Development and validation of an XGBoost model to predict 5-year survival in elderly patients with intrahepatic cholangiocarcinoma after surgery: a SEER-based study

Abstract

Background: Nomograms have been established to predict survival in postoperative or elderly intrahepatic cholangiocarcinoma (ICC) patients. There are no models to predict postoperative survival in elderly ICC patients. Extreme gradient boosting (XGBoost) can adjust the errors generated by existing models. This retrospective cohort study aimed to develop and validate an XGBoost model to predict postoperative 5-year survival in elderly ICC patients. Methods: The Surveillance, Epidemiology, and End Results (SEER) program provided data on elderly ICC patients aged 60 years or older and undergoing surgery. The median follow-up time was 20 months. Totally 1,055 patients were classified as training (n=738) and testing (n=317) sets at a ratio of 7:3. The outcome was postoperative 5-year survival. Demographic, tumor-related and treatment-related variables were collected. Variables were screened using the XGBoost model. The predictive performance of the model was assessed by the area under the receiver operating characteristic (ROC) curve (AUC), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and Kaplan-Meier curve. Cox regression analysis was conducted to estimate the risk of death in the predicted populations. The predictive abilities of the XGBoost model and the American Joint Commission on Cancer (AJCC) system (7th edition) were compared.Results: The XGBoost model achieved an AUC of 0.811, a sensitivity of 0.573, a specificity of 0.890, and a PPV of 0.849 in the training set. In the testing set, the model had an AUC of 0.713, a sensitivity of 0.478, a specificity of 0.814, and a PPV of 0.726. The 5-year mortality risk of patients predicted to die was 2.91 times that of patients predicted to survive [hazard ratio (HR) =2.91, 95% confidence interval (CI): 2.42-3.50]. The XGBoost model showed a better predictive performance than the AJCC staging system both in the training and testing sets. AJCC stage, multiple (satellite) tumors/nodules, tumor-node-metastasis (TNM) stage, more than one lobe invaded, direct invasion of adjacent organs, tumor size, and radiotherapy were relatively important features in survival prediction. Conclusions:The XGBoost model exhibited some predictive capacity, which may be applied to predict postoperative 5-year survival for elderly ICC patients.

Development of a machine learning-based model to predict prognosis of alpha-fetoprotein-positive hepatocellular carcinoma

“…In addition, Zhong et al [ 16 ] applied the XGBoost algorithm to create a prognostic model for patients with breast cancer with bone metastasis and showed AUC values of 0.88 and 0.80 in the training and test sets. Consistent with the previous studies [ 14 – 16 ], our present study also revealed that the XGBoost model showed good performance in prognostic survival prediction models, showing AUCs greater than 0.7 and even the 5-year AUC value over 0.8 (training data). Generally, an AUC ≥ 0.7 indicates that the model has an adequate predictive ability [ 22 ].…”

“…In recent years, machine learning-based AI models attracted increasing attention in clinical practice [ 14 , 20 , 21 ]. Especially, AI-based technologies have made a significant contribution to the field of cancer research [ 21 ].…”

“…Recent studies have examined the use of the XGBoost model in predicting the survival of cancer patients, and verified that this model is of better prediction ability in various types of cancer. In a recent study, Xu et al [ 14 ] reported that the XGBoost model exhibited a better performance than the AJCC staging system in predict postoperative survival in elderly intrahepatic cholangiocarcinoma patients, with the AUCs of more than 0.7 both in the training and test sets. Li et al [ 15 ] found that the XGBoost model behaved efficiently and successfully in predict the survival of patients with breast cancer brain metastases, with an AUC of 0.8 or above (test data).…”

Construction of an Early Alert System for Intradialytic Hypotension before Initiating Hemodialysis Based on Machine Learning

“…Four filling methods and three feature screening methods were used to obtain 12 datasets. Eighteen machine learning algorithms, including logistic regression [27, 28], Latent Dirichlet allocation [29], Quadratic Discriminant Analysis [30], Stochastic Gradient Descent [31], k-Nearest Neighbor [32], Decision Tree [33], Naive Bayes [34], Gaussian Naïve Bayes [35], Multinomial Naive Bayes [36], Bernoulli Naïve Bayes [37], Support Vector Machine [38], passive-aggressive [39], AdaBoost [40], bagging, Random Forest [41], Extremely Randomized Trees [42], gradient boosting [41], XGBoost [43], and ensemble learning [44], were used to train 216 models. The process of building the models was as follows:The dataset was randomly divided into a training and a test set in a ratio of 8:2.The training set data were entered into the machine learning model, and the 10-fold cross-validation method was used to continuously adjust the model parameters, so that the parameters had the largest area under the receiver operating characteristic curve (AUC) value on the training set.…”

Machine Learning Approach for Improved Longitudinal Prediction of Progression from Mild Cognitive Impairment to Alzheimer’s Disease

“…The predictive performance of each trained model (for each of the 7 prediction windows, as well as for the averaging approach) on its corresponding test set was evaluated using several standard performance measures [ 45 , 46 ]. These performance metrics were AUROC (the probability that a classifier will be able to distinguish between an instance of the positive class and one of the negative class), sensitivity (true positive rate), specificity (true negative rate), positive predictive value (PPV, the proportion of true positives among all positive predictions), negative predictive value (NPV, the proportion of true negatives among all negative predictions), and accuracy (the proportion of correct predictions among all predictions).…”