Table 1:

Model performance of the integrated model and Pathomic Fusion at the different missing rates of data. (All values represent mean value ± standard deviation computed from the testing sets of the 15 splits. The methods with the best performance are highlighted in bold)

Methods	Use data with missing modalities?	Evaluation Metrics	Missing Rate*
			0%	25%	50%	75%	90%	95%	99%
Proposed Integrated model (Pathomic + CPM-Nets)	Yes	AUC ↑	0.896 ± 0.009	0.895 ± 0.008	0.888 ± 0.011	0.887 ± 0.010	0.890 ± 0.010	0.886 ± 0.010	0.885 ± 0.011
		F1 score ↑	0.731 ± 0.018	0.738 ± 0.023	0.719 ± 0.023	0.716 ± 0.018	0.721 ± 0.024	0.702 ± 0.023	0.696 ± 0.023
		F1 Grade IV ↑	0.926 ± 0.011	0.922 ± 0.014	0.918 ± 0.010	0.910 ± 0.014	0.909 ± 0.015	0.896 ± 0.017	0.898 ± 0.016
Pathomic Fushion	No	AUC ↑	0.911 ± 0.010	0.905 ± 0.009	0.894 ± 0.010	0.883 ± 0.014	0.880 ± 0.012	0.874 ± 0.017	0.751 ± 0.076
		F1 score ↑	0.749 ± 0.021	0.740 ± 0.019	0.729 ± 0.017	0.711 ± 0.020	0.723 ± 0.018	0.699 ± 0.024	0.579 ± 0.084
		F1 Grade IV ↑	0.933 ± 0.015	0.928 ± 0.013	0.919 ± 0.013	0.903 ± 0.011	0.903 ± 0.018	0.887 ± 0.018	0.723 ± 0.172
Proposed Integrated model (Pathomic + CPM-Nets)	No	AUC ↑	0.896 ± 0.009	0.891 ± 0.009	0.893 ± 0.013	0.884 ± 0.011	0.881 ± 0.009	0.869 ± 0.011	0.786 ± 0.063
		F1 score ↑	0.731 ± 0.018	0.727 ± 0.021	0.723 ± 0.019	0.715 ± 0.021	0.705 ± 0.017	0.705 ± 0.017	0.660 ± 0.011
		F1 Grade IV ↑	0.926 ± 0.011	0.926 ± 0.009	0.921 ± 0.012	0.905 ± 0.013	0.901 ± 0.024	0.888 ± 0.020	0.815 ± 0.126

^*Indicates n% data in the training set had the complete two modalities pathological image and genomic data, while the rest of data in the training set had only one modality. Using data with missing modalities means using all available data, including the data with missing modalities at the feature fusion stage.