Three-class machine learning model based on 18F-FDG PET/CT for predicting EGFR mutation subtypes in lung adenocarcinoma
10.3760/cma.j.cn321828-20240815-00293
- VernacularTitle:基于PET/CT的三分类机器学习模型预测肺腺癌EGFR突变亚型
- Author:
Xinyu GE
1
;
Jianxiong GAO
;
Rong NIU
;
Yunmei SHI
;
Zhenxing JIANG
;
Yan SUN
;
Jinbao FENG
;
Yuetao WANG
;
Xiaonan SHAO
Author Information
1. 苏州大学附属第三医院、常州市第一人民医院核医学科，苏州大学核医学与分子影像临床转化研究所，常州市分子影像重点实验室，常州 213003
- Publication Type:Journal Article
- Keywords:
Lung neoplasms;
Adenocarcinoma;
Genes, erbB-1;
Mutation;
Machine learning;
Positron-emission tomography;
Tomography, X-ray computed;
Fluorodeoxyglucose F18
- From:
Chinese Journal of Nuclear Medicine and Molecular Imaging
2025;45(9):530-536
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To develop and assess a three-class machine learning model for predicting wild-type, 19 del, and 21 L858R mutations of the epidermal growth factor receptor (EGFR) in lung adenocarcinoma using 18F-FDG PET/CT radiomic features and clinical features. Methods:The retrospective data was collected from 703 patients (346 males, 357 females; age (64.3±9.0) years) with lung adenocarcinoma at the First People′s Hospital of Changzhou from January 2018 to June 2023. Patients were divided into the training set (563 cases) and test set (140 cases) at the ratio of 8∶2. Clinical features were selected using recursive feature elimination (RFE). Radiomic features were extracted from PET and CT images, and the optimal feature sets were selected using minimum redundancy maximum relevance (mRMR) and least absolute shrinkage and selection operator (LASSO) methods. Base models were constructed by using random forest (RF), logistic regression (LR), support vector machine (SVM), K-nearest neighbors (KNN), and multi-layer perceptron (MLP), and the stacking method was applied to establish the CT and PET ensemble models. Delong test was used to compare the AUC differences between the PET/CT combined model and the clinical + PET/CT integrated model.Results:Among 703 patients, 273 were with EGFR wild-type, 202 were with 19 del mutation, and 228 were with 21 L858R mutation. In the single-modal analysis, the AUCs of CT ensemble model in the training and test sets were 0.893 and 0.667, respectively, while the AUCs of PET ensemble model were 0.692 and 0.660. The AUC of PET/CT combined model were 0.897 in training set and 0.672 in test set. The AUC of clinical + PET/CT integrated model showed further improvement, with AUCs of 0.902 and 0.721 in training and test sets, respectively. Notably, the clinical + PET/CT integrated model outperformed PET/CT combined model in predicting wild-type EGFR (test set AUC: 0.784 vs 0.707; Z=3.28, P=0.001). Conclusion:The three-class model (clinical + PET/CT integrated model) based on 18F-FDG PET/CT radiomics and clinical features effectively predicts EGFR mutation subtypes in lung adenocarcinoma.
