文献速递神经肿瘤功能神经外科

基于经颅磁刺激的机器学习预测运动区胶质瘤肿瘤分级

Transcranial Magnetic Stimulation-Based Machine Learning Prediction of Tumor Grading in Motor-Eloquent Gliomas.
PubMed 文献 · Neurosurgery · 2024-03-21

该回顾性队列研究纳入2018至2022年接受手术的运动区胶质瘤患者177例及健康受试者10例,采集导航经颅磁刺激(nTMS)衍生变量并用机器学习模型预测肿瘤分级。皮层兴奋性评分与WHO分级显著相关(P=.04),模型预测WHO 4级准确率0.57,区分高级别与低级别胶质瘤准确率0.83,加入IDH突变和1p19q共缺失数据后准确率分别升至0.95和0.74。

要点

  • 上肢异常半球间静息运动阈值比在WHO 2、3、4级中分别占22.7%、62.5%、75.4%。
  • 皮层兴奋性评分与WHO分级显著相关(P=.04)。
  • 模型区分高级别与低级别胶质瘤准确率0.83,加入分子数据后达0.95和0.74。

核心结论

  • 基于nTMS衍生半球间兴奋性评估的机器学习算法能准确预测累及运动通路的高级别胶质瘤,拟合分子数据后准确率进一步提高,为术前影像基因组学联合方法铺路。

研究设计与样本

  • 回顾性队列研究,纳入2018至2022年接受手术的运动区胶质瘤患者177例,另纳入健康受试者10例。
  • 采集术前nTMS衍生变量:静息运动阈值(RMT)、半球间RMT比值(iRMTr,高于10%为异常)及皮层兴奋性评分(异常iRMTr数量)。
  • 收集WHO分级和分子分型以表征每个肿瘤,经统计特征选择后拟合机器学习模型预测肿瘤分级。

主要发现

  • 共纳入177例患者:WHO 2级32例、WHO 3级65例、WHO 4级80例。
  • 上肢异常iRMTr在WHO 2级占22.7%、WHO 3级占62.5%、WHO 4级占75.4%。
  • 下肢iRMTr异常在WHO 2级占23.1%、WHO 3级占67.6%、WHO 4级占63.6%。
  • 皮层兴奋性评分与WHO分级有统计学显著相关(P=.04)。
  • 以这些变量为预测因子,机器学习模型预测WHO 4级病变的准确率为0.57;在高级别胶质瘤 vs 低级别胶质瘤的亚组分析中,预测高级别胶质瘤的准确率增至0.83。

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: Navigated transcranial magnetic stimulation (nTMS) is a well-established preoperative mapping tool for motor-eloquent glioma surgery. Machine learning (ML) and nTMS may improve clinical outcome prediction and histological correlation. METHODS: This was a retrospective cohort study of patients who underwent surgery for motor-eloquent gliomas between 2018 and 2022. Ten healthy subjects were included. Preoperative nTMS-derived variables were collected: resting motor threshold (RMT), interhemispheric RMT ratio (iRMTr)-abnormal if above 10%-and cortical excitability score-number of abnormal iRMTrs. World Health Organization (WHO) grade and molecular profile were collected to characterize each tumor. ML models were fitted to the data after statistical feature selection to predict tumor grade. RESULTS: A total of 177 patients were recruited: WHO grade 2-32 patients, WHO grade 3-65 patients, and WHO grade 4-80 patients. For the upper limb, abnormal iRMTr were identified in 22.7% of WHO grade 2, 62.5% of WHO grade 3, and 75.4% of WHO grade 4 patients. For the lower limb, iRMTr was abnormal in 23.1% of WHO grade 2, 67.6% of WHO grade 3%, and 63.6% of WHO grade 4 patients. Cortical excitability score ( P = .04) was statistically significantly related with WHO grading. Using these variables as predictors, the ML model had an accuracy of 0.57 to predict WHO grade 4 lesions. In subgroup analysis of high-grade gliomas vs low-grade gliomas, the accuracy for high-grade gliomas prediction increased to 0.83. The inclusion of molecular data into the model-IDH mutation and 1p19q codeletion status-increases the accuracy of the model in predicting tumor grading (0.95 and 0.74, respectively). CONCLUSION: ML algorithms based on nTMS-derived interhemispheric excitability assessment provide accurate predictions of HGGs affecting the motor pathway. Their accuracy is further increased when molecular data are fitted onto the model paving the way for a joint preoperative approach with radiogenomics.

来源:Neurosurgery,查看原文。版权归原作者所有。

引用

  1. Lavrador JP, Mirallave-Pescador A, Soumpasis C, et al. Transcranial Magnetic Stimulation-Based Machine Learning Prediction of Tumor Grading in Motor-Eloquent Gliomas. Neurosurgery. 2024. PMID: 38511960. doi:10.1227/neu.0000000000002902

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。