TABLE 3. Description of the Features Used for Developing the Decision Tree Classifiers.

Feature category (no. of features)	Feature name abbreviation	Feature description
Infant metadata (3)	GA	Gestational age in days
	BW	Birthweight in gram
	PNA	Postnatal age in days
Yellow alarm category (1)	Y_Alarm_Catb	The yellow alarm category could be one of the following – HR-low, HR-high, SpO2-low, SpO2-high, BP-low, and BP-high.
Alarm counts (2)	Count_Y_Alarm	No. of yellow alarms that occurred in the pre-alarm window
	Count_R_Alarm	No. of red alarms that occurred in the pre-alarm window
HR, BR and SpO2 based features (), with HR, BR, and SpO2 described by Parameter	Parameter_Occ	Feature value at the moment of occurrence of the yellow alarm
	Parameter_Min	Minimum value
	Parameter_Mean	Mean value
	Parameter_Std	Standard deviation
	Parameter_NTC_Y	No. of times the parameter crossed the yellow alarm threshold.c
	Parameter_NTC_R	No. of times the parameter crossed the red alarm threshold.
	Parameter_TUR	Cumulative time under red alarm threshold
	Parameter_DI	Delta index – the average of the absolute differences between the mean values of 2 successive and non-overlapping 12s intervals. [31]
	Parameter_CTM	Central tendency measure – the sum of distances to the origin of a second order difference plot of all points except the furthest 5% of all points.
	Parameter_ApEn	Approximate entropy – calculated using a run-length of 2 with a tolerance of 25% of the standard deviation of the data. [32]
	Parameter_LZC	Lempel-Ziv complexity – the median value was used as a threshold for binarization. [33]
	Parameter_Sloped	The slope of the regression line fitting the last 50s of data preceding the yellow alarm.
	Parameter_Rvaluee	Coefficient of correlation between actual values and those predicted by regression.
Correlation features of HR, BR and SpO2 ()	Max_Corr	The maximum cross-correlation between a parameter of window length one-third the length of the pre-alarm window immediately preceding the yellow alarm with the entire pre-alarm window of the other parameter at 5s intervals, without padding. Prior to cross-correlation, the parameters were normalized using the standard score.
	Lag_At_Max_Corr	The lag corresponding to Max_Corr
HRV based featurese	NN_Occf	NN-interval at occurring moment
	NN_AUCg	Area under the NN-intervals.
	SDNN_Occ	Standard deviation of NN-intervals at occurring moment.
	SDNN_AUC	Area under the SDNN time-series.
	RMSSD_Occ	Root mean square of the standard deviation of NN-intervals at occurring moment.
	RMSSD_AUC	Area under the RMSSD time-series.
	pNN50_Occ	Percentage of NN-intervals longer than 50ms at occurring moment.
	pNN50_AUC	Area under the pNN50 time-series.
	pDec_Occ	Percentage of NN-intervals longer than mean value of NN intervals[21]
	pDec_AUC	Area under the pDec time-series.
	SDDec_Occ	Standard deviation of all NN-intervals contributing to pDec[21]
	SDDec_AUC	Area under the SDDec time-series.

^aAll features were calculated using the entire pre-alarm window unless explicitly stated otherwise.

^bHR-low, HR-high, SpO2-low, SpO2-high, BP-low, and BP-high constituted all the yellow alarm categories.

^cThe thresholds for HR and SpO2 were acquired from the alarm logs themselves. For BR since there are no threshold-based alarms, thresholds of 25 and 30 were used for red and yellow alarms respectively, in discussion with clinicians.

^dThe choice of 50s was based on visual observations of data, as exemplified by Fig. 2 as well.

^eHRV features were calculated every 10s using a moving average window of the preceding 30s.

^fNN_Occ, in effect, is the mean value of NN-intervals in the 30s leading up to the yellow alarm. The same holds for other HRV feature based on the occurring moment.

^gThe area under the curve is calculated using the 10^th percentile value of NN-intervals acting as a baseline. The same holds for the other HRV feature based on AUC.