初診外来肝臓専門外来における最適化XGBoostモデルを用いた肝硬変における肝不全の早期予測

Hepatic decompensation represents a critical transition in cirrhosis, leading to increased morbidity, mortality and healthcare utilisation. Identifying patients at risk of decompensation remains a clinical challenge. We aimed to develop and validate XGBoost models to predict hepatic decompensation at multiple time points using clinical data available at a patient's initial outpatient hepatology visit.

We conducted a retrospective cohort study including 2208 adult patients with cirrhosis or its complications seen in hepatology clinics between 1985 and 2022. Patients were classified as compensated or decompensated based on a keyword search of the Philip and Ellie Kives Clinical Database, with decompensation dates confirmed by chart review. Sixteen routinely available variables including demographics, biochemical parameters and disease aetiology were used as predictors. Logistic regression and XGBoost models were trained to predict hepatic decompensation at 1, 3, 5 and 10 years, with a random 20% holdout test set used for validation. XGBoost models were tuned to optimise the precision-recall area under the curve (PR-AUC). Performance was evaluated using AUROC, precision, recall and F1 scores.

XGBoost models outperformed logistic regression at most time points, demonstrating strong performance at 3 and 5 years. Recall was 0.42, 0.98, 0.98 and 0.82 at 1, 3, 5 and 10 years respectively. Corresponding AUROC values were 0.85, 0.88, 0.81 and 0.89.

Optimised XGBoost models demonstrated robust predictive accuracy for medium- and long-term hepatic decompensation among patients with compensated cirrhosis. These models may support early risk stratification and enable personalised management strategies to prevent clinical deterioration.