Interpretable Sub-phenotype Identification in Acute Kidney Injury

Abstract

Acute kidney injury (AKI) is a life-threatening and heterogeneous syndrome. Timely and etiology-based personalized treatment is crucial. AKI sub-phenotyping can lead to better understanding of the pathophysiology of AKI and help developing more targeted intervention. Current dimensionality reduction and similarity-based clustering for AKI sub-phenotyping suffer from limited interpretability and specificity. To address these limitations, we propose a pattern mining approach with multiobjective evolutionary algorithm (MOEA) for AKI sub-phenotyping. AKI sub-phenotypes are presented as explicit rules, so no post-hoc explanation is needed. Also, our method can search feature subspace efficiently for minor and highly specific sub-phenotypes. We aimed to discover sub-phenotypes for AKI patients against non-AKI patients (AKI vs non-AKI) and moderate-to-severe AKI patients against mild AKI patients (AKI-2/3 vs AKI-1). We identified 174(178) significant sub-phenotypes with average confidence of 0.33(0.33). Our method can assign patients to a sub-phenotype with higher confidence than k-means clustering, with average improvement of 0.20(0.23).