Table 1.

Performance comparison between best performing models

Model	C-index (CI 95%)	P-value vs. linear	P-value vs. NN	P-value vs. RF
Linear Cox model	0.618 (0.609–0.627)	—	—	—
Neural network	0.634 (0.626–0.642)	7.8e−2	—	—
Random forest	0.674 (0.666–0.681)	1.5e−4	3.2e−3	—
Gradient boosting	0.676 (0.668–0.684)	2.0e−5	1.1e−3	6.1e−1

Based on their performance during the nested cross-validation on the training set, we selected the best performing models and evaluated them on the test set. Comparing the best models of each category to one another, we found that tree-based models significatively improve the C-index. In terms of Net Reclassification Improvement (NRI), the best model (gradient boosting) improved patients’ reclassification by 31.6% (19.0%, 40.7%) compared to the linear Cox model.