Evaluation of Four Deep Learning-Based Postoperative Survival Prediction Models for Hepatocellular Carcinoma Based on SEER

Abstract

Abstract Accurate prognosis prediction is crucial for treatment decisions in HCC patients, but there is limited research investigating the combination of deep learning with time-to-event analysis. This study assessed four models, including deep learning survival neural network (DeepSurv), neural multi-task logistic regression model (N-MTLR), random survival forest (RSF), and traditional Cox proportional hazards (Cox-PH) models in predicting postoperative survival in hepatocellular carcinoma (HCC) patients. Utilizing data from the US SEER database 2004–2015 to, extract and analyze 5420 patients’ baseline demographic and tumor characteristics. The fellow was randomly divided into a training set and an internal testing set in a ratio of 8:2. Four algorithms were employed to build the predictive models after variable selection and was internally validated using metrics such as Harrell's concordance index (C-index), Brier Score, Receiver Operating Characteristic curve (ROC) curve, and calibration curve. DeepSurv, N-MTLR, RSF exhibited greater robustness compared to traditional Cox-PH models. These models aid in identifying patients who can benefit from HCC surgery and facilitates early intervention and the reliability of individual treatment recommendations.