Scholarlink

Background <p>Biliary atresia (BA), a rare but life-threatening neonatal cholestatic disorder, necessitates Kasai portoenterostomy (KPE) as the primary surgical intervention. However, despite timely KPE within the recommended 60-day window, approximately 60% of infants ultimately require liver transplantation before age 2. Current prognostic tools lack the capacity to predict long-term native liver survival (NLS) using early perioperative biomarkers.</p> Methods <p>The study subjects were BA patients treated by a senior KPE surgeon at our center (93 cases) and another senior KPE surgeon at Jiangxi Children’s Hospital (61 cases) from January 2017 to December 2022. Clinical and pathological follow-up data were collected. Five traditional machine learning (ML) algorithms (LRM, DET, MLP, SVC, RF) were trained on preoperative, intraoperative, and postoperative biomarkers up to 3 months post-KPE, encompassing biochemical profiles. The best-performing RF model was further enhanced using transfer learning (TL).</p> Results <p>For a combined cohort of 154 patients, traditional statistical analysis revealed significant differences in perioperative factors between NLS and non-NLS groups, with multivariate logistic regression identifying surgical age (<i>P</i> = 0.029), globular proteins at 1 week postoperatively (GLO_1W, <i>P</i> &lt; 0.001), aspartate aminotransferase at 1 month postoperatively (AST_1M, <i>P</i> = 0.007), albumin at 3 months postoperatively (ALB_3M, <i>P</i> = 0.029), gamma-glutamyl transpeptidase at 3 months postoperatively (GGT_3M, <i>P</i> &lt; 0.001), direct bilirubin at 3 months postoperatively (DBIL_3M, <i>P</i> = 0.005), and prealbumin at 3 months postoperatively (PALB_3M, <i>P</i> = 0.004) as independent risk factors. The five traditional ML models showed varying performance, with RF demonstrating the highest average AUC in the training set (AUC = 0.795, 95% CI 0.709–0.881). The TL model, built on the RF base, achieved superior performance in the training set (AUC = 0.888, 95% CI 0.819–0.943) and validation set (AUC = 0.856, 95% CI 0.738–0.947). GGT_3M, ALB_3M, DBIL_3M, DBIL_1M and GLO_1W were identified as the most critical predictive features. A TL-based clinical decision support system was deployed, providing dynamic visualization of individualized 24-month NLS trajectories.</p> Conclusion <p>Early perioperative biochemical data-driven ML models, enhanced by transfer learning, enable accurate prediction of 2-year NLS in BA. GGT_3M, ALB_3M, DBIL_3M, DBIL_1M, and GLO_1W are among the most influential predictors for 2-year NLS after KPE procedure.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Transfer learning prediction of 2-year native liver survival in biliary atresia following Kasai portoenterostomy using early perioperative biochemical data up to 3 months postoperatively

  • 期刊论文

  • 发表日期：

    2026年7月27日

查看全文
  • Chuanping Xie,
  • Wenqing Yan,
  • Yong Zhao,
  • Shuangshuang Li,
  • Junmin Liao,
  • Yanan Zhang,
  • Kaiyun Hua,
  • Yichao Gu,
  • Peize Wang,
  • Jie Sun,
  • Yuyan Jin,
  • Zhaozhou Liu,
  • Dayan Sun,
  • Dingding Wang,
  • Jinshi Huang

摘要

Background

Biliary atresia (BA), a rare but life-threatening neonatal cholestatic disorder, necessitates Kasai portoenterostomy (KPE) as the primary surgical intervention. However, despite timely KPE within the recommended 60-day window, approximately 60% of infants ultimately require liver transplantation before age 2. Current prognostic tools lack the capacity to predict long-term native liver survival (NLS) using early perioperative biomarkers.

Methods

The study subjects were BA patients treated by a senior KPE surgeon at our center (93 cases) and another senior KPE surgeon at Jiangxi Children’s Hospital (61 cases) from January 2017 to December 2022. Clinical and pathological follow-up data were collected. Five traditional machine learning (ML) algorithms (LRM, DET, MLP, SVC, RF) were trained on preoperative, intraoperative, and postoperative biomarkers up to 3 months post-KPE, encompassing biochemical profiles. The best-performing RF model was further enhanced using transfer learning (TL).

Results

For a combined cohort of 154 patients, traditional statistical analysis revealed significant differences in perioperative factors between NLS and non-NLS groups, with multivariate logistic regression identifying surgical age (P = 0.029), globular proteins at 1 week postoperatively (GLO_1W, P < 0.001), aspartate aminotransferase at 1 month postoperatively (AST_1M, P = 0.007), albumin at 3 months postoperatively (ALB_3M, P = 0.029), gamma-glutamyl transpeptidase at 3 months postoperatively (GGT_3M, P < 0.001), direct bilirubin at 3 months postoperatively (DBIL_3M, P = 0.005), and prealbumin at 3 months postoperatively (PALB_3M, P = 0.004) as independent risk factors. The five traditional ML models showed varying performance, with RF demonstrating the highest average AUC in the training set (AUC = 0.795, 95% CI 0.709–0.881). The TL model, built on the RF base, achieved superior performance in the training set (AUC = 0.888, 95% CI 0.819–0.943) and validation set (AUC = 0.856, 95% CI 0.738–0.947). GGT_3M, ALB_3M, DBIL_3M, DBIL_1M and GLO_1W were identified as the most critical predictive features. A TL-based clinical decision support system was deployed, providing dynamic visualization of individualized 24-month NLS trajectories.

Conclusion

Early perioperative biochemical data-driven ML models, enhanced by transfer learning, enable accurate prediction of 2-year NLS in BA. GGT_3M, ALB_3M, DBIL_3M, DBIL_1M, and GLO_1W are among the most influential predictors for 2-year NLS after KPE procedure.