Prematurity is the main cause of perinatal mortality and accounts for 50% of cases of disability in childhood.1 In Spain, approximately 30,000 children are born preterm each year. Of these children, 10% are born at or before 32 weeks’ gestation or with a birth weight of 1500g or less.2,3

The subset of infants born extremely preterm (gestational age [GA] ≤28 weeks) and especially those born at the threshold of viability (GA, 23–25 weeks) pose a significant challenge for the professionals involved in their care and treatment. Due to their considerable immaturity, some of these patients may have medical problems for which there is no curative system and require palliative care. The corresponding decision-making process is long and complex for both health professionals and families.4 The main difficulty is making an accurate prognosis for the infant.

Prediction models are used to estimate the probability or risk of a certain condition occurring in the future in a given individual through the combination of different prognostic factors.5 Thus, while the diagnosis of a condition is essential to clinical decision-making, the incorporation of new prognostic tools allows a more accurate prediction of future health outcomes and can therefore guide decision-making to improve management on a case-by-case basis.6 Numerous factors influence the prognosis of a disease besides the diagnosis. For instance, in extremely preterm infants, gestational age can be used to predict the level and duration of care that they will receive (intensive, high-dependence or special care) and the probability of death or discharge home throughout the hospital stay.7

Extreme prematurity is a condition associated with a high mortality in which the risk of death varies based on the associated complications. It is important for both clinicians and family members to be able to estimate the individualised probability of death at different time points, preferably before birth, in the first hours post birth and during the hospital stay. In 2011, Medlock et al. published a systematic review that included 41 studies that evaluated mortality prediction models applicable to extremely preterm infants at one or more of these time points.8

The aim of the study was to carry out a systematic review of prediction models of mortality in premature infants developed and validated in the last 10 years.

The first search identified 7744 references and the second 1435 references. After reading the titles and abstracts, the reviewers selected 31 articles published after May 2010 to be read in full, and subsequently excluded 23. We finally reviewed the 8 new prediction models of mortality for preterm infants published since 201010–17 (Fig. 1).

Figure 1.

Flowchart of the selection of articles out of the total identified in the literature search.

(0.15MB).

Table 1 presents the study populations, times of model application and outcomes in the 8 new mortality prediction models. Of this total, 5 (62.5%) were developed in North America (3 in Canada), 2 (25%) in Europe (Netherlands and Italy) and 1 was an international multicentre project. All included preterm newborns of different BWs and GAs exclusively except for the model developed in Mexico,14 which can also be applied to full-term newborns. The sample size varied widely, and the studies that included the greatest number of preterm infants were 1 conducted in Italy (Podda et al., n=23,747),17 1 in Canada (Shah et al., n=17,048)10 and 1 in the USA (Ambalavanan et al., n=13,085).11 When it came to the time point that the model was applied, 2 (25%) were prenatal models,13,16 6 (62.5%) were applied at birth or in the first day of life10–12,14,17 and 2 (25%) involved assessments at different times during the stay in the NICU.11,15 Of these new models, we ought to highlight the sequential model of Ambalavanan et al.,11 which makes it possible to adjust the prognosis depending on the clinical trajectory of the newborn during the stay in the neonatal unit by making predictions at 7 and 28 days post birth and at 36 weeks of postmenstrual age. The outcome variable was different in every study. Three studies (37.5%) considered mortality alone at different times during the stay,13,14,17 and the rest combined mortality with survival with different degrees of morbidity during hospitalisation or after discharge.10,15,16

Table 2 summarises the statistical characteristics of the 8 new prediction models. We found that 87.5% used different multiple logistic regression approaches and only the most recent study17 used a neural network approach. The methods used to measure discrimination were the OR (75%) and the AUC (50%). The Hosmer–Lemeshow test was used as a measure of calibration in 3 studies10,12,13 (37.5%), of which 2 also used the Brier score to assess accuracy12,13 (25%). The input (independent) variables in the different models varied widely both in nature and number (Table 2).

The validation of the models consisted of internal validation in 5 of the studies (62.5%).11–14,17 Two models (25%) were validated in specific cohorts different from the cohorts used in the development phase,14,17 one by analysing a subset of 30% of the total sample11 (12.5%) and another 2 (25%) using a bootstrap approach in small subsets.12,13 Three models have online tools.11,16,17

The other 23 studies initially selected after reading the full text that were ultimately excluded (Table 3), 12 external validation studies of previously published models, 7 studies that analysed different outcomes associated with a poor prognosis, 1 study on the use of the Score for Neonatal Acute Physiology (SNAP) to predict mortality or bronchopulmonary dysplasia (BPD), 1 to predict the length of stay, 1 that compared outcomes in different cohorts of preterm newborns and 1 on the use of the Respiratory Severity Score to predict mortality or BPD.

Predictive methods are based on epidemiological population data on the selected outcomes, while health professionals and families need specific information on the individual outcomes of a particular child.28 Thus, the applicability of the model may be limited by the fact that the model may be based on data that could be obsolete by the time the model is validated. When we analysed models published after the review by Medlock et al.,8 we found that some are based on population-based data for infants born preterm up to 15 years before publication of the model.11,13,15,16 Furthermore, some of these classical prediction models, such as the Clinical Risk Index for Babies (CRIB) score19 or the model published by Draper et al.20 were developed using data for preterm newborns collected before the use of surfactant became widespread, which may compromise their current applicability.

Even in recent models, it is important to take into account that models developed in one population may not be applicable to a different population. This is why models such as the simplified age-weight (SAW) or the simplified age-weight-sex (SAWS) classifications,21 designed for developing countries where it would not make sense to apply other types of variables, are important. Thus, it is important that a prediction model undergoes not only internal validation, but also external validation in other populations if it is expected to be applied to populations other than the source population. Of all the prognostic models developed to date for preterm infants, less than half have undergone internal validation and only 4 have undergone external validation in more than 1 study (National Institute of Child Health and Human Development [NICHD], CRIB, SNAP II and Neonatal Acute Physiology Perinatal Extension [SNAPPE] II).8 Our analysis of the 8 models published most recently revealed that half had undergone internal validation,12,17 and to our knowledge none have undergone external validation, which limits their applicability in other geographical areas.

In Spain, there are currently no prediction models for preterm infants developed with data collected in the Spanish population, and the instruments most widely used in clinical practice are the CRIB score18 and the NICHD model published by Tyson et al.22

To facilitate the widespread use of a model, it is helpful to make online instruments available that can be easily accessed and interpreted by health care professionals, as has been done for some of the models.11,16,17,22 Another aspect that can be useful would be to be able to adjust the data used in the model to the specific characteristics of a single child. In this regard, we ought to highlight the effort made in the 2012 NICHD model11 to adjust the prognosis based on different events that may take place during the hospital stay of the preterm infant by offering sequential prediction models applied at birth, 7 and 28 days post birth and 36 weeks of postmenstrual age. Most existing models make the prediction at a given time point, predominantly at birth or in the first day of hospitalisation,10,12,14,16,17 which limits their usefulness.

Another important aspect is the outcome or dependent variable that is evaluated. Some of the outcomes investigated in recent studies, in addition to mortality or survival at different time points (from 7 days post birth to hospital discharge), include morbidity during the hospital stay12,16 or neurodevelopmental abnormalities at age 18–22 months.11

In nearly all of the prediction models analysed in this review, the statistical methods used during the development phase were based on multivariate logistic regression,10–16 with the study by Podda et al.17 being the exception that used a neural network approach. An equal number of studies performed internal validation by means of specific cohorts14,17 and bootstrap samples.12,13 Validation in different cohorts offers the advantage of allowing inclusion of a greater number of patients and determining whether the model continues to be useful as time passes.

Whenever prediction models are used, it is important to take into account the self-fulfilling prophecy phenomenon.11 By giving professional access to these models, they may perpetuate the values of the model itself, precluding outcomes different from those already validated in the model. For example, a hospital with a proportion of survival of 10% in preterm infants born at 23 weeks’ gestation is likely to perpetuate this proportion if families and providers make decisions based on this figure in subsequent years. It is known that when the providers that manage a newborn are pessimistic about the outcome, outcomes tend to be worse.23 In light of this, we need to periodically review prediction models and adapt them to observed changes in survival through the years, as has been done by the NICH11,22 and the Canadian Neonatal Network.10,12

It is important to assess the quality of any new prediction model. With this purpose, the Prognosis Methods Group of the Cochrane Collaboration developed the Checklist for Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modelling Studies (CHARMS) with a list of elements to be used for critical evaluation of each predictive model,24 including aspects like the source of data, participant description, predicted outcomes, sample size, handling of missing data, model development, model performance, applicability and risk of bias. Recently, a different group published the Prediction model Risk Of Bias Assessment Tool (PROBAST),25 developed to assess risk of bias and applicability of models in studies that develop, validate or update prediction models by means of 20 items organised into 4 domains (participants, predictors, outcome and analysis).

At present, there are no studies in the impact of the use of prediction models in decision-making in preterm infants,8 and also no studies analysing which outcomes families would consider most important for the purpose of decision-making or how they would like to have the available evidence presented to them.26

Table 4 presents the main prediction models currently used at the international level,11,17,19–22,27–31 some of which have been developed in the same population over periods of more than 2 decades,19,20 and most of them not externally validated in a different population.8

In Spain, the SEN 1500 Network has been collecting data on infants born before 30 weeks’ gestation or with a BW of 1500g or lower since 2002.32,33 The objectives of this Network are to assess and compare the quality of neonatal care of very low birth weight infants born in Spain and to provide data to professionals for comparison with other populations34 and to develop specific improvement measures in their care settings.35–38 The data obtained by the network is clearly an ideal source for the development of mortality and morbidity prediction models in this population and specifically in Spain.

One of the strengths of the study is that we performed 2 complementary literature searches of studies that included preterm infants and death or survival as an endpoint. One of the searches was based on predominantly clinical criteria and the other on statistical methods. One possible limitation is that both searches focused exclusively on the PubMed database.

In conclusion, decision-making based on prediction models allows a higher degree of individualisation of the care delivered to a child and optimises the use of available resources. It is important to be aware of the risks involved in using these models. At present, there are no externally validated models for predicting outcomes of preterm infants in Spain, nor models developed and validated with Spanish data. The SEN 1500 Network provides an opportunity to develop mortality prediction models for Spanish preterm infants at different time points.

The authors have no conflicts of interest to declare.