Determining an Optimal Oxygen Saturation Target Range Based on Neonatal Maturity: Demonstration of a Decision Tree Analytic

The utility of decision tree machine learning in exploring the interactions among the SpO2 target range, neonatal maturity, and oxemic-risk is demonstrated. METHODS: This observational study used 3 years of paired age-SpO2-PaO2 data from a neonatal ICU. The CHAID decision tree method was used to explore the interaction of postmenstrual age (PMA) on the risk of extreme arterial oxygen levels at six different potential SpO2 target ranges (88–92%, 89–93%, 90–94%, 91–95%, 92–96% and 93–97%). Risk was calculated using a severity-weighted average of arterial oxygen outside the normal range for neonates (50–80 mmHg). RESULTS: In total, 7500 paired data points within the potential target range envelope were analyzed. The two lowest target ranges were associated with the highest risk, and the ranges of 91–95% and 92–96% were associated with the lowest risk. There were shifts in the risk associated with PMA. All the target ranges showed the lowest risk at ≥42 weeks PMA. The lowest risk for preterm infants was within a target range of 92–96% with a PMA of ≤34 weeks. CONCLUSIONS: This study demonstrates the utility of decision tree analytics. These results suggest that SpO2 target ranges that are different from typical range might reduce morbidity and mortality. Further research, including prospective randomized trials, is warranted.


Background
Preterm infants spend many weeks after birth in the neonatal intensive care unit (NICU).As a consequence of the instability of their underdeveloped respiratory system, supplemental oxygen and ventilatory support are usually required for prolonged periods.Further, their immaturity also makes them highly susceptible to excess mortality and morbidity.Pulmonary, retinal and neurodevelopmental morbidity are specifically associated with not only inadequate oxygenation but also excessive oxygen [1][2][3].
Due to their respiratory instability, managing neonatal oxygenation is challenging, requiring frequent adjustments to the inspired oxygen [4].While intermittent arterial blood gas measurements are feasible, with limited arterial access, they are not an acceptable tool for monitoring oxygenation in unstable neonates.Rather, arterial sampling is used primarily for periodic assessments of the acid-base balance.The standard of care for managing oxygenation in neonates is to obtain continuous noninvasive peripheral oxygen saturation (SpO 2 ) measurements [5,6].While SpO 2 is continuous, it is not as accurate as arterial measurements and is relatively insensitive to changes in higher levels of arterial oxygen (PaO 2 ) [7].Because these infants are highly susceptible to both high (hyperoxemia) and low (hypoxemia) oxygen, SpO 2 is managed in a tight target range.Currently, the usual SpO 2 target range in the neonatal ICU is 90-95% [8][9][10][11].However, European guidelines recommend 90-94% for extremely preterm infants [5].Finally, the target range has changed dramatically in the last several decades.Based on pragmatic outcome studies, it has shifted from previously higher norms (95-98%) [12] to rather lower levels (85-93%) [13,14].While it is now between 90 and 95%, this range is far from a consensus.Some think the optimum range should be either slightly lower [15] or slightly higher [16,17].
Most also believe that the maturity of the infant should be a factor in determining the target range, with slightly higher oxygen levels being the norm.The traditional transition from prematurity is above 36 weeks postmenstrual age (PMA).Maturity is a factor for two reasons.The first is that excess oxygenation in significantly premature infants has long been associated with blindness [3].While this is now rare, retinal morbidity remains a common problem.However, increased exposure to intermittent low oxygenation has been shown to increase mortality [2] and has even been associated with retinal morbidity [18].Secondly, the relationship between arterial oxygenation (PaO 2 ) and oxygen saturation varies in preterm infants as they transition from fetal to adult hemoglobin.This gestational age hemoglobin transition effect has its primary impact via the increased risk of high arterial PaO 2 associated with adult hemoglobin.Clinically, relevant differences in the risk of high oxygen between preterm and near-term infants have been demonstrated.[19].
According to standard nomenclature, infants born between 37 and 39 weeks are characterized as late preterm, those born ≤36 weeks gestation are characterized as preterm and those born ≥40 weeks are characterized as term [20].To normalize for differences in gestational and postnatal ages, PMA is commonly used in neonatology for the assessment of outcomes.It is determined by adding the gestational age to the postnatal age.The assessment of neonatal outcomes, to normalize for gestational age, is based on status at 36 weeks PMA.Thus, an infant born at 24 weeks gestational age would be assessed when they are 12 weeks old, while an infant born at 35 weeks would be assessed when they are one week old.The use of 36 weeks PMA is based on the premise that an infant's physiological development continues in the same trajectory after birth as during intrauterine maturation.However, applying this traditional outcome assessment criteria to the selection of maturitybased SpO 2 targets might not be appropriate.
The roots of machine learning or artificial intelligence were established over half a century ago.The exponential growth of computer processing power, coupled with the development of alternative analytical approaches, has resulted in it becoming embedded into our daily lives.Its use in healthcare has become essential for administrative tasks such as optimizing scheduling and resource management.Further, the opportunities for computer-enhanced medical devices and diagnostic/therapeutic decision support are apparent.Originating from decision support practice, automated decision trees have become common tools in machine learning [21].Given a data set, they parcel it into distinctly different subsets called leaves.Different decision tree methods use distinct approaches.However, the simplicity of the approach and the comprehensible nature of the results are common attractive traits.This is in contrast to the "black box" methodology of other machine learning approaches, such as fuzzy logic, random forests and neural networks.
The aim of this study was to demonstrate how a machine learning decision tree analysis could be used to determine optimal cutoffs for maturity to minimize the risk of inappropriate oxygenation in different potential target SpO 2 ranges.

Design
This was an observational study using data from a data repository at a single NICU.The collection and use of the data were approved by the requisite hospital research ethics committee.The need for informed consent was waived as there was no research intervention and the patient identities are masked.

Population
The Children's Hospital Los Angeles, University of Southern California Keck School of Medicine, is a major regional tertiary care referral center.Its neonatal services include a 58-bed NICU.The hospital does not include an obstetrics service, and admissions to the level 4 NICU are referred from other NICUs for a higher level of care.The patient data repository represents 3 years of experience of neonates receiving supplemental oxygen and mechanical ventilatory support.It includes 26,162 paired PMA-SpO 2 -PaO 2 measurements.

Data Collection
Details of the collection of the requisite study data from the patient data repository were provided previously described.[19].In summary, the query strategy included several steps.First, an average SpO 2 value was calculated for each arterial blood gas measurement.SpO 2 data from the monitoring system were stored every 30 s.The average value was the mean of all values within 120 s of the arterial sample.Second, the gestational age was accessed for each neonate.The PMA was then calculated as the sum of the birth gestational age and the postnatal age at the time of the arterial sample.

Endpoints
For this evaluation, six potential SpO 2 target ranges were defined that spanned the range of most practices [8][9][10][11].These ranges were centered at six median SpO 2 levels (90, 91, 92, 93, 94, 95%) with a span of ±2% SpO 2 .While target ranges often have a span wider than ±2, this narrower span was selected to facilitate the detection of subtle differences across the possible target ranges.Thus, the SpO 2 ranges evaluated were 88-92%, 89-93%, 90-94%, 91-95% 92-96% and 93-97%.Maturity was defined as the estimated gestational age at birth plus the postnatal age (PMA) and was described in weeks and days (week days ).The paired PMA-SpO 2 -PaO 2 in the database reflect when blood gas measurements were ordered, the timing of which might be associated with a standing order, a clinical exacerbation or an assessment of the effectiveness of weaning from respiratory support.Thus, it was not a random sample of oxygenation when stable.Therefore, 750 paired measurements were randomly selected from each of the 10 SpO 2 bins (88-97%).
We labeled the primary dependent variable oxemic-risk, which was quantified using a risk score assigned to each measurement.The score was based on a risk-weighted average of measurements of PaO 2 outside the normal range for neonates (50-80 mm of Hg).The weights were based on a subjective estimate.More relative weight was placed on hypoxemia than on hyperoxemia, based on recent outcome studies [1,2].These studies found that lower levels of SpO 2 were associated with increased mortality, while exposure to higher levels was associated with poorer developmental outcomes.Significant levels of hypoxemia (PaO 2 ≤ 40 mmHg) were scored as 6, and those between 41 and 49 mmHg were scored as 2. Oxemic-risk associated with hyperoxemia was scored as follows: ≥200 mmHg as 6, 100-199 mmHg as 3 and 81-99 mmHg as 1.

Statistical Considerations
The primary aim of this analysis was to determine the PMA thresholds, if any, at which the oxemic-risk shifted for each of the six different potential target ranges.The CHAID decision tree method was used to determine these PMA break points [22].CHAID (Chisquared Automatic Interaction Detector) was selected from other decision tree methods for two reasons.First, other decision tree machine learning analytics are binary and multitiered and thus provide hierarchical output.The CHAID analysis is nonbinary.Thus, it can identify the number of significantly different subgroups, whether there are zero or many.Further, it uses a traditional chi-squared method to partition the data into the relevant number of statistically significant categories.This avoids the lack of confidence associated with some machine learning approaches that utilize unintelligible or "black box" analysis methods.For illustrative purposes, Figure 1 shows the decision tree for oxemicrisk for all 21,162 paired measurements and was created using CHAID.In this case, Four unique PMA partitions were identified.The number of measurements, their proportion and the oxemic-risk are shown for each.As an example, for the lowest of the four PMA partitions (≤37 weeks), there were 5256 measurements (20.1%), with an oxemic-risk of 1.87.The oxemic-risk values for the four partitions (1.87, 2.45, 2.67 and 2.26) are statistically significantly different.
Diagnostics 2023, 13, x FOR PEER REVIEW 4 of 10 "black box" analysis methods.For illustrative purposes, Figure 1 shows the decision tree for oxemic-risk for all 21,162 paired measurements and was created using CHAID.In this case, Four unique PMA partitions were identified.The number of measurements, their proportion and the oxemic-risk are shown for each.As an example, for the lowest of the four PMA partitions (≤37 weeks), there were 5256 measurements (20.1%), with an oxemicrisk of 1.87.The oxemic-risk values for the four partitions (1.87, 2.45, 2.67 and 2.26) are statistically significantly different.Populations included all data points regardless of SpO2.PMA is in weeks days .Subsequent to the primary analysis, two descriptive analyses were conducted.The first post hoc analysis was calculation of the oxemic-risk score for a common recommendation (90-94% SpO2 target range for PMA partitions ≤36 weeks) and contrast it with those with >36 weeks PMA.The second was a calculation of the oxemic-risk for those SpO2 measurement outside the target range envelop.These were hypoxemic SpO2 measurements (<88%) and hyperoxemic SpO2 measurements (>97%).
Demographics of the neonates, and a variety of cross tabulations of SpO2 and PaO2 were compiled.The CHAID-derived partitions were, by definition, statistically significantly different.Confidence limits of the descriptive oxemic-risk scores were calculated, with statistical significance inferred if they did not overlap.The CHAID and other analyses were conducted using XLSTAT v11.5 (Lumivero, New York, NY, USA).

Results
In the analysis, there were 3750 paired data points in each of the six specified target ranges and 7500 in the envelope of normoxemia between 88 and 97% SpO2.These data came from 888 infants with a median gestational age of 37 weeks (IQR 32-39) at birth.The median number of samples per subject was five (IQR 2-10).The median postnatal age of the infants at the time of the blood gas measurements was 2.0 weeks (IQR 0.9-4.9).The PMA, SpO2 and PaO2 levels for these neonates are shown in Table 1.For comparison, the corresponding values above and below the SpO2 target range envelope (88-97%) are also shown.The PMA values were comparable, and in confirmation of the expectation, the oxygenations levels are markedly different (SpO2: 94, 80 and 99%, and PaO2: 42, 74 and 129 mmHg).Populations included all data points regardless of SpO 2 .PMA is in weeks days .Subsequent to the primary analysis, two descriptive analyses were conducted.The first post hoc analysis was calculation of the oxemic-risk score for a common recommendation (90-94% SpO 2 target range for PMA partitions ≤36 weeks) and contrast it with those with >36 weeks PMA.The second was a calculation of the oxemic-risk for those SpO 2 measurement outside the target range envelop.These were hypoxemic SpO 2 measurements (<88%) and hyperoxemic SpO 2 measurements (>97%).
Demographics of the neonates, and a variety of cross tabulations of SpO 2 and PaO 2 were compiled.The CHAID-derived partitions were, by definition, statistically significantly different.Confidence limits of the descriptive oxemic-risk scores were calculated, with statistical significance inferred if they did not overlap.The CHAID and other analyses were conducted using XLSTAT v11.5 (Lumivero, New York, NY, USA).

Results
In the analysis, there were 3750 paired data points in each of the six specified target ranges and 7500 in the envelope of normoxemia between 88 and 97% SpO 2 .These data came from 888 infants with a median gestational age of 37 weeks (IQR 32-39) at birth.The median number of samples per subject was five (IQR 2-10).The median postnatal age of the infants at the time of the blood gas measurements was 2.0 weeks (IQR 0.9-4.9).The PMA, SpO 2 and PaO 2 levels for these neonates are shown in Table 1.For comparison, the corresponding values above and below the SpO 2 target range envelope (88-97%) are also shown.The PMA values were comparable, and in confirmation of the expectation, the oxygenations levels are markedly different (SpO 2 : 94, 80 and 99%, and PaO 2 : 42, 74 and 129 mmHg).The PaO 2 and SpO 2 values for the measurements in the target range (SpO 2 88-97%) for each of the six PaO 2 oxemic-risk strata are provided in Table 2.The PaO 2 levels are consistent with the six oxygenation strata, increasing from a low of 37 ± 4 to a high of 273 ± 64 mmHg.In contrast, while the SpO 2 increased by one mmHg for each of the two steps of hypoxemia, it is insensitive to increases in PaO 2 between the three higher PaO 2 risk strata that are associated with hyperoxemia.It is also of note that nearly half (42%) of the PaO 2 readings associated with SpO 2 readings between 87 and 97%, are outside the normal of PaO 2 range of 50-80 mmHg.This is reflective of a known limitation in the precision of SpO 2 measurements.Table 3 describes the SpO 2, PaO 2 and associated oxemic-risk scores for the six SpO 2 target ranges.The PaO 2 increases from 56 ± 24 mmHg for the target range of 88-92% SpO 2 , with increasingly higher SpO 2 target ranges of up to 80 ± 41 mmHg for the highest target range (93-97% SpO 2 ).The large standard deviations of the PaO 2 in each target range also reflect the lack of precision in the SpO 2 -PaO 2 relationship.These oxemic-risk scores are "U shaped".The highest (1.24) is associated with the lowest SpO 2 target range (88-93%).SpO 2 target ranges of 91-95% (0.83) and 92-96% (0.80) are at the nadir, and the 95% confidence limits of their oxemic-risk scores indicate that they are not different.The highest SpO 2 target range (93-97%), just 1% SpO 2 higher, has a higher oxemic-risk score of 0.88.Table 4 summarizes the ideal PMA oxemic-risk partitions derived from the CHAID decision tree for each SpO 2 target range.For all six target ranges the upper partition cut off was during the 42nd week of PMA.The oxemic-risk for these older neonates was lower at each SpO 2 target range than the oxemic-risk for the less mature infants.Three of these PMA partitions were dichotomous (89-93%, 90-94% and 91-95%), that is there was no significant additional partitions below 42 weeks.An additional partition was identified for the other 3 target ranges, though at different PMA thresholds (38, 39, 34 weeks).The lowest oxemic-risk for immature infants was with the 92-96% target range and ≤34 weeks PMA.In addition, the oxemic-risk was markedly higher in the two lower target ranges (88-92%, 89-93%) than the other 4 target ranges for each PMA partition.Interestingly the oxemic-risk for the middle PMA partition, when present, is higher than the low and high PMA partitions.
We also determined the oxemic-risk for the recommended target range of 90-94%, above and below the usual transition at 36 weeks (≤36 weeks 0.97 (95% CI 0.89-1.06),>36 weeks 0.94 (0.88-1.00)).These were not statistically different and were also markedly higher risks than the optimal target ranges identified via the decision tree analysis.Shown graphically in Figure 2 are these oxemic-risk scores and PMA partitions, along with the optimal 92-96% target range, as well as those for SpO 2 < 88% and SpO 2 > 97%.The oxemicrisk when the SpO 2 is not in the target range is nearly an order of magnitude larger than when it is in the target range.The difference between the two target ranges for PMA ≤ 34 and ≥42 is also marked.We also determined the oxemic-risk for the recommended target range of 90-94%, above and below the usual transition at 36 weeks (≤36 weeks 0.97 (95% CI 0.89-1.06),>36 weeks 0.94 (0.88-1.00)).These were not statistically different and were also markedly higher risks than the optimal target ranges identified via the decision tree analysis.Shown graphically in Figure 2 are these oxemic-risk scores and PMA partitions, along with the optimal 92-96% target range, as well as those for SpO2 < 88% and SpO2 > 97%.The oxemicrisk when the SpO2 is not in the target range is nearly an order of magnitude larger than when it is in the target range.The difference between the two target ranges for PMA ≤ 34 and ≥42 is also marked.

Discussion
We used a machine learning decision tree to evaluate the complex interaction of two parameters: infant maturity and six different potential SpO2 target ranges.The goal was to determine the best target range for infants of differing maturities based on minimizing oxemic-risk.Our results suggest that current practice of using a target range of 90-95%

Discussion
We used a machine learning decision tree to evaluate the complex interaction of two parameters: infant maturity and six different potential SpO 2 target ranges.The goal was to determine the best target range for infants of differing maturities based on minimizing oxemic-risk.Our results suggest that current practice of using a target range of 90-95% and transitioning to higher levels at >36 weeks might not be ideal for minimizing the risk associated with excessive and inadequate oxygen.
We found that the lowest oxemic-risk occurred in a target range of 92-96% SpO 2 but, as expected, the risk varied depending on the neonate's maturity.The lowest oxemicrisk occurred in term infants (PMA ≥ 42 weeks) regardless of the target range.Among these mature term infants, a target range of 91-95% resulted in a reduced oxemic-risk ranging between 5 and 72% compared to the other five target ranges.We believe this target range is lower than in typical practice.Though little has been published to support our perception, the finding is consistent with another analysis.[19] For extremely preterm infants (PMA ≤ 34 weeks), a target range of 92-96% was associated with a reduced oxemicrisk ranging between 11 and 70% compared to the other five target ranges.Specifically, it was a 55% reduction from the European recommendation of 90-94% [5].This 34-week cut off was also two weeks less than the standard practice of ≥36 weeks.For those infants between 35-41 weeks PMA, the oxemic-risk was the lowest for the target ranges of 91-95% and 92-96%.
In summary, the more common practice is to use a target range of 90-95% for preterm infants [8][9][10], and some advocate using lower targets until 36 weeks PMA [13,14].In contrast, our analysis of oxemic-risk suggests a target range of 92-96% is better, particularly for extremely preterm infants.Nevertheless, this was a focused analysis and only one part of the puzzle.Other factors also ought to be considered [6].One of these factors is the percent of hemoglobin that is fetal.Another is the ability of the clinical staff to be compliant with the intended target range.Several other factors beyond the scope of this discussion are body temperature, congenital anomalies, the severity of pulmonary dysfunction, and the total hemoglobin level.For this reason, the recommendation of the American Academy of Pediatrics suggests that targeting should be more patient-specific rather than uniform [6].
The relationship between arterial oxygen saturation and oxygen partial pressure is complex.First the relationship shifts according to the concentration of fetal hemoglobin.Thus, the concentration of fetal hemoglobin, associated with neonatal development, results in a need for maturity-based target ranges.However, the predicable transition over several months shifts with blood transfusions.In addition, the acid-base balance, temperature and other factors also result in clinically relevant shifts.Second, the SpO 2 is a noninvasive measurement of arterial oxygen saturation (SaO 2 ).These two factors account for much of the scatter between the predicted PaO 2 and the measured SpO 2 .Finally, the relationship is not linear but rather sigmoid, and the clinically relevant part, above 50% saturation, has been described by a seventh order polynomial [23].This sigmoid relationship explains the poor sensitivity of high SpO 2 levels to increasingly higher PaO 2 s above the target range as the relationships becomes asymptotic with 100% SpO 2 .In contrast, there is linear drop in PaO 2 and SpO 2 values below the target range, making it reliable for detecting hypoxemia, even if the exact value is not precise.Nevertheless, SpO 2 is the standard of care for monitoring oxygenation.When considering trade-offs between target ranges it is also highly relevant that lower SpO 2 target ranges, as the result of a linear drop in the PaO 2 , have been shown to limit the oxygenation reserve of the infant and thus result in more frequent episodes of hypoxemia [24,25].Therefore, lower target ranges, while reducing the risk of hyperoxemia with a lower average PaO 2 , also increase the prevalence of hypoxemic episodes.This consideration provides additional support to our findings that higher target ranges for less mature neonates would reduce oxemic-risk.
Until recently, oxygen levels in the NICU were managed exclusively by manual transient increases in the inspired oxygen in response to desaturations.This approach has been shown to result in only about half of the time being spent in the desired SpO 2 target range, with an additional marked bias between hypoxemia and hyperoxemia among centers and caregivers [4,11].Of late, many new neonatal ventilators offer closed-loop control over inspired oxygen based on SpO 2 .This approach to titrating inspired oxygen has consistently been shown to improve time spent in the intended target range and to reduced time spent at SpO 2 extremes [26].Aside from enhanced general effectiveness, these automatic control systems not only make the use of narrower target ranges practical but also offer the possibility of regular changes to target ranges that are individualized for each neonate.
The design of this study and its results are somewhat unique.This is, to our knowledge, the first study to use a severity-weighted risk score in addition to one that balances the risk between hypoxemia and hyperoxemia.In addition, only one other study has examined the exposure to aberrant arterial PaO 2 levels when within the intended target range [19].Studies of neonatal SpO 2 targeting have primarily investigated maintaining compliance with the SpO 2 target and the amount of time spent outside a target range and at SpO 2 extremes [26].Our results focus on this balanced tradeoff of oxemic-risk when in the target range and also provide a contrast to situations in which saturations are above or below normoxemia.The unique use of machine learning also suggests that traditional transitions in risk, based on maturity, ought to be reexamined.This approach also identified an interim maturity level with elevated risk.While it is well understood that preterm neonates experience a period of increased instability when they are 2-4 weeks old, which clearly affects the ability to maintain normoxemia, our finding suggests that even in the target range, they have an increased oxemic-risk.
Another similar analysis which only considered the maintenance of PaO 2 between 50 and 80 mmHg found no difference in similar ranges and in preterm vs. near term infants.[19] Our analysis considered the severity of the risk of values outside the desired PaO 2 range rather than the prevalence within.Our analysis used one scheme to assign weights to different levels of hyperoxemia and hypoxemia for all infants.Though defendable, it can be challenged as perhaps overly simplistic.Nevertheless, it is not clear what a better weighting strategy might be.Some experts argue that extremely preterm infants have no mechanisms to protect themselves from excessive oxygenation and that prior to delivery, their fetal oxygen was <40 mmHg, and they are thus are at added risk [14,15].In contrast others suggest that fetal hemoglobin limits the likelihood of high arterial oxygenation and thus is protective [16,19].
Of note, the data used in this study came from a level 4 tertiary care center.This permitted an analysis of the experience of neonates on extended respiratory support over a three-year period that might only be treated over decades in most of the referring NICUs.We believe it is generally applicable to other neonates receiving respiratory support, albeit for shorter periods of time.It is not generally applicable to neonates just receiving supplement oxygen.However, consideration should be given to the increased likelihood that the neonates in our study population received one or more transfusions of adult blood and thus their shift to adult hemoglobin was accelerated.Nevertheless, we would expect that this would increase exposure to hyperoxemia in our population and thus probably reinforce our overall finding that higher ranges for preterm infants might be preferable.However, it might well change the relative oxemic-risk scores and the PMA transition points that our study identified.
In addition to the issues discussed above, our study has other limitations.First, this was an observational study, the results of which should always be considered with caution.Second, the level of inspired oxygen might also play a role in the risk of hyperoxemia and was not considered.Severe PaO 2 levels are much less likely, though possible, when inspired oxygen is weaned towards room air.However, most hyperoxemia in infants is a result of administering excess oxygen following a period of hypoxemia, which can happen at any baseline inspired oxygen level [27].Finally, the analysis was based on a traditional normal PaO 2 range of 50-80 mmHg.This range has not been, and never will be, carefully correlated to normal outcomes.Rather, SpO 2 monitoring is the standard of care, and studies are needed to explore the impact of SpO 2 target ranges higher than 90-94% or 90-95%.
Our findings have some important implications related to needed research.Further prospectively designed observational studies in other subject populations are needed.With the increasing number of centers collecting and archiving data from integrated patientmonitoring systems, this is practical and efficient.Specific study populations should include extremely preterm neonates in the early stage following birth as a well as neonates who have not been referred to tertiary centers for a higher level of care.Further analyses of the sensitivity of changes in oxemic-risk weighting as well as oxemic-risk scores that change with maturity are warranted.These results would contribute to the design of randomized controlled multicenter trials essential to supporting rationale recommendations for SpO 2 target ranges not just for extremely preterm neonates but for all neonates requiring respiratory support.

Conclusions
Our study demonstrated the usefulness of an automated decision tree to quickly partition data into statistically different groups, independent of historical rubrics.It was found to be an effective analytical tool.Our results also suggest that SpO 2 target ranges slightly

Figure 1 .
Figure 1.Decision tree for oxemic-risk based on PMA.

Figure 1 .
Figure 1.Decision tree for oxemic-risk based on PMA.

Figure 2 .
Figure 2. Comparison of oxemic-risk within and outside target ranges.

Figure 2 .
Figure 2. Comparison of oxemic-risk within and outside target ranges.

Table 1 .
PMA, PaO2 and SpO2 in, below and above the target ranges.

Table 1 .
PMA, PaO 2 and SpO 2 in, below and above the target ranges.

Table 4 .
PMA ranges and associated oxemic-risks for each SpO 2 target range.The three center SpO 2 target ranges are dichotomous.PMA weeks days .The three center SpO2 target ranges are dichotomous.PMA weeks days .