Machine learning based survival prediction for liver cancer using cancer registry linked study from the cancer public library database

Abstract

Prognostic factors differ between young and older patients with liver cancer, but research on predictive models is lacking. We aimed to develop mortality prediction models for younger and older patients with liver cancer and identify key variables. We included data on 4,510 patients diagnosed with liver cancer (2014–2016) in the Cancer Public Library Database, with 5-year mortality as the dependent variable. Models were developed using Extreme Gradient Boosting, LightGBM, ExtraTree, Random Forest, and Gradient Boosting, and influential variables were investigated. Gradient Boosting showed the highest AUC in the older group (0.854), whereas Random Forest showed the highest AUC in the younger group (0.843). In the prediction models, high-density lipoprotein level and estimated glomerular filtration rate strongly influenced mortality among younger patients. Age, topography code C221, and gamma glutamyl transpeptidase level strongly influenced mortality among older patients. Among prediction models, the most influential variables were the SEER (Surveillance, Epidemiology, and End Results) stage and absence of chronic viral hepatitis. The models showed good predictive performance for young and older patients and identified differences in the most influential variables for prediction models between younger and older patients.