Fig. 4

Performance comparison single- versus multi-center for two overall survival classes. Shown for non-robust feature sets, robust features with sequence prior, and hand-picked feature selection. The results show the trade-off between single-center performance and the drop when moving to multi-center data. Introducing priors helped reduce performance drop. The arrows indicate whether a prior increased performance on multi-center data when compared to the non-robust features. The benefit of robust features highly depends on the class boundary used, since different feature selections and machine learning methods were used. The supplementary material contains corresponding plots for further classification performance metrics, as well as the results for the experiments with three overall survival classes