinical prediction models: Protocol for a systematic review

Abstract

Explainable artificial intelligence (XAI) methods such as SHAP, LIME and permutation feature importance are widely used to justify machine learning models for clinical prediction, yet an explanation is useful only if it is reproducible. Explanations may change when a model is refitted on resampled data, when inputs are slightly perturbed, when a different but equally accurate model is chosen, or when a different explanation method is applied. It is not known how often published clinical prediction studies assess these forms of stability, how they measure them, or how stable the reported explanations are. This protocol describes a systematic review that will address these questions. Scopus, Web of Science Core Collection, PubMed/MEDLINE and IEEE Xplore will be searched for primary studies published from January 2016 onwards that develop or validate a clinical prediction model using machine learning and report a quantitative assessment of the stability, robustness, consistency or agreement of feature-attribution explanations. Two reviewers will independently screen records and full texts, extract data with a form based on the CHARMS checklist, and assess risk of bias with PROBAST. Outcomes will include the type of stability assessed, the metrics used, the magnitude of reported stability, and factors associated with instability, such as sample size, feature correlation and model family. Results will be synthesised narratively and, where at least five studies report comparable metrics, summarised quantitatively. The review will be reported according to PRISMA 2020 and PRISMA-S and is expected to inform reporting standards for explanations in clinical machine learning.