Dynamic landmark models predicting native liver survival after kasai portoenterostomy: a bicentric study

Native liver survival (NLS) after Kasai portoenterostomy (KPE) for biliary atresia (BA) is highly heterogeneous, and clinically practical approaches that update prognostic estimates across prespecified postoperative follow-up time points remain limited.

In this retrospective bicentric cohort study (training, n = 168; validation, n = 80), Cox proportional hazards landmark models were developed at five prespecified time points (preoperative and 2 weeks, 1, 3, and 6 months post-KPE) using a LASSO/Boruta dual-algorithm intersection approach with variance inflation factor (VIF)-based collinearity pruning. Internal validation used 500-replicate bootstrap resampling; external validation was performed in an independent cohort.

Preoperative and 2-week models showed limited discrimination (C-index ≤0.576). The 3-month two-variable model (TBA + TBil clearance ratio [TBil-CR]) achieved a training C-index of 0.812 (optimism 0.003) and an external C-index of 0.778 (95% CI 0.679-0.878). Risk stratification yielded a 35-percentage-point difference in 2-year cumulative event rates (41.7% vs. 7.0%; log-rank P < 0.001), consistently reproduced in the validation cohort. TBil-CR HR declined progressively across postoperative landmarks, and its single-variable C-index exceeded the jaundice clearance criterion by 0.103 at 3 months.

A parsimonious two-variable landmark model incorporating TBil-CR and TBA at 3 months post-KPE accurately predicts long-term NLS in children with BA, outperforming traditional binary bilirubin thresholds.