Table 1.

ReceptorNet performs well across different splits of the data based on other patient characteristics and also outperforms baseline methods.

Algorithm (data split)	AUC (95% CI)	Algorithm (data split)	AUC (95% CI)
ReceptorNet (cross-validation set)	0.899 (0.884–0.913)	ReceptorNet (test set)	0.920 (0.892–0.946)
Baselines
Meanpool (test set)	0.827 (0.786–0.866)	Maxpool (test set)	0.880 (0.846–0.912)
Individual patch (test set)	0.760 (0.726–0.794)	Logit on type and grade (test set)	0.809 (0.766–0.848)
Data splits based on other hormone receptors and grade
ReceptorNet (HER2+, test set)	0.768 (0.719–0.813)	ReceptorNet (HER2−, test set)	0.927 (0.912–0.943)
ReceptorNet (PR+, test set)	0.906 (0.869–0.940)	ReceptorNet (PR-, test set)	0.827 (0.795–0.855)
ReceptorNet (Grade 1, test set)	0.949 (0.925–0.973)	ReceptorNet (Grade 2, test set)	0.810 (0.716–0.888)
ReceptorNet (Grade 3, test set)	0.865 (0.840, 0.887)
Data splits based on data source
ReceptorNet (TCGA, cross-validation set)	0.861 (0.828–0.893)	ReceptorNet (TCGA, trained on ABCTB alone)	0.850 (0.830–0.868)
ReceptorNet (ABCTB, cross-validation set)	0.905 (0.889–0.921)	ReceptorNet (University of Pittsburgh, trained on rest)	0.910 (0.836–0.969)
Data splits based on demographics
ReceptorNet (postmenopausal women, TCGA)	0.872 (0.832–0.908)	ReceptorNet (premenopausal women, TCGA)	0.838 (0.779–0.893)
ReceptorNet (African-American patients, TCGA)	0.859 (0.785–0.921)	ReceptorNet (Non-African-American patients, TCGA)	0.858 (0.817–0.896)