. 2019 Nov 8;7(4):e14340. doi: 10.2196/14340

Table 2.

Performance of the SVM (support vector machine) baseline and HYPE (High-Performing System for Automatically Detecting Hypoglycemic Events) based on different kinds of neural networks.

Performance measures	SVM	P value^a	LSTM^b	P value	Bi-LSTM^c	P value	TCN^d	P value	CNN^e	P value
Precision, mean (SD)	0.74 (0.07)	<.001	0.91 (0.02)	<.001	0.91 (0.02)	<.001	0.92 (0.03)	.05	0.96 (0.03)	N/A^f
Recall, mean (SD)	0.57 (0.05)	<.001	0.86 (0.02)	.02	0.87 (0.04)	.10	0.89 (0.04)	N/A	0.86 (0.03)	.10
F1, mean (SD)	0.64 (0.03)	<.001	0.88 (0.02)	<.001	0.88 (0.02)	.001	0.90 (0.02)	.30	0.91 (0.02)	N/A
PR-AUC^g	0.745	N/A	0.934	N/A	0.942	N/A	0.964	N/A	0.966	N/A
ROC-AUC^h	0.970	N/A	0.996	N/A	0.997	N/A	0.998	N/A	0.998	N/A

^aP values are based on two-sample t tests between the performance of the system and the best-performing system; values <.05 are significant.

^bLSTM: long short-term memory.

^cbi-LSTM: bidirectional long short-term memory.

^dTCN: temporal convolutional neural network.

^eCNN: convolutional neural network.

^fN/A: not applicable.

^gPR-AUC: precision-recall area under the curve.

^hROC-AUC: receiver operating characteristic area under the curve.