Scholarlink

Objectives <p>This study aimed to construct a machine learning (ML) model to facilitate non-invasive identification of moderate-to-severe intrahepatic venovenous shunts (MS-IHVS) before hepatic venous pressure gradient (HVPG) measurement.</p> Methods <p>A total of 605 patients receiving HVPG measurement were retrospectively included for model development. Two validation cohorts were used: an FN-validation cohort (<i>n</i> = 194; retrospectively enrolled) with both wedged hepatic venous pressure (WHVP) and fine-needle (21–22G) portal venous pressure (FN-PVP) measurements, and a TIPS-validation cohort (<i>n</i> = 92; prospectively enrolled with WHVP and direct PVP measurements. Nine ML and eleven stacking ensemble models were developed. Performance was evaluated using area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, PPV, NPV, F1-score, calibration plots, confusion matrices, and decision-curve analysis. The Shapley Additive Explanation was used for model interpretation.</p> Results <p>Four predictors were selected for model development, including portal vein thrombosis, spontaneous portosystemic shunts, prothrombin time, and number of endoscopic treatments. Across all single and stacking models, the stacking ensemble model combining Random Forest (RF) and Gradient Boosting (GB) performed best (AUC 0.907, 95%CI 0.858–0.945). A Comprehensive Rank also placed RF + GB first. AUCs were 0.781 in FN-validation and 0.889 in TIPS-validation. Model-based risk stratification showed markedly reduced WHVP–PVP/FN-PVP agreement and correlation (intraclass correlation coefficient for agreement 0.314 and 0.143; <i>r</i> = 0.579 and 0.347) in high-risk group compared with low-risk group.</p> Conclusions <p>This RF + GB stacking model effectively identified the risk of MS-IHVS in patients with sinusoidal PHT, providing a powerful tool for early detection and avoiding unnecessary invasive examination.</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

A noninvasive model for predicting intrahepatic venovenous shunts before HVPG measurement: a multiple cohort study

  • 期刊论文

  • 发表日期：

    2026年3月5日

查看全文
  • Rufeng Chen,
  • Li Ma,
  • Yaozu Liu,
  • Wen Zhang,
  • Minjie Yang,
  • Jiaze Yu,
  • Yongjie Zhou,
  • Lingxiao Liu,
  • Junan Lin,
  • Qipeng Wang,
  • Zhiping Yan,
  • Jingqin Ma,
  • Jianjun Luo

摘要

Objectives

This study aimed to construct a machine learning (ML) model to facilitate non-invasive identification of moderate-to-severe intrahepatic venovenous shunts (MS-IHVS) before hepatic venous pressure gradient (HVPG) measurement.

Methods

A total of 605 patients receiving HVPG measurement were retrospectively included for model development. Two validation cohorts were used: an FN-validation cohort (n = 194; retrospectively enrolled) with both wedged hepatic venous pressure (WHVP) and fine-needle (21–22G) portal venous pressure (FN-PVP) measurements, and a TIPS-validation cohort (n = 92; prospectively enrolled with WHVP and direct PVP measurements. Nine ML and eleven stacking ensemble models were developed. Performance was evaluated using area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, PPV, NPV, F1-score, calibration plots, confusion matrices, and decision-curve analysis. The Shapley Additive Explanation was used for model interpretation.

Results

Four predictors were selected for model development, including portal vein thrombosis, spontaneous portosystemic shunts, prothrombin time, and number of endoscopic treatments. Across all single and stacking models, the stacking ensemble model combining Random Forest (RF) and Gradient Boosting (GB) performed best (AUC 0.907, 95%CI 0.858–0.945). A Comprehensive Rank also placed RF + GB first. AUCs were 0.781 in FN-validation and 0.889 in TIPS-validation. Model-based risk stratification showed markedly reduced WHVP–PVP/FN-PVP agreement and correlation (intraclass correlation coefficient for agreement 0.314 and 0.143; r = 0.579 and 0.347) in high-risk group compared with low-risk group.

Conclusions

This RF + GB stacking model effectively identified the risk of MS-IHVS in patients with sinusoidal PHT, providing a powerful tool for early detection and avoiding unnecessary invasive examination.