OBJECTIVES: This review aimed to identify AI-based methods used to and evaluate their performance.

METHODS: A search strategy was developed to identify relevant publications on AI-based predictive modeling in oncology published since 2020. The search was conducted in Embase and Medline via the EMBASE platform, and the retrieved citations were screened based on pre-specified selection criteria. Data from the selected articles were extracted using a pre-specified data extraction grid that included study objectives, AI methods used, and performance assessments.

RESULTS: A total of 1,127 citations were retrieved, and 30 publications were selected for data extraction: most studies included applications of survival prediction (n=28) and the review also included a systematic literature review of machine learning models for survival prediction in breast cancer, and a survey of methods used in predictive modelling. Most applications related to breast cancer(n=7), lung cancer (n=5) and hepatocarcinoma (n=5). The most widely used methods included neural network survival models (n=10), random forest survival models (n=6) and support vector machine models (n=6). Additional methods included survival trees and neural multitask logistic regression, and gradient boosting methods were used to optimize the performance of the algorithm. Random forest survival models and deep learning systems demonstrated good performance in predicting survival outcomes according to AUC and C-index values. Compared to standard methods, AI methods frequently matched or exceeded performance, with random forest survival models models often showing superior predictive accuracy.

CONCLUSIONS: This review demonstrates the potential of AI-based methods to enhance survival predictions and therefore improve the accuracy of survival extrapolation, a crucial step in providing reliable cost-effectiveness assessments of new health technologies.