Table 3.

False-positive rates on the test set using threshold selection based on ROC optimization (ROC) or on a significance-level requirement (FPR) of α ≤ 10⁻⁴

Scheme	NSR	LP-NSR	AF	SVT	VT	Vent. big.	Vent. trig	IR	At. big.	Brady.
ROC	10−4	0.31	0.03	2 · 10−3	7 · 10−4	2 · 10−3	10−3	0	10−3	5 · 10−3
FPR	10−5	0.41	10−5	5·10−5	10−5	4 · 10−5	6·10−5	0	10−5	10−5

Thresholds were selected based solely on a validation-set (disjoint patient-wise from the train set). The model with ROC scheme fails to uphold the required FPR for all but one rhythm (IR), while the model with the FPR scheme successfully upholds the requirement for all rhythms but LP-NSR, which the model occasionally confuses with NSR (recall these are morphologically identical for cardiologists). A failure to uphold the statistical requirement is highlighted in red. SVT, supraventricular tachycardia; Vent. big, ventricular bigeminy; Vent. trig., ventricular trigeminy; IR, idioventricular rhythm; At. big., atrial bigeminy; Brady., sinus bradycardia.