Stratified neural networks in a time-to-event setting

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in oncology

Background:

• Deep neural networks (DNNs) are used for survival prediction with omics data, often using Cox proportional hazards partial likelihood as a loss function.
• Combining clinical datasets can improve DNN parameter learning but may violate Cox model assumptions if baseline hazards differ.
• Existing machine learning methods often use ranking loss, which may not adequately handle differing baseline hazards.

Purpose of the Study:

• To develop and evaluate a stratified partial likelihood loss function for DNNs to accommodate varying baseline hazards across pooled clinical datasets.
• To compare the performance of stratified loss functions against non-stratified partial likelihood and ranking loss.
• To investigate the impact of different loss functions on the identification of genes crucial for survival prediction.

Main Methods:

• Implemented a deep learning framework using a stratified partial likelihood loss function.
• Utilized high-dimensional transcriptome profiles (RNA-seq) from the Cancer Genome Atlas (TCGA) for analysis.
• Compared stratified partial likelihood with standard partial likelihood and ranking loss functions.
• Assessed discriminatory power and prediction error, and analyzed gene importance for survival prediction.

Main Results:

• Stratified loss functions significantly improved discriminatory power and reduced prediction error compared to non-stratified counterparts.
• Both stratified and non-stratified methods identified similar genes, but stratified loss functions assigned higher importance to known prognostic genes.
• The study demonstrated the benefit of pooling data with stratified loss functions for robust DNN parameter learning.

Conclusions:

• Stratified partial likelihood loss functions are crucial for accurate survival prediction using deep learning when pooling data from sources with different baseline hazards.
• This approach enhances the reliability of omics-driven survival models and improves the identification of clinically relevant biomarkers.
• The findings support the use of stratified loss functions for maximizing the benefits of multi-dataset integration in precision medicine research.