GNN-surv: Discrete-Time Survival Prediction Using Graph Neural Networks

Area of Science:

• Oncology
• Bioinformatics
• Machine Learning

Background:

• Accurate survival prediction is crucial for patient prognosis and personalized cancer treatment.
• Traditional models can be enhanced by integrating patient similarity networks to capture complex data patterns.
• Graph Neural Networks (GNNs) offer a powerful approach for leveraging network structures in data.

Purpose of the Study:

• To develop and evaluate Graph Neural Network-based survival prediction models (GNN-surv) for enhanced accuracy.
• To leverage patient similarity networks constructed from genomic and clinical data for improved survival analysis.
• To assess the performance of GNN-surv models against traditional models in urologic cancer datasets.

Main Methods:

• Construction of patient similarity networks using genomic and clinical data from cancer patients.
• Training and evaluation of various GNN models integrated with Logistic Hazard and Probability Mass Function (PMF) survival models.
• Comparison of GNN-surv model performance against Multilayer Perceptron (MLP) models using time-dependent concordance index and integrated Brier score.

Main Results:

• GNN-surv models significantly outperformed traditional MLP models in survival prediction on BLCA and KIRC datasets.
• Performance improvements included up to 14.6% and 7.9% increases in the time-dependent concordance index.
• Reductions in the integrated Brier score were observed, reaching 26.7% and 24.1% for BLCA and KIRC, respectively.
• Models demonstrated robustness across varying graph construction hyperparameters and effectiveness across different GNN architectures.

Conclusions:

• GNN-surv models provide a superior approach for discrete-time survival prediction by effectively utilizing patient similarity networks.
• The enhanced accuracy and robustness of GNN-surv models offer a valuable tool for clinicians in oncology and personalized medicine.
• The adaptability of these models suggests broad applicability across different cancer types and potential for integration with other survival models or data modalities.