机器学习识别儿童急性诊疗环境中分流管故障及ShuntGPT的开发
本研究回顾性分析2017至2022年1167例因疑似分流管故障就诊的儿童神经外科会诊,收集24个变量,比较多种机器学习模型与ShuntGPT(SGPT)预测分流管翻修需求的能力。285例(24.4%)接受了翻修;SGPT表现最佳(准确率87%,AUC 0.927),且在去除影像数据后仍保持较高判别力(AUC 0.84)。
要点
- 在1167例儿童急诊会诊中,285例(24.4%)需分流管翻修。
- SGPT预测翻修的准确率87%、AUC 0.927,优于人工神经网络等模型。
- 去除影像数据后SGPT仍保持AUC 0.84,提示其可依赖文本信息。
核心结论
- 先进机器学习模型总体优于传统统计分析,但大量变量集存在过拟合担忧。
- SGPT因能解读细微文本而表现更优,有望辅助儿童急性诊疗中的分流故障决策。
研究设计与样本
- 回顾性分析2017至2022年1167例因疑似分流管故障的儿童神经外科会诊。
- 收集24个历史、临床和影像学变量,使用支持向量机、K近邻、随机森林、人工神经网络和专有大语言模型ShuntGPT进行分类。
主要发现
- 1167例会诊中285例(24.4%)接受了翻修。
- 多变量分析识别出脑室增大、分流管系列异常、嗜睡、精神状态改变和心动过缓为翻修预测因素。
- 人工神经网络准确率84%、AUC 0.88、敏感度71%、特异度88%。
- SGPT准确率87%、精确率0.80、召回率0.64、AUC 0.927。
- 传统分类器AUC为0.81至0.86,敏感度与精确率权衡各异。
关键数值与推荐
- 去除影像结果后所有模型性能下降,但SGPT仍保持高判别力(AUC 0.84)。
- SGPT在独立验证队列中准确率83%。
临床怎么用
- SGPT可作为儿童急诊疑似分流管故障时辅助判断翻修需求的决策工具。
- 在缺乏影像结果时SGPT仍具判别力,可支持临床文本信息的快速评估。
局限性
- 大量变量集可能带来过拟合担忧。
- 本研究为回顾性分析,需进一步外部验证。
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 AND OBJECTIVES: Identifying predictors of shunt failure in acute care settings is challenging. Previous studies largely rely on traditional statistical analyses to evaluate risk factors associated with shunt failure. However, these methods have not been adapted for emergency department (ED) presentations, where machine learning (ML) techniques may offer enhanced predictive power. Thus, we explored the application of ML and artificial intelligence in predicting necessity for shunt revision in pediatric ED visits. METHODS: A retrospective analysis was conducted on 1167 pediatric neurosurgical consults for suspected shunt malfunction from 2017 to 2022. We collected 24 historical, clinical, and radiographic variables. Various ML techniques were used, including support vector machine, K-nearest neighbors, random forest, an artificial neural network (ANN), and a proprietary large language model named ShuntGPT (SGPT), which classifies the need for shunt revision. RESULTS: Among 1167 ED consults for shunt malfunction, 285 patients (24.4%) underwent revision. Multivariate analysis identified increased ventricle size, abnormal shunt series, lethargy, altered mental status, and bradycardia as predictors of need for revision. The best-performing ML models included ANN (accuracy 84%, area under the curve [AUC] 0.88, 71% sensitivity, 88% specificity) and SGPT (accuracy 87%, precision 0.80, recall 0.64, AUC 0.927). Traditional classifiers (support vector machine, K-nearest neighbors, random forest) achieved AUCs of 0.81 to 0.86 with varying trade-offs in sensitivity and precision. With imaging results removed from training data, the performance of all models suffered, however, SGPT retained a high level of discrimination (AUC 0.84). SGPT continued to show high accuracy (83%) on a separate validation cohort. CONCLUSION: Advanced ML models generally outperformed traditional statistical analyses, albeit with concerns about overfitting due to the extensive variable set. SGPT, in particular, showed superior performance, likely due to its capacity to interpret nuanced text. This model represents a promising tool to enhance decision-making in pediatric acute care settings regarding shunt malfunctions.
引用
- Lehner KR, Kalluri AL, Jiang K, et al. Machine Learning Methods to Identify Pediatric Shunt Malfunction in the Acute Care Setting and the Development of ShuntGPT. Neurosurgery. 2026. PMID: 41697038. doi:10.1227/neu.0000000000003962
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。