DOI: 10.1097/js9.0000000000005689 ISSN: 1743-9159

Optimal duration of endoscopic biliary stenting after living donor liver transplantation: a multicenter retrospective cohort study using machine learning–based risk stratification

Ji Won Han, In Seok Lee, Soon Kyu Lee, Jung Hyun Kwon, Soon Woo Nam, Jaejun Lee, Hyun Yang, Si Hyun Bae, Kyu-Hyun Paik, Pil Soo Sung, Jeong Won Jang, Eui Soo Han, Ho Joong Choi, Young Kyoung You, Jong Young Choi
Show PDF Cite

Background:

Anastomotic biliary strictures (ABSs) are among the most common complications after living-donor liver transplantation (LDLT). Endoscopic retrograde biliary drainage (ERBD) is the primary treatment, but the optimal stent duration remains unclear. This study aimed to identify risk factors for ABS recurrence and to develop a machine learning (ML) model to personalize stent removal.

Methods:

Among 1267 LDLT recipients, 173 patients who underwent ERBD stent removal for ABS were analyzed at three tertiary centers (internal cohort n = 131; validation cohort n = 42). Clinical and biochemical variables were examined using both logistic and Cox regression. A logistic regression-based ML-logit classifier that integrates stent duration, number of endoscopic retrograde cholangiopancreatography (ERCP) procedures, and laboratory values at removal was developed and compared with fixed-duration cutoffs.

Results:

The overall recurrence rate was 25.4% (44/173), with a median recurrence time of 297 days. Independent predictors included shorter stent duration [hazard ratio (HR), 3.45; 95% confidence interval (CI), 1.52–7.83; P = 0.003], higher ERCP count (HR, 2.12; P = 0.036), elevated bilirubin (HR, 2.58; P = 0.011), and elevated alkaline phosphatase (HR, 2.21; P = 0.030). The recurrence rate was significantly lower when the stent duration exceeded 12 months (10.7% vs. 47.1%; HR, 4.76; P < 0.001). The ML-logit classifier achieved an area under the receiver operating characteristic curve (AUROC) of 0.901, significantly outperforming thresholds at 12 months (AUROC, 0.688; P = 0.008) and 396 days (AUROC, 0.730; P = 0.012). Model stratification separated patients into low- and high-risk groups, with 5-year recurrence-free survival of 83.4% and 11.4%, respectively, and an HR of 9.06 ( P < 0.001). Model discrimination was broadly similar in the internal and external cohorts, although sensitivity was limited, particularly in the external cohort.

Conclusion:

While maintaining stents for at least 12 months remains an important principle, this proof-of-concept study suggests that the logistic regression-based ML-logit classifier may provide promising risk stratification, potentially helping to reduce unnecessarily prolonged stenting in selected low-risk patients under close surveillance. Prospective validation in larger, independent cohorts is required before clinical implementation.

More from our Archive