Table 3.

Epoch-specific cut-ff cross-validation.

	0–6 h			6–24 h			24–48 h			48–72 h
Fold	Cut-off	Sens	Spec	Cut-off	Sens	Spec	Cut-off	Sens	Spec	Cut-off	Sens	Spec
	(80% Training)	(20% Validation)		(80% Training)	(20% Validation)		(80% Training)	(20% Validation)		(80% Training)	(20% Validation)
1	33.0°C	92%	50%	31.0°C	75%	27%	32.5°C	45%	59%	32.0°C	64%	50%
2	33.0°C	83%	30%	35.0°C	25%	76%	29.0°C	83%	24%	32.0°C	64%	36%
3	33.0°C	88%	45%	31.0°C	100%	29%	32.5°C	29%	75%	31.5°C	100%	45%
4	33.0°C	100%	57%	31.0°C	100%	38%	31.0°C	80%	57%	32.0°C	90%	57%
5	33.0°C	100%	58%	31.0°C	100%	0%	31.0°C	85%	50%	31.5°C	85%	38%
	Cut-off	Sens	Spec	Cut-off	Sens	Spec	Cut-off	Sens	Spec	Cut-off	Sens	Spec
Overall	33.0°C	93%	49%	31.0°C	91%	30%	31.0°C	77%	54%	32.0°C	79%	48%

Epoch-specific cut-offs to predict the primary outcome were developed using 5-fold cross-validation. The data were randomly split into five folds such that each infant for whom at least 50% of expected MTs were available in that epoch appeared in four training folds and one validation fold. Within each fold in each epoch, average MT cut-offs were selected optimizing for the sum of sensitivity and specificity to predict outcome in unadjusted models. For each of the five training folds, the optimal MT cut-off and the sensitivity and specificity for predicting outcome in the validation fold is shown. To select a final cut-off for each epoch for outcome prediction, the cut-off identified in the most folds was selected. In the 24–48 h epoch two different MT cut-offs were selected in the same number of training folds (≥32.5°C and ≥31°C, n = 2 each), so the cut-off of ≥31°C was selected as it had a higher sensitivity for predicting the primary outcome in the associated validation folds. Overall sensitivity and specificity for the cut-offs across the entire cohort are also shown.