Table 1.

This table shows the F1-score performance of the DeepTRIAGE attention model for luminal sub-type classification according to a single test set

	Logistic Regression	Linear SVM	DeepTRIAGE
GO (BP) annotations	0.87	0.89	0.90
KEGG annotations	0.86	0.84	0.87
Ensembl genes	0.85	0.85	0.87

We benchmark its performance as compared to a logistic regression and support vector machine (SVM), using both gene and gene set annotation features. From this, we see that our model, which adds a level of interpretability at the individual level, does not sacrifice classification accuracy. The objective of DeepTRIAGE is to improve interpretability, not accuracy. Yet, this method appears to perform marginally better for the given test set