可解释机器学习模型预测超声引导下经皮肾活检后出血：多中心研究
该多中心回顾性研究基于664例开发与内部验证、137例外部队列验证，构建可解释机器学习模型预测超声引导下经皮肾活检后出血。随机森林模型表现最佳，训练集、内部验证集和外部验证集AUC分别为0.831、0.784和0.776，准确度0.742、敏感度0.675、特异度0.805、F1分数0.691。SHAP分析显示凝血酶原时间延长、肾实质厚度变薄、穿刺点至肾脏下缘距离减小、eGFR降低和病程延长增加出血风险。
要点
- 随机森林模型外部验证AUC为0.776，可用于辅助预测肾活检后出血风险。
- 肾实质厚度、穿刺点至肾脏下缘距离及其与肾长径比值是重要超声预测因子。
- 凝血酶原时间延长、eGFR降低和病程延长提示出血风险增高。
核心结论
- 随机森林模型能有效预测超声引导下经皮肾活检后出血，支持早期诊断和决策。
研究设计与样本
- 回顾性多中心研究，开发与内部验证队列664例，外部验证队列137例。
- 收集关键超声参数和临床数据，采用Boruta算法筛选预测因子。
- 八种机器学习模型经5折交叉验证超参数调优，以AUC确定最佳模型。
- 使用SHAP值增强模型解释性。
诊断标准与关键数值
- 随机森林模型训练集AUC 0.831、内部验证集AUC 0.784、外部验证集AUC 0.776。
- 模型准确度0.742、敏感度0.675、特异度0.805、F1分数0.691。
- 重要预测因子包括肾实质厚度、穿刺点至肾脏下缘距离、D1与肾长径比值、病程、eGFR和凝血酶原时间。
临床应用与阅片要点
- 阅片时可关注肾实质厚度及穿刺点至肾脏下缘距离，辅助评估活检后出血风险。
- 报告时可结合凝血酶原时间、eGFR和病程等临床信息综合提示风险。
- 模型预测结果可用于围手术期管理决策支持。
局限性
- 原文未明确报告研究局限性。
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）。
英文原文（摘要）· 点击展开
OBJECTIVE: This study aims to develop and validate an interpretable machine learning (ML) model for predicting post-procedural hemorrhage (PH) after ultrasound-guided percutaneous renal biopsy (PRB), aiding perioperative management.
MATERIALS AND METHODS: A retrospective analysis was conducted using data from 664 patients to develop and internally validate the predictive model. Key ultrasound parameters and clinical data were collected. The Boruta algorithm selected predictive factors. Eight ML models underwent hyperparameter tuning with 5-fold cross-validation, using the area under the curve (AUC) to determine the best model, which was further validated externally (n = 137). Model interpretation was enhanced using SHAP values.
RESULTS: Important predictive factors identified included renal parenchymal thickness (RPT), distance from the puncture point to the lower edge of the kidney (D1), and the ratio of D1 to the Renal length (R1), disease course, eGFR and prothrombin time (PT). The random forest model outperformed others with AUCs of 0.831, 0.784, and 0.776 for the training, internal validation, and external validation sets, respectively, displaying an accuracy of 0.742, sensitivity of 0.675, specificity of 0.805, and F1 score of 0.691. SHAP analysis highlighted prolonged PT, thinner RPT, smaller D1, lower eGFR, and extended disease course as increasing PH risk.
CONCLUSION: The random forest model effectively predicts PH post-PRB, supporting early diagnosis and decision-making.
引用
- Su L, Nie J, Xie Q, et al. Interpretable machine learning model for predicting hemorrhage following ultrasound-guided percutaneous renal biopsy: a multicenter study. Abdom Radiol (NY). 2026. PMID: 42207295. doi:10.1007/s00261-026-05582-2
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。