TABLE III.

In-hospital mortality prediction results by benchmarking LSTM model [5] under different losses. BCE: Binary Cross Entropy; CBCE: Contrastive Binary Cross Entropy; CSCE: Contrastive Softmax Cross Entropy; SCR: Supervised Contrastive Regularizer. We highlight the best performance w.r.t different metrics. We also report the standard deviation (std.) of bootstrapped results by re-sampling the test set 100 times with replacement in round brackets for reference.

	AUROC	AUPRC	Accuracy	min(Se, P+)
${ℒ}_{\text{BCE}}$	0.854(0.010)	0.483(0.031)	0.896(0.005)	0.487(0.026)
${ℒ}_{\text{BCE}}+\lambda {ℒ}_{\text{SCR}}$	0.858(0.009)	0.489(0.028)	0.892 (0.005)	0.487 (0.023)
${ℒ}_{\mathbf{C}\mathbf{B}\mathbf{C}\mathbf{E}}+\lambda {ℒ}_{\mathbf{S}\mathbf{C}\mathbf{R}}$	0.860(0.009)	0.504(0.031)	0.897(0.005)	0.482 (0.025)
${ℒ}_{\mathbf{C}\mathbf{S}\mathbf{C}\mathbf{E}}+\lambda {ℒ}_{\mathbf{S}\mathbf{C}\mathbf{R}}$	0.860(0.010)	0.501 (0.030)	0.893 (0.005)	0.505(0.024)