文献速递脊柱外科骨肿瘤

利用机器学习预测转移性脊柱肿瘤术后神经功能结局

Predicting Postoperative Neurological Outcomes in Metastatic Spinal Tumor Surgery Using Machine Learning.
PubMed 文献 · Spine (Phila Pa 1976) · 2025-03-03

本研究为多中心回顾性分析,纳入38家机构共244例接受手术的转移性脊柱肿瘤患者,用随机森林、XGBoost、LightGBM和CatBoost四种机器学习算法预测术后1个月Frankel分级D或E级(可行走)的神经功能结局。随机森林模型ROC曲线下面积最高,为0.8516;SHAP分析显示术前Frankel分级、转移能力、炎症标志物和手术时机是最重要的预测因素。

要点

  • 244例患者术后1个月可行走比例由术前36.8%升至63.1%。
  • 随机森林模型ROC曲线下面积最高,为0.8516。
  • 术前Frankel分级、转移能力、炎症标志物和手术时机为最重要预测因素。

核心结论

  • 机器学习模型在评估转移性脊柱肿瘤患者术后神经功能状态方面显示出较强的预测性能。
  • 术前神经功能、功能状态和炎症标志物等关键因素显著影响结局。
  • 这些发现可为手术决策提供参考,并帮助设定合理的术后预期。

研究设计与样本

  • 多中心回顾性数据分析,纳入38家机构共244例因转移性疾病接受脊柱手术的患者。
  • 主要结局为术后1个月的功能性行走,定义为Frankel分级D或E级。
  • 采用随机森林、XGBoost、LightGBM和CatBoost四种机器学习算法构建预测模型。
  • 特征选择采用Boruta算法和方差膨胀因子分析以减少多重共线性。

主要发现

  • 244例患者中,可行走患者(Frankel分级D或E级)比例由术前36.8%升至术后1个月63.1%。
  • 随机森林模型ROC曲线下面积最高,为0.8516,其次为XGBoost(0.8351)、CatBoost(0.8331)和LightGBM(0.8098)。
  • SHAP分析识别术前Frankel分级、转移能力、炎症标志物(C反应蛋白和白细胞-淋巴细胞)及手术时机为术后结局最重要的预测因素。

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

STUDY DESIGN: Retrospective analysis of data collected across multiple centers. OBJECTIVE: To develop machine learning models for predicting neurological outcomes 1 month postoperatively in patients with metastatic spinal tumors undergoing surgery, and to identify key factors influencing neurological recovery. BACKGROUND: The increasing prevalence of spinal metastases has led to a growing need for surgical intervention to address mechanical instability and neurological deficits. Predicting postoperative neurological status, as assessed by the Frankel classification, can provide valuable insights for surgical planning and patient counseling. Traditional prognostic models have shown limitations in capturing the complexity of neurological recovery patterns. PATIENTS AND METHODS: We analyzed data from 244 patients who underwent spinal surgery for metastatic disease across 38 institutions. The primary outcome was functional ambulation, defined as Frankel grades D or E at 1 month postoperatively. Four machine learning algorithms (random forest, XGBoost, LightGBM, and CatBoost) were used to build predictive models. Feature selection employed the Boruta algorithm and variance inflation factor analysis to reduce multicollinearity. RESULTS: Among the 244 patients, the proportion of ambulatory patients (Frankel grades D or E) increased from 36.8% preoperatively to 63.1% at 1 month postoperatively. The random forest model achieved the highest area under the receiver operating characteristic curve of 0.8516, followed by XGBoost (0.8351), CatBoost (0.8331), and LightGBM (0.8098). SHapley Additive exPlanations analysis identified preoperative Frankel classification, transfer ability, inflammatory markers (C-reactive protein and white blood cell-lymphocyte), and surgical timing as the most important predictors of postoperative outcomes. CONCLUSIONS: Machine learning models showed strong predictive performance in assessing postoperative neurological status of patients with metastatic spinal tumors. Key factors, including preoperative neurological function, functional ability, and inflammation markers, significantly influenced outcomes. These findings could inform surgical decision-making and help set realistic postoperative expectations while potentially improving patient care through more accurate outcome prediction.

来源:Spine (Phila Pa 1976),查看原文。版权归原作者所有。

引用

  1. Maki S, Shiratani Y, Orita S, et al. Predicting Postoperative Neurological Outcomes in Metastatic Spinal Tumor Surgery Using Machine Learning. Spine (Phila Pa 1976). 2025. PMID: 40085125. doi:10.1097/BRS.0000000000005322

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。