人工智能模型预测经蝶垂体瘤术后不良结局的诊断试验准确性综述
该综述遵循PRISMA方案,纳入8篇观察性或随机设计研究,评估机器学习模型预测内镜经蝶垂体瘤术后尿崩症和脑脊液漏的效能。随机森林表现最佳,平均AUC为0.832,神经网络次之为0.8,各模型平均AUC均大于0.7。
要点
- 随机森林是预测术后尿崩症和脑脊液漏的最佳机器学习模型,平均AUC为0.832。
- 神经网络模型平均AUC为0.8,表现次之。
- 所研究模型均报告了灵敏度、特异度及平均AUC大于0.7。
核心结论
- 在已识别的研究中,识别尿崩症和脑脊液漏风险患者的最佳机器学习方法是随机森林。
研究设计与样本
- 遵循PRISMA方案的诊断试验准确性综述,最终纳入8篇文章。
- 纳入标准为观察性或随机设计、报告经蝶手术的垂体腺瘤患者、使用机器学习预测术后主要结局并报告至少一项相关结局。
主要发现
- 分析的2个结局为尿崩症和脑脊液漏。
- 研究的模型包括随机森林、神经网络、逻辑回归和决策树,均报告了灵敏度、特异度及平均AUC大于0.7。
- 随机森林平均AUC为0.832,神经网络平均AUC为0.8。
关键数值与推荐
- 随机森林平均AUC为0.832。
- 神经网络平均AUC为0.8。
- 各模型平均AUC均大于0.7。
临床怎么用
- 机器学习模型可作为经蝶垂体瘤手术的术前规划工具,用于预测术后尿崩症和脑脊液漏风险。
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
Pituitary tumors are the second most common type of brain tumor in adults, accounting for 17% of adult tumors. Over the years, endoscopic transsphenoidal surgery has become widely used in the treatment of pituitary tumors. With the recent advancement of Artificial Intelligence (AI) and Machine Learning (ML) technologies to generate clinical predictions that aid in diagnostic medicine, they have stood out as preoperative planning tools with the aim of predicting clinical outcomes, including postoperative ones. This article aims to evaluate the applicability of ML models in predicting postoperative complications in the postoperative period of endoscopic transsphenoidal surgeries for pituitary lesions. Following the PRISMA protocol. Articles were included if they had an observational or randomized design, reported on patients with skull base tumors (pituitary adenomas) who underwent transsphenoidal surgery, used machine learning-based prediction for major outcomes after surgery, and reported at least one relevant outcome. Eight articles were included in this analysis. The 2 outcomes analyzed were diabetes insipidus and cerebrospinal fluid (CSF) leak. The models studied were Random Forest, Neural Network, Logistic Regression, and Decision Tree, and they presented sensitivity, specificity, and an average area under the curve greater than 0.7. The best model was Random Forest with an average area under the curve of 0.832, followed by the Neural Network model, which had an average area under the curve of 0.8. In the studies we identified, the best machine learning methodology for identifying patients at risk of Diabetes Insipidus and cerebrospinal fluid leak was Random Forest.
引用
- Bacelar Ferreira V, Batista da Hora DA, Vargas de Mendonça KL, et al. Artificial Intelligence Models as Predictors of Poor Outcomes in Transsphenoidal Pituitary Tumor Surgeries: A Diagnostic Test Accuracy Review. World Neurosurg. 2025. PMID: 41241048. doi:10.1016/j.wneu.2025.124640
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。