文献速递肝病

新型机器学习模型在代谢功能障碍相关脂肪性肝炎伴纤维化中优于无创检测

Novel Machine-Learning Models Outperform Noninvasive Tests in Metabolic Dysfunction-Associated Steatohepatitis With Fibrosis.
PubMed 文献 · Clin Gastroenterol Hepatol · 2025-11-07

该研究在905例经活检证实的个体中评估28种无创检测对MASH伴纤维化F2-F3及肝硬化的识别性能,并开发验证机器学习模型。既往NITs对F2-F3检测最佳AUC仅0.67,而新型ML模型验证AUC达0.86,阳性预测值提升至65%,对肝硬化ML模型平均AUC达0.93,优于传统工具。

要点

  • 既往无创检测对MASH F2-F3识别性能有限,最佳FAST的AUC仅0.67。
  • 新型ML模型对F2-F3验证AUC达0.86,阳性预测值提升至65%。
  • 对肝硬化ML模型平均AUC达0.93,优于传统无创工具。

核心结论

  • 新型机器学习模型在检测MASH F2-F3和肝硬化方面显著优于现有无创检测,可弥补FDA定义治疗窗口的诊断缺口。

研究设计与样本

  • 分析905例经活检证实的个体(含健康对照),评估28种生物标志物、影像或算法类无创检测。
  • 开发并验证基于分类梯度提升的机器学习模型,选用常规变量和激素标志物。
  • 在留出验证队列及训练/特征选择与两个独立验证的二次设置中评估性能。

主要发现

  • 既往NITs对F2-F3检测最佳为FAST,AUC 0.67,阴性预测值92%,阳性预测值30%。
  • 新型ML模型对F2-F3验证AUC最高达0.86,阳性预测值提升至65%,总体准确率最高87%。
  • 对肝硬化,AGILE4+在已知NITs中灵敏度最高92%、阴性预测值99.7%。
  • ML模型对肝硬化平均AUC 0.93(准确率86%),优于传统工具。

关键数值与推荐

  • F2-F3检测:FAST的AUC 0.67,阴性预测值92%,阳性预测值30%。
  • F2-F3检测:ML模型验证AUC最高0.86,阳性预测值最高65%,总体准确率最高87%。
  • 肝硬化检测:AGILE4+灵敏度92%,阴性预测值99.7%。
  • 肝硬化检测:ML模型平均AUC 0.93,准确率86%。

临床怎么用

  • ML模型可能支持新兴疗法的患者选择及临床试验入组。

以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。

原文

BACKGROUND & AIMS: Noninvasive tests (NITs) have not yet been systematically evaluated or optimized to detect metabolic dysfunction-associated steatohepatitis (MASH) with fibrosis stages F2 to F3, the current treatment-eligible group defined by the United States Food and Drug Administration for emerging treatments, or, separately, for cirrhosis (F4). We aimed to assess the performance of existing and novel approaches in identifying the target F2 to F3 population and excluding cirrhosis. METHODS: We analyzed a cohort of 905 biopsy-proven individuals (including healthy controls) and evaluated 28 biomarker-, imaging-, or algorithm-based NITs that have been previously tested for MASH F2 to F4, using published and newly optimized cutoffs. In parallel, we developed and validated machine learning (ML) models using categorical gradient boosting that selected routine variables and hormonal markers. Performance was assessed in a held-out validation cohort and a secondary setting using one population for training/feature selection and two for independent validation. RESULTS: Previous NITs did not perform optimally for F2 to F3 detection, the highest being fibroscan-AST (FAST) with an area under the curve (AUC) of 0.67; negative predictive value of 92%; and positive predictive value of 30%. In contrast, novel ML models achieved AUCs of up to 0.86 for F2 to F3 in validation, with improved positive predictive value up to 65% and overall accuracy up to 87%. For cirrhosis (F4), AGILE4+ showed the highest sensitivity (92%) and negative predictive value (99.7%) among known NITs, whereas ML models reached a mean AUC of 0.93 (accuracy = 86%), and outperformed conventional tools. CONCLUSIONS: Novel ML models significantly improve the detection of MASH F2 to F3 and cirrhosis over existing NITs and address a key diagnostic gap in the United States Food and Drug Administration-defined treatment window. These tools may support patient selection for emerging therapies, as well as for clinical trial enrolment.

来源:Clin Gastroenterol Hepatol,查看原文。版权归原作者所有。

引用

  1. Stefanakis K, Mingrone G, George J, et al. Novel Machine-Learning Models Outperform Noninvasive Tests in Metabolic Dysfunction-Associated Steatohepatitis With Fibrosis. Clin Gastroenterol Hepatol. 2025. PMID: 41207519. doi:10.1016/j.cgh.2025.10.029

供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。