机器学习在正常压力脑积水管理中的应用:系统综述
本系统综述检索 PubMed、Embase 和 Web of Science 数据库,纳入 22 项研究共 53 个机器学习模型,评估其在正常压力脑积水管理中的应用。模型总体平均准确率为 77%,诊断模型与干预模型的准确率及曲线下面积存在统计学显著差异;PROBAST 评估显示 59.09% 的研究存在高偏倚风险。
要点
- 纳入 22 项研究共 53 个模型,卷积神经网络为最常用模型。
- 模型总体平均准确率 77%,卷积神经网络最高 98%,决策树最低 55%(P=0.176)。
- 诊断模型与干预模型准确率 83.4% vs 69.4%、AUC 0.882 vs 0.729(P<0.001)。
核心结论
- 尽管机器学习模型准确率较高,但当前模型面临诸多挑战,需要标准化机器学习模型以便跨研究比较并提升正常压力脑积水的决策与诊疗。
研究设计与样本
- 系统综述,检索 PubMed、Embase 和 Web of Science 数据库中报告机器学习在正常压力脑积水应用的研究。
- 采用 PROBAST 进行质量评估,并参照 TRIPOD 报告规范评估依从性,统计分析显著性水平为 <0.05。
- 共纳入 22 项研究、53 个模型。
主要发现
- 卷积神经网络是最常用的模型。
- 用于训练各模型的输入包括临床特征、计算机断层扫描、磁共振成像、颅内脉冲波形特征和灌注输注。
- 模型总体平均准确率为 77%,卷积神经网络最高为 98%,决策树最低为 55%(P=0.176)。
- 诊断模型与干预模型的准确率和曲线下面积差异有统计学意义(准确率:83.4% vs 69.4%,曲线下面积:0.882 vs 0.729;P<0.001)。
- PROBAST 评估中 59.09%(n=13)的研究存在高偏倚风险,81.82%(n=18)的研究具有高适用性;仅 55.15% 的研究遵循 TRIPOD 声明。
局限性
- PROBAST 评估显示 59.09%(n=13)的研究存在高偏倚风险。
- 仅 55.15% 的研究遵循 TRIPOD 声明。
- 当前机器学习模型面临诸多挑战,需要标准化以便跨研究比较。
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: In the past decade, many machine learning (ML) models have been used in the management of normal pressure hydrocephalus (NPH). This study aims at systematically reviewing those ML models. METHODS: The PubMed, Embase, and Web of Science databases were searched for studies reporting applications of ML in NPH. Quality assessment was performed using Prediction model Risk Of Bias ASsessment Tool (PROBAST) and Transparent Reporting of a multivariable predication model for Individual Prognosis Or Diagnosis (TRIPOD) adherence reporting guidelines, and statistical analysis was performed with the level of significance of <0.05. RESULTS: A total of 22 studies with 53 models were included in the review, of which the convolutional neural network was the most used model. Inputs used to train various models included clinical features, computed tomography scan, magnetic resonance imaging, intracranial pulse waveform characteristics, and perfusion infusion. The overall mean accuracy of the models was 77% (highest for the convolutional neural network, 98%, while lowest for decision tree, 55%; P = 0.176). There was a statistically significant difference in the accuracy and area under the curve of diagnostic and interventional models (accuracy: 83.4% vs. 69.4%, area under the curve: 0.882 vs. 0.729; P < 0.001). Overall, 59.09% (n = 13) and 81.82% (n = 18) of the studies had high-risk bias and high-applicability, respectively, on PROBAST assessment; however, only 55.15% of the studies adhered to the TRIPOD statement. CONCLUSIONS: Though highly accurate, there are many challenges to current ML models necessitating the need to standardize the ML models to enable comparison across the studies and enhance the NPH decision-making and care.
引用
- Pahwa B, Tayal A, Shukla A, et al. Utility of Machine Learning in the Management of Normal Pressure Hydrocephalus: A Systematic Review. World Neurosurg. 2023. PMID: 37356488. doi:10.1016/j.wneu.2023.06.080
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。