Objective:To compare three models for predicting the risk of tumor treatment-related cardiac dysfunction(CTRCD)in aging colorectal cancer patients using deep learning methods.Methods:This was a retrospective cohort study.From January 2010 to December 2018,there were 382 aging patients with colorectal cancer who were hospitalized and treated at the Second People's Hospital of Lianyungang City.All those patients were randomly divided into the training group(255 cases)and the validation group(127 cases)based on a 2∶1 ratio.The baseline and prognostic data between two groups were compared.All patients were followed up for at least 2 years.The primary endpoint was CTRCD.The deep learning method was used to establish and verify three models to predict the risk of CTRCD,including logical regression model,random forest model(RF),and extreme gradient lifting model(XG Boost).We used the AUC and ROC to compare the predictive performance of three models for CTRCD,and compare the accuracy and effectiveness in predicting CTRCD using calibration curves and clinical decision curves.Results:Among the 382 patients,the incidence of CTRCD was 15.2％(58 cases),with 37 cases(14.5％)in the training group and 21 cases(16.5％)in the validation group,respectively.Compared with the training group,the smoking proportion of patients in the validation group was higher(P＜0.05),and other baseline data were similar between the two groups.In the training group,the AUC of the logistic regression model,RF,and XG Boost models for predicting CTRCD were 0.75,0.79,and 0.84,respectively.However,the specificity and accuracy of the logistic regression model were very low(0.56)and not high(0.66),and the prediction accuracy of the RF model was also lower than that of the XGBBoost model(0.74 vs.0.81).In the validation group,the AUC predicted by the three models for CTRCD were 0.72,0.77,and 0.82,respectively.Meanwhile,the calibration curves in the training group and validation group showed good correlation between the actual observation results of the XGBoost model predicting CTRCD and the predicted results of the column chart.The clinical decision curves of both the training and validation sets show that the XGBoost model has a higher net benefit value and better effectiveness.Conclusions:The XGBoost model has the best accuracy and stability in predicting the occurrence of CRTCD in colorectal cancer patients.Important related factors include global longitudinal strain,left ventricular ejection fraction,cTnI,natriuretic peptide,and age.
关键词
直肠癌/肿瘤相关心力衰竭/深度学习/预测价值
Key words
Colorectal cancer/Cancer therapy-related cardiac dysfunction/Deep learning/Prognostic value
维普
万方数据