Gamma-glutamyl transpeptidase has different efficacy on biliary atresia diagnosis in different hospital patient groups: an application of machine learning approach

Abstract

Objectives: Gamma-glutamyl transpeptidase (GGT) is the most widely used biomarker in the early diagnosis of biliary atresia (BA), but its diagnostic efficacy is questionable in different sub-populations. We aim to identify subgroups that are defined by specific variables with cut-offs and can significantly affect the diagnostic efficacy of GGT for detecting BA.

Methods: Clinical data from 1273 infants with neonatal obstructive jaundice (NOJ) diagnosed between January 2012 and December 2017 at the Children's Hospital of Fudan University were enrolled, reviewed, and analyzed. Random forest-based Virtual Twins method was used to identify potential subgroups.

Results: Hemoglobin (HGB) and fasting gallbladder filling were selected as defining variables. The diagnostic efficacy of GGT was significantly better (AUC = 0.855) for patients with hemoglobin (HGB) ≤ 105 g/L and a gallbladder that was not or poorly filled. Diagnostic efficacy was worst in the subgroup defined by HGB > 105 g/L (AUC = 0.722). The inclusion of interaction terms between GGT and the subgroups in a logistic regression model significantly improved (p = 0.002) prediction performance.

Conclusions: This study provides evidence that the diagnostic efficacy of GGT can differ significantly across different subgroups. Therefore, a GGT diagnostic result should be interpreted cautiously when patients belong to subgroups with low diagnostic efficacy. The development of a prediction model and/or clinical diagnostic pathway for early detection of BA should also account for the heterogeneous diagnostic efficacy of GGT.

Keywords: Biliary atresia; Gamma-glutamyl transpeptidase; Machine learning; Subgroup analysis.