Sunmin Kim

Comparative Performance of Machine Learning Models for Genomic-Based Survival Prediction in Breast, Lung, and Liver Cancers: A Network Meta-Analysis

Sunmin Kim, Undergraduate Student, Johns Hopkins University

Purpose: Accurate survival prediction is fundamental in oncology, supporting therapeutic stratification, clinical trial design, and personalized treatment. With the growing integration of high-dimensional genomic data into clinical care, machine learning (ML) methods provide opportunities to capture complex biological interactions beyond conventional statistical approaches. Despite their growing use, rigorous comparative evidence across algorithms remains limited. This study addresses this gap by systematically evaluating the performance of Cox proportional hazards (Cox-PH), Random Survival Forest (RSF), XGBoost (XGB), and Support Vector Machine (SVM) models for genomic-based survival prediction in breast, lung, and liver cancers. Methods: A systematic search of PubMed, Google Scholar, Embase, and Web of Science was conducted through June 2025 following PRISMA guidelines. Eligible studies applied at least two of the four algorithms to genomic datasets, with or without clinical covariates, and reported Harrell's C-index with standard errors. A frequentist network meta-analysis (NMA) was used to synthesize both direct and indirect evidence. Between-study heterogeneity was evaluated using Cochran's Q and I² statistics, and network consistency was assessed with node-splitting analyses. The primary outcome was standardized mean difference (SMD) in C-index relative to Cox-PH. Results: Six studies met eligibility criteria. In breast cancer, RSF (SMD = 8.48) and XGB (26.09) outperformed Cox-PH, while SVM (−11.35) underperformed. In lung cancer, RSF demonstrated the greatest advantage (122.23), with SVM (44.74) and XGB (35.75) also exceeding Cox-PH. In liver cancer, RSF (2.95) and XGB (1.47) showed modest improvements, while SVM (−17.21) was inferior. Conclusions: RSF and XGB consistently outperformed Cox-PH across cancer types, while SVM showed inferior performance. The magnitude of improvement varied by cancer type, being most pronounced in lung cancer and least in liver cancer. These findings provide evidence-based guidance for selecting ML algorithms in oncologic prognostic modeling and demonstrate the engineering potential of ensemble and boosting methods to enhance genomic-based survival prediction.