Predicting better: the first step to building a personalized model for the prediction of Kasai operation outcome in biliary atresia (BA) patients

Abstract

Background: Performing Kasai within 60-75days is critical to successful outcome for BA. Despite early operation, predictors for failed outcome are lacking. We used mixed graphical models to integrate biomedical datasets to explore genetic factors predisposing to failed Kasai outcome. Materials and methods: A causal model was trained by clinical variables (gender/age-at-Kasai/Kasai outcome/ allergy) and 35 BA-associated single nucleotide polymorphisms (SNPs) in a Chinese cohort (181 BA, 481 controls) with ”causalMGM” algorithm on 60 BA (with complete clinical information). Cross-validated using ten bootstrapped sub-samples. A causal relationship was generated (Figure 1). A new set of 39 BA was used for model testing. Results: The model captured both direct/indirect relationships between age and Kasai outcome. Three SNPs (rs749619/rs2510389/rs11101722) were causal to outcome: two are also directly related to age, with predicting sensitivity 30%, specificity 84%. Sensitivity was improved to 50% (specificity 76%) when the SNP only related to Kasai outcome was added as a predictor. All except one correctly predicted failed Kasai outcome were operated on early (30-66days). Conclusion: The first ”Kasai outcome prediction model” was generated to predict outcome in ”early operation”. Expanding the training cohort can further improve sensitivity/specificity. This facilitates clinical decision for BA patients such as post-operative adjunctive steroids use, closer post-operative follow-ups.