Table 2.

A summary of the AUROC for the forecasting tasks

Backbone model	End-to-end supervised training	CEF-CL	Improvement
VUMC (181 tasks)
RNN	0.607 (0.584–0.629)	0.689 (0.677–0.702)	13.5%
GRU	0.637 (0.624–0.650)	0.680 (0.668–0.693)	6.8%
LSTM	0.638 (0.617–0.657)	0.684 (0.670–0.697)	7.2%
CONAN	0.633 (0.620–0.647)	0.668 (0.652–0.679)	5.5%
LSAN	0.618 (0.597–0.637)	0.662 (0.649–0.677)	7.1%
All of Us (120 tasks)
RNN	0.572 (0.544–0.609)	0.773 (0.758–0.791)	35.1%
GRU	0.683 (0.660–0.701)	0.772 (0.753–0.788)	13.0%
LSTM	0.690 (0.656–0.719)	0.784 (0.767–0.798)	13.6%
CONAN	0.717 (0.700–0.735)	0.761 (0.742–0.779)	6.1%
LSAN	0.670 (0.628–0.702)	0.764 (0.746–0.780)	14.0%

Note: In this table, the results are depicted as a (b−c), where a represents the average performance score calculated of three independent runs, while b and c represent the minimum score and the maximum score.