A deep learning approach identifies new ECG features in congenital long QT syndrome

Table 2 Model performances on Leuven data

Training	Type	External validation (Leuven data)
Training	Type	Sensitivity ± SD	Specificity ± SD	AUC ± SD
First ECG approach (Amsterdam data)	LQTS1	80 ± 2%	94 ± 2%	0.86 ± 0.01
First ECG approach (Amsterdam data)	LQTS2	92 ± 3%	81 ± 6%	0.87 ± 0.02
All ECG approach (Amsterdam data)	LQTS1	87 ± 4%	93 ± 3%	0.90 ± 0.01
All ECG approach (Amsterdam data)	LQTS2	90 ± 4%	88 ± 3%	0.89 ± 0.01

The whole set of LQTS1 (n = 32), LQTS2 (n = 80), and controls (n = 2280) from the Leuven dataset was used to validate our DL models, which were trained on the Amsterdam data using the only the first acquired 12-lead ECGs (i.e., First ECG approach) (top) or all acquired 12-lead ECGs (i.e., ALL ECG approach) (bottom). The mean and standard deviation (SD) of the collected metrics is reported

ISSN: 1741-7015