Short-term prognosis for hepatocellular carcinoma patients with lung metastasis: A retrospective cohort study based on the SEER database

Abstract

Our study aimed to develop a prediction model to predict the short-term mortality of hepatocellular carcinoma (HCC) patients with lung metastasis. The retrospective data of HCC patients with lung metastasis was from the Surveillance, Epidemiology, and End Results registration database between 2010 and 2015. 1905 patients were randomly divided into training set (n = 1333) and validation set (n = 572). There were 1092 patients extracted from the Surveillance, Epidemiology, and End Results database 2015 to 2019 as the validation set. The variable importance was calculated to screen predictors. The constructed prediction models of logistic regression, random forest, broad learning system, deep neural network, support vector machine, and naïve Bayes were compared through the predictive performance. The mortality of HCC patients with lung metastasis was 51.65% within 1 month. The screened prognostic factors (age, N stage, T stage, tumor size, surgery, grade, radiation, and chemotherapy) and gender were used to construct prediction models. The area under curve (0.853 vs. 0.771) of random forest model was more optimized than that of logistic regression model in the training set. But, there were no significant differences in testing and validation sets between random forest and logistic regression models. The value of area under curve in the logistic regression model was significantly higher than that of the broad learning system model (0.763 vs. 0.745), support vector machine model (0.763 vs. 0.689) in the validation set, and higher than that of the naïve Bayes model (0.775 vs. 0.744) in the testing model. We further chose the logistic regression prediction model and built the prognostic nomogram. We have developed a prediction model for predicting short-term mortality with 9 easily acquired predictors of HCC patients with lung metastasis, which performed well in the internal and external validation. It could assist clinicians to adjust treatment strategies in time to improve the prognosis.

Figure 1.

Flowchart of the systematic selection process.

Figure 2.

Variable importance from the random forest model.

Figure 3.

Nomogram for predicting 1-month survival of HCC patients with lung metastasis. HCC = hepatocellular carcinoma.

Figure 4.

ROC curve of the training set, testing set, and validation set in logistic regression model (A); random forest model (B); BLS model (C); DNN model (D); SVM model (E); and naïve Bayes (F). BLS = broad learning system, DNN = deep neural network, ROC = the receiver operating characteristic, SVM = support vector machine.