基于机器学习的创伤性脑损伤后早期死亡预测预后模型:与显著颅脑损伤后皮质类固醇随机化(CRASH)模型的比较
本研究回顾性纳入2017年2月至2021年8月的创伤性脑损伤患者,随机分为训练集和测试集,开发随机森林、支持向量机(SVM)和逻辑回归三种机器学习模型预测早期死亡,并与基于10008例患者开发的CRASH模型比较。测试集106例中,SVM的曲线下面积最高(0.952,95%CI 0.906-0.990),CRASH模型最低(0.942,95%CI 0.886-0.999),各机器学习模型与CRASH模型的曲线下面积无显著差异。作者认为机器学习模型尽管样本量较小,但性能可能与CRASH模型相当。
要点
- 共纳入423例创伤性脑损伤患者,训练集317例,测试集106例。
- SVM的曲线下面积最高(0.952,95%CI 0.906-0.990),CRASH模型最低(0.942,95%CI 0.886-0.999)。
- 机器学习模型与CRASH模型的曲线下面积无显著差异(随机森林 vs CRASH P=0.899,SVM vs CRASH P=0.760,逻辑回归 vs CRASH P=0.806)。
核心结论
- 机器学习模型尽管样本量较小,但性能可能与CRASH模型相当。
研究设计与样本
- 回顾性研究,纳入2017年2月至2021年8月的创伤性脑损伤患者,随机分为训练集(317例)和测试集(106例)。
- 预测变量包括临床表现、实验室值和计算机断层扫描(CT)表现。
- 采用训练集开发随机森林、支持向量机(SVM)和逻辑回归三种机器学习模型。
- CRASH模型是基于CRASH试验中10008例患者开发的预后模型。
- 将机器学习模型和CRASH模型应用于测试集评估性能。
主要发现
- 共纳入423例患者,训练集317例,测试集106例。
- 测试集中曲线下面积最高为SVM(0.952,95%CI 0.906-0.990),最低为CRASH模型(0.942,95%CI 0.886-0.999)。
- 机器学习模型与CRASH模型的曲线下面积无显著差异(随机森林 vs CRASH P=0.899,SVM vs CRASH P=0.760,逻辑回归 vs CRASH P=0.806)。
局限性
- 原文未报告局限性。
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)。
原文
BACKGROUND: Machine learning (ML) has been used to predict the outcomes of traumatic brain injury. However, few studies have reported the use of ML models to predict early death. This study aimed to develop ML models for early death prediction and to compare performance with the corticosteroid randomization after significant head injury (CRASH) model. METHODS: We retrospectively reviewed traumatic brain injury patients between February 2017 and August 2021. The patients were randomly assigned to a training set and a test set. Predictive variables included clinical findings, laboratory values, and computed tomography findings. The ML models (random forest, support vector machine [SVM], logistic regression) were developed with the training set. The CRASH model is a prognostic model that was developed based on 10,008 patients included in the CRASH trial. The ML and CRASH models were applied to the test set to evaluate the performance. RESULTS: A total of 423 patients were included; 317 and 106 patients were randomly assigned to the training and test sets, respectively. The area under the curve was highest in the SVM (0.952, 95% confidence interval = 0.906-0.990) and lowest in the CRASH model (0.942, 95% confidence interval = 0.886-0.999). There were no significant differences between the area under the curves of the ML and CRASH models (P = 0.899 for random forest vs. the CRASH model, P = 0.760 for SVM vs. the CRASH model, P = 0.806 for logistic regression vs. the CRASH model). CONCLUSIONS: The ML models may have comparable performances compared to the CRASH model despite being developed with a smaller sample size.
引用
- Lee SH, Lee CH, Hwang SH, et al. A Machine Learning-Based Prognostic Model for the Prediction of Early Death After Traumatic Brain Injury: Comparison with the Corticosteroid Randomization After Significant Head Injury (CRASH) Model. World Neurosurg. 2022. PMID: 35787963. doi:10.1016/j.wneu.2022.06.130
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。