Scholarlink

Background <p>Glioma is characterized by a poor prognosis and limited possibilities for treatment. Previous studies have developed prediction models for glioma using genetic, clinical, pathological, imaging and other aspects; however, few studies have combined these data. The current study is intended to fully utilize medical data from the routine practice of glioma care and develop a model with the assistance of machine learning.</p> Methods <p>Multiple factors—including demographic features, radiomic features, laboratory biomarkers, and pathological features—were collected from two Class Three hospitals in China. Preoperative images and blood tests were quantified with machine learning methods. The survival time was documented during follow-up. Multivariate Cox regression and seven machine learning algorithms were used for modeling.</p> Results <p>A total of 674 glioma patients from two centers were enrolled. Fifteen radiomic features (RFs) and ten laboratory biomarkers were used to create the RF score and blood score. A clinicopathological–radiomic–blood model (CRBM) was created to stratify the mortality risk of glioma patients (<i>P</i> &lt; 0.0001). The AUC of the Cox-based model was 0.913 (0.886-0.940) on the training dataset and 0.802 (0.738-0.865) on the validation dataset, and the AUCs of the XGBoost model on the same datasets were 0.954 (0.935-0.973) and 0.761 (0.693-0.829), respectively. The SHapley Additive exPlanations (SHAP) method suggested the contribution of preoperative imaging and laboratory data to the model.</p> Conclusion <p>The CRBM is able to predict the survival of glioma patients with acceptable accuracy. Our work suggests the considerable potential of combined clinically derived data in predicting glioma survival and the utility of machine learning in variable selection and model construction.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Integrated clinicopathological–radiomic–blood model for glioma survival prediction via machine learning: a multicenter cohort study

  • 期刊论文

  • 发表日期：

    2025年7月12日

查看全文
  • Zhihao Wang,
  • Tao Chang,
  • Jing Yang,
  • Chaodong Xiang,
  • Xianqi Wang,
  • Pinzhen Chen,
  • Yunhui Zeng,
  • Lanqin Deng,
  • Wenhao Li,
  • Yuhang Ou,
  • Siliang Chen,
  • Hao Ren,
  • Yuan Yang,
  • Xiaofei Hu,
  • Qing Mao,
  • Wei Chen,
  • Yanhui Liu

摘要

Background

Glioma is characterized by a poor prognosis and limited possibilities for treatment. Previous studies have developed prediction models for glioma using genetic, clinical, pathological, imaging and other aspects; however, few studies have combined these data. The current study is intended to fully utilize medical data from the routine practice of glioma care and develop a model with the assistance of machine learning.

Methods

Multiple factors—including demographic features, radiomic features, laboratory biomarkers, and pathological features—were collected from two Class Three hospitals in China. Preoperative images and blood tests were quantified with machine learning methods. The survival time was documented during follow-up. Multivariate Cox regression and seven machine learning algorithms were used for modeling.

Results

A total of 674 glioma patients from two centers were enrolled. Fifteen radiomic features (RFs) and ten laboratory biomarkers were used to create the RF score and blood score. A clinicopathological–radiomic–blood model (CRBM) was created to stratify the mortality risk of glioma patients (P < 0.0001). The AUC of the Cox-based model was 0.913 (0.886-0.940) on the training dataset and 0.802 (0.738-0.865) on the validation dataset, and the AUCs of the XGBoost model on the same datasets were 0.954 (0.935-0.973) and 0.761 (0.693-0.829), respectively. The SHapley Additive exPlanations (SHAP) method suggested the contribution of preoperative imaging and laboratory data to the model.

Conclusion

The CRBM is able to predict the survival of glioma patients with acceptable accuracy. Our work suggests the considerable potential of combined clinically derived data in predicting glioma survival and the utility of machine learning in variable selection and model construction.