AUTHOR=Deng Zhikang , Dong Wentao , Xiong Situ , Jin Di , Zhou Hongzhang , Zhang Ling , Xie LiHan , Deng Yaohong , Xu Rong , Fan Bing TITLE=Machine learning models combining computed tomography semantic features and selected clinical variables for accurate prediction of the pathological grade of bladder cancer JOURNAL=Frontiers in Oncology VOLUME=13 YEAR=2023 URL=https://www.frontiersin.org/journals/oncology/articles/10.3389/fonc.2023.1166245 DOI=10.3389/fonc.2023.1166245 ISSN=2234-943X ABSTRACT=Objective

The purpose of this research was to develop a radiomics model that combines several clinical features for preoperative prediction of the pathological grade of bladder cancer (BCa) using non-enhanced computed tomography (NE-CT) scanning images.

Materials and methods

The computed tomography (CT), clinical, and pathological data of 105 BCa patients attending our hospital between January 2017 and August 2022 were retrospectively evaluated. The study cohort comprised 44 low-grade BCa and 61 high-grade BCa patients. The subjects were randomly divided into training (n = 73) and validation (n = 32) cohorts at a ratio of 7:3. Radiomic features were extracted from NE-CT images. A total of 15 representative features were screened using the least absolute shrinkage and selection operator (LASSO) algorithm. Based on these characteristics, six models for predicting BCa pathological grade, including support vector machine (SVM), k-nearest neighbor (KNN), gradient boosting decision tree (GBDT), logical regression (LR), random forest (RF), and extreme gradient boosting (XGBOOST) were constructed. The model combining radiomics score and clinical factors was further constructed. The predictive performance of the models was evaluated based on the area under the receiver operating characteristic (ROC) curve, DeLong test, and decision curve analysis (DCA).

Results

The selected clinical factors for the model included age and tumor size. LASSO regression analysis identified 15 features most linked to BCa grade, which were included in the machine learning model. The SVM analysis revealed that the highest AUC of the model was 0.842. A nomogram combining the radiomics signature and selected clinical variables showed accurate prediction of the pathological grade of BCa preoperatively. The AUC of the training cohort was 0.919, whereas that of the validation cohort was 0.854. The clinical value of the combined radiomics nomogram was validated using calibration curve and DCA.

Conclusion

Machine learning models combining CT semantic features and the selected clinical variables can accurately predict the pathological grade of BCa, offering a non-invasive and accurate approach for predicting the pathological grade of BCa preoperatively.