Table 2.

Comparison of best-performing feature selectors (by AUC). Due to the high class imbalance for class boundaries at high and low overall survival values, the balanced accuracy is reported. The single-center metrics are listed as the mean across splits of stratified 10-fold cross-validation. A large performance drop can be observed when the model is tested on unseen multi-center data

Prior	Center	Robustness	Class Boundary	Selector	ML model	AUC	Bal. Acc.
–	S	Non-robust	304.20	MRMR	Gaussian Process	1.00	70%
–	S	Non-robust	365.00	MIFS	MLP	0.98	93%
–	S	Non-robust	425.80	CIFE	Adaboost	1.00	94%
–	S	Non-robust	540.00	MRMR	MLP	1.00	95%
–	M	Non-robust	304.20	MRMR	Gaussian Process	0.50	52%
–	M	Non-robust	365.00	MIFS	MLP	0.42	50%
–	M	Non-robust	425.80	CIFE	AdaBoost	0.49	51%
–	M	Non-robust	540.00	MRMR	MLP	0.51	55%
MR	S	Robust	304.20	RELF	Random Forest	0.76	69%
MR	S	Robust	365.00	RELF	Nearest Neighbors	0.72	52%
MR	S	Robust	425.80	RELF	XGBoost	0.81	71%
MR	S	Robust	540.00	GINI	Decision Tree	0.69	68%
MR	M	Robust	304.20	RELF	Random Forest	0.54	50%
MR	M	Robust	365.00	RELF	Nearest Neighbors	0.57	51%
MR	M	Robust	425.80	RELF	XGBoost	0.49	45%
MR	M	Robust	540.00	GINI	Decision Tree	0.46	43%
H	S	Robust	304.20	CIFE	XGBoost	0.90	75%
H	S	Robust	365.00	MRMR	AdaBoost	0.78	69%
H	S	Robust	425.80	MRMR	AdaBoost	0.82	67%
H	S	Robust	540.00	GINI	AdaBoost	0.74	68%
H	M	Robust	304.20	CIFE	XGBoost	0.66	57%
H	M	Robust	365.00	MRMR	AdaBoost	0.54	58%
H	M	Robust	425.80	MRMR	AdaBoost	0.51	50%
H	M	Robust	540.00	GINI	Adaboost	0.48	43%

Abbreviations: MR sequence prior, H hand-picked, S single-center, M multi-center, Bal. Acc. balanced accuracy, MRMR minimum redundancy maximum relevance, MIFS mutual information feature selection, CIFE conditional infomax feature extraction, RELF ReliefF, GINI Gini index, CMIM conditional mutual information maximization, MLP multi-layer perceptron, RBF SVC support vector classifier with radial basis function kernel. The full table with all performance metrics is reported in the supplementary material