Neonatal death prediction scores: a systematic review and meta-analysis

Abstract

Objective

To compare, through a systematic review and meta-analysis of observational accuracy studies, the main existing neonatal death prediction scores.

Method

Systematic review and meta-analysis of observational accuracy studies. The databases accessed were MEDLINE, ELSEVIER, LILACS, SciELO, OpenGrey, Open Access Thesis and Dissertations, EMBASE, Web of Science, SCOPUS and Cochrane Library. For qualitative analysis, Quality Assessment of Diagnostic Accuracy Studies 2 was used. For the quantitative analysis, the area under the curve and the SE were used, as well as the inverse of the variance as a weight measure, DerSimonian and Laird as a measure of random effects, Higgins’ I² as an estimate of heterogeneity, Z as a final measure with a 95% confidence level.

Results

55 studies were analysed, 8 scores were compared in a total of 193 849 newborns included. The most accurate neonatal death prediction score was Score for Neonatal Acute Physiology Perinatal Extension II (SNAPPE II) (0.89 (95% CI 0.86 to 0.92)) and the least accurate was gestational age (0.75 (95% CI 0.71 to 0.79)).

Conclusion

SNAPPE II was the most accurate score found in this study. Despite this, the choice of score depends on the situation and setting in which the newborn is inserted, and it is up to the researcher to analyse and decide which one to use based on practicality and the possibility of local implementation. Given this, it is interesting to carry out new prospective studies to improve the prediction of neonatal deaths around the world.

PROSPERO registration number

CRD42023462425.

WHAT IS ALREADY KNOWN ON THIS TOPIC

• There are eight main neonatal death prediction scores. There are several studies showing the advantages and disadvantages of each one, as well as comparisons between scores, but using restricted groups of newborns. As neonatal death does not only affect premature or low birth weight newborns, a comparison of the main scores in a broader population is needed.

WHAT THIS STUDY ADDS

• The most accurate score was the Score for Neonatal Acute Physiology Perinatal Extension II, but it has some difficult criteria to follow in many settings.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• Identifying the most accurate score can help health professionals, as well as institutions, to reduce the high numbers of neonatal deaths.

Introduction

Neonatal death corresponds to the death of a newborn between birth and the 28th day of life.^{1 2} Although it is a short period of time, it contains most deaths in childhood, with the highest risk of death occurring in the first week of the newborn’s life.³ The outcome of this period, early neonatal death, represents approximately 73% of all postnatal deaths recorded worldwide.⁴ This is why early neonatal death is a public health problem that is still far from being resolved.⁵

The main reason for neonatal death is the low birth weight (BW), followed by prematurity, inadequate prenatal care, the presence of congenital anomalies, the presence of neonatal asphyxia and the type of delivery.^{5 6} Given the various causes and rising numbers, there was a need to create systems to identify early neonatal deaths, such as the Clinical Risk Index for Babies (CRIB), Score for Neonatal Acute Physiology (SNAP) and Score for Neonatal Acute Physiology Perinatal Extension (SNAPPE).⁷

Despite the existence of a systematic review and meta-analysis comparing the main scores in premature babies,⁸ there is a need to expand the population studied, not being restricted only to premature babies or low BW newborns, to better understand the global impact of different neonatal death prediction scores in all situations. Therefore, this study aims, through a systematic review and meta-analysis of observational accuracy studies, to compare the main neonatal death prediction scores.

Methods

The PROSPERO protocol (https://www.crd.york.ac.uk/prospero/) has been published under number CRD42023462425.

Study selection, data extraction, risk of bias analysis, data analysis and quality of evidence analysis were carried out by two authors independently. Disagreements were resolved through an agreement between the parties. If there was no consensus, a third author decided the impasse.

Eligibility criteria

The eligibility criteria were made up of the following items: observational studies on neonatal death, studies that involved the main prediction scores (BW, CRIB, CRIB II, gestational age (GA), SNAP, SNAP II, SNAPPE), studies that used the area under the curve (AUC) and their respective CIs (95% CI) and/or SE, studies that had neonatal death as the main outcome. There were no restrictions on language, period or location.

Databases

The databases were MEDLINE via PubMed, ELSEVIER via ScienceDirect, LILACS via BVS, SciELO, OpenGrey, Open Access Thesis and Dissertations, EMBASE, Web of Science, SCOPUS and Cochrane Library.

The searches were carried out between November 2023 and February 2024. In June 2024, before submission, the searches were repeated to ensure that all possible studies were included.

Search strategy

The search strategy was defined based on the PECOS structure (Population, Exposure, Comparison, Outcome, Type of study). With this, a search strategy was constructed using MeSH terms (https://www.ncbi.nlm.nih.gov/mesh) and free terms.

Study selection

The selection of studies was divided into four stages: titles, abstracts, full articles and references. Furthermore, the selection of studies was complemented by seeking help from experts in the field to find studies relevant to the topic.

The removal of duplicates was carried out with the help of the Rayyan application (Qatar Foundation, Doha, Qatar) (https://rayyan.ai/).

Data extraction

Data extraction was organised in a spreadsheet using Microsoft Excel V.16.0 software (Microsoft, Redmond, USA) based on the following elements: name of authors, year of publication, prediction score used, AUC, 95% CI and SE.

Neonatal death prediction scores

The neonatal death prediction scores included were BW, CRIB, CRIB II, GA, SNAP, SNAP II, SNAPPE and SNAPPE II.

Risk of bias analysis

The risk of bias analysis was performed using the Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS-2) tool to evaluate accuracy studies. QUADAS-2 analyses four domains (patient selection, index test, reference standard, flow and time) and is applied in four phases: summary of the research question, adaptation of the tool and production of specific analyses, construction of a flow chart and judgement of bias and applicability.⁹

Data analysis

Data analysis was performed using RStudio software version 2023.12.1+402 (Posit Software, Boston, USA). The data analysed were AUC and SE. For studies that did not have SE, but rather 95% CI, the conversion form (SE=(upper limit of the CI−lower limit of the CI)/3.92)) was used.

Using the Meta package and Metagen documentation, the pooled AUC and its CI were found. Furthermore, the inverse of the variance was used as a weight measure, DerSimonian and Laird as a measure of random effects, Higgins’ I² as an estimate of heterogeneity and Z as a final measure. A confidence level of 95% was considered.

Heterogeneity was explored when Higgins’ I² was greater than 50%.¹⁰ The analysis was performed in subgroups. The evaluation of publication bias was carried out using the funnel scatterplot, in addition to the Egger test, only when the analysis had ten or more studies. Finally, if there was an asymmetry in the funnel scatterplot, the Trim-and-Fill method was used to nullify the asymmetry found.

Quality of evidence analysis

The quality of evidence analysis was carried out using the Grading of Recommendations, Assessment, Development and Evaluations, a tool composed of four levels of evidence to recommend a clinical decision.¹¹

Patient and public involvement

No patient is involved.

Results

Study selection

The database search identified 3087 articles, of which 346 were duplicates. After the eligibility phase, 2686 studies were excluded, leaving 55 articles¹²,⁶⁵ to be included in our study (figure 1).

Figure 1. Study selection flow chart. AUC, area under the curve.

The characteristics of the 55 included studies are available in table 1 and those relevant to the quantitative analysis are shown in table 2.

Table 1. General characteristics of the included studies.

Author	Year	Country	Type of study	Number of neonates analysed	Score
Alshafei et al12	2024	United Arab Emirates	Retrospective cohort	404	CRIB II, BW, GA
Bastos13	1997	Portugal	Retrospective cohort	186	CRIB, SNAP, SNAPPE, BW
Bayen et al14	2023	India	Prospective cohort	57	CRIB II, SNAPPE II
Beltempo et al	2018	Canada	Retrospective cohort	9240	SNAP II
Bhandekar et al16	2024	India	Prospective cohort	44	CRIB II
Brito et al17	2003	Brazil	Prospective cohort	284	CRIB, BW, GA
Bührer et al18	2008	Germany	Retrospective cohort	1358	CRIB, CRIB II, BW, GA
Bührer et al19	2000	Germany	Retrospective cohort	430	CRIB, BW, GA
Eriksson et al20	2002	Sweden	Retrospective cohort	240	CRIB, SNAP, SNAPPE, BW, GA
Escobar et al21	1995	USA	Prospective cohort	1251	SNAPPE, BW
Ezz-Ezdin et al22	2015	Egypt	Prospective cohort	113	CRIB II, BW, GA
De Felice et al23	2005	Italy	Retrospective cohort	147	CRIB, CRIB II, BW, GA
Fontenele et al24	2020	Brazil	Prospective cohort	247	SNAPPE II
Gagliardi et al25	2004	Italy	Prospective cohort	720	CRIB, CRIB II, SNAPPE II
Gooden et al26	2014	Jamaica	Retrospective cohort	109	CRIB II, BW, GA
Guenther et al27	2015	Germany	Retrospective cohort	5340	CRIB, BW, GA
Guzmán Cabañas et al28	2009	Spain	Prospective cohort	10 608	CRIB, BW, GA
Harsha nd Archana29	2015	Bangalore	Prospective cohort	248	SNAPPE II
Heidarzadeh30	2016	Iran	Prospective cohort	215	CRIB II
International Neonatal Network31	1993	UK	Retrospective cohort	735	CRIB, BW
Jasic32	2016	Croatia	Prospective cohort	153	CRIB II, BW, GA
Lago et al33	1999	Italy	Prospective cohort	81	CRIB, BW, GA
Lee et al	2019	South Korea	Prospective cohort	636	CRIB II
Lokraj35	2023	Nepal	Prospective cohort	260	SNAP II, SNAPPE II
Medvedev et al7	2020	UK and Gambia	Retrospective cohort	110 726	CRIB II
Menéndez36	2018	Ecuador	Retrospective cohort	227	CRIB, CRIB II, SNAP II, SNAPPE II
Mesquita37	2011	Paraguay	Retrospective cohort	288	SNAP II, SNAPPE II
Moreira et al38	2022	Sweden	Prospective cohort	3752	CRIB II, BW, GA
Muktan et al39	2019	Nepal	Prospective cohort	255	SNAPPE II
Muñoz-Garcia et al40	2014	Spain	Retrospective cohort	81	CRIB II, SNAP II, SNAPPE II
Park et al41	2018	South Korea	Retrospective cohort	564	CRIB II, BW, GA
Patra and Karmakar42	2019	India	Prospective cohort	143	CRIB II
Phillips et al	2010	UK	Retrospective cohort	408	CRIB
Radfar et al44	2018	Iran	Prospective cohort	191	SNAP II, SNAPPE II
Rastogi et al45	2010	India	Prospective cohort	86	CRIB II
Rautonen et al46	1994	Finland	Prospective cohort	222	CRIB, SNAP, SNAPPE, GA
Ray et al47	2019	India	Prospective cohort	961	SNAPPE II
Reid et al	2014	Australia	Prospective cohort	1607	CRIB II, SNAPPE II
Richardson et al49	1993	USA	Retrospective cohort	1621	SNAP, SNAPPE, BW
Richardson et al50	2001	Canada	Prospective cohort	14 610	SNAPPE II, BW
Rosenberg et al51	2008	Bangladesh, Egypt and Nepal	Retrospective cohort	467	CRIB II
Ruiz et al52	2007	Spain	Prospective cohort	163	CRIB, BW, GA
Sarquis et al53	2002	Brazil	Prospective cohort	100	CRIB, BW, GA
Shrestha et al54	2017	Nepal	Prospective cohort	126	SNAP II, SNAPPE II
Silveira et al55	2001	Brazil	Retrospective cohort	553	SNAP, SNAPPE
Singh56	2018	Nepal	Prospective cohort	255	SNAPPE II
Sotodate et al	2020	Japan	Retrospective cohort	171	CRIB II, SNAP II, SNAPPE II, BW, GA
Thimoty et al58	2009	Indonesia	Prospective cohort	40	SNAPPE II
Tyagi et al59	2022	India	Prospective cohort	100	SNAPPE II
Vardhelli et al60	2022	India	Prospective cohort	419	CRIB II, SNAPPE II
Vardhelli et al	2023	India	Prospective cohort	669	SNAPPE II
Vasudevan et al62	2006	India	Retrospective cohort	97	SNAP
Weirich et al63	2005	Brazil	Retrospective cohort	20 286	BW
Yang et al64	2002	China	Prospective cohort	192	CRIB, CRIB II, SNAP II, SNAPPE II
Zhang et al65	2023	China	Prospective cohort	1363	CRIB II

BWbirth weightCRIB IIClinical Risk Index for Babies IIGAgestational ageSNAPScore for Neonatal Acute PhysiologySNAPPE IIScore for Neonatal Acute Physiology Perinatal Extension II

Table 2. Characteristics relevant to the quantitative analysis of the included studies.

Scale	Studies included	Pooled AUC	SE
BW	22	0.76	0.015
CRIB	17	0.87	0.010
CRIB II	25	0.82	0.020
GA	17	0.75	0.020
SNAP	6	0.82	0.020
SNAP II	9	0.83	0.035
SNAPPE	6	0.88	0.025
SNAPPE II	21	0.89	0.015

AUCarea under the curveBWbirth weightCRIB IIClinical Risk Index for Babies IIGAgestational ageSNAP IIScore for Neonatal Acute Physiology IISNAPPE IIScore for Neonatal Acute Physiology Perinatal Extension II

Risk of bias analysis

The studies were analysed using the QUADAS-2 tool. In the ‘patient selection’ domain, 14 studies^{12 13 15 18 20 23 31 36 43 49 55 57 62 63} were classified as having a high risk of bias. In the ‘index test’ and ‘flow and time’ domains, 28^{1213 15 19},^{21 23 7 26 27 31 33 34 36 37 40 41 43 4647 51 55},^{57 62 63} and 14 studies718 19 31 34 37 38 41 45,^{47 49} had the same classification, respectively. Finally, in the ‘reference standard’ domain, all studies were classified as low risk of bias. Likewise, in relation to applicability, all studies were classified as not worrying. The graphs relating to the risk of bias and concerns about applicability are illustrated in figure 2.

Figure 2. Risk of bias analysis of the included studies. QUADAS-2, Quality Assessment of Diagnostic Accuracy Studies 2.

Data analysis

Figure 3 illustrates the relationship between neonatal death prediction scores represented by pooled AUC and their 95% CI. Below, separately, the results of the neonatal death prediction scores analysed. The figures for each of the scores analysed are available in online supplemental material.

Figure 3. Summary of neonatal death prediction scores. AUC, area under the curve; BW, birth weight; CRIB II, Clinical Risk Index for Babies II; GA, gestational age; SNAP II, Score for Neonatal Acute Physiology II; SNAPPE II, Score for Neonatal Acute Physiology Perinatal Extension II.

Birth weight

BW was highly accurate in predicting neonatal death (pooled AUC=0.76; 95% CI (0.73 to 0.79); p=0.000; I² = 93% (p<0.01)).^{1213 18},^{23 26} When exploring heterogeneity, it was noticed that heterogeneity was maintained in four of the subgroups analysed (Germany^{18 19 27}: I²=77% (p=0.01); Italy^{23 33}: I²=92% (p<0.01); South Korea⁴¹: I²=90% (p<0.01); Canada^{21 50}: I²=99% (p<0.01)).

Regarding publication bias, there was symmetry in the funnel graph (Egger’s test: t=−1.53; p=0.13).

Clinical Risk Index for Babies

CRIB showed high accuracy in predicting neonatal death (pooled AUC=0.87; 95% CI (0.85 to 0.89); p=0.00; I²=87% (p<0.01)).¹³¹⁷,20 23 25 27 28 31 33 36 43 46 52 53 64 When exploring heterogeneity, it was noticed that heterogeneity was maintained in two of the subgroups analysed (Germany^{18 19 27}: I²=88% (p<0.01) (Spain)^{13 28 52}: I²=73% (p=0.03)).

Regarding publication bias, there was symmetry in the funnel plot (Egger’s test: t=0.52; p=0.60).

Clinical Risk Index for Babies II

CRIB II showed high accuracy in predicting neonatal death (pooled AUC=0.82; 95% CI (0.78 to 0.86); p<0.01; I²=99% (p=0.00)).^{7 12 14 16 18 20 23 25 26 30 32 34 36 38 40},4345 48 51 57 60 64 65 When exploring heterogeneity, it was noticed that heterogeneity was maintained in three of the subgroups analysed (India^{14 16 42 45 60}: I²=93% (p<0.01); South Korea^{34 41}: I²=98% (p<0.01); China^{64 65}: I²=96% (p<0.01)).

Regarding publication bias, there was asymmetry in the funnel graph (Egger’s test: t=−4.46; p=0.001). To adjust the graph, it was necessary to impute 15 studies to eliminate the asymmetry (Egger’s test: t=−0.61; p=0.54).

Gestational age

GA showed good accuracy in predicting neonatal death (pooled AUC=0.75; 95% CI (0.71 to 0.79); p<0.01; I²=94% (p<0.01)).^{1217 18 26},28 32 33 38 41 46 52 53 When exploring heterogeneity, it was noticed that there was a decrease in heterogeneity in two of the subgroups analysed (Germany^{18 19 27}: I²=92% (p<0.01); Spain^{28 52}: I²=75% (p=0.05); South Korea⁴¹: I²=78% (p=0.03)).

Regarding publication bias, there was symmetry in the funnel graph (Egger’s test: t=−1.53; p=0.14).

Score for Neonatal Acute Physiology

SNAP showed high accuracy in predicting neonatal death (pooled AUC=0.82; 95% CI (0.78 to 0.86); p=0.00; I²=44% (p=0.11)).^{13 20 46 49 55 62}

Score for Neonatal Acute Physiology II

SNAP II showed good accuracy in predicting neonatal death (pooled AUC=0.83; 95% CI (0.76 to 0.90); p<0.01; I²=99% (p<0.01)).¹⁵³⁵,37 40 44 54 57 64 When exploring heterogeneity, it was noticed that heterogeneity was maintained in one of the subgroups analysed (Nepal^{35 54}: I^²=80% (p=0.03)).

Score for Neonatal Acute Physiology Perinatal Extension

SNAPPE showed high accuracy in predicting neonatal death (pooled AUC=0.88; 95% CI (0.83 to 0.93); p<0.01; I^²=87% (p<0.01)).^{13 20 21 46 49 55} It was not possible to analyse heterogeneity due to the unitary nature of each subgroup analysed.

Score for Neonatal Acute Physiology Perinatal Extension II

SNAPPE II showed high accuracy in predicting neonatal death (Pooled AUC=0.89; 95% CI (0.86 to 0.92); p=0.00; I^²=94% (p<0.01)).1424 25 29 35,37 39 40 44 47 48 50 54 56 When exploring heterogeneity, it was noticed that heterogeneity was maintained in two of the subgroups analysed (India^{1447 59},⁶¹: I^²=84% (p<0.01); Nepal^{35 39 54 56}: I²=75% (p<0.01)).

Regarding publication bias, there was asymmetry in the funnel plot (Egger’s test: t=−3.01; p=0.006). To adjust the graph, it was necessary to impute nine studies to eliminate the asymmetry (Egger’s test: t=−0.37; p=0.71).

Table 3 illustrates the compilation of results obtained in the qualitative and quantitative analysis.

Table 3. Summary of results obtained in the qualitative and quantitative analysis.

Scores	Studies included	Number of participants analysed	Pooled AUC	95% CI	I²	Publication bias
BW	22	62 706	0.76	(0.73 to 0.79)	93%	p=0.13
CRIB	17	21 441	0.87	(0.85 to 0.89)	87%	p=0.60
CRIB II	25	123 754	0.82	(0.78 to 0.86)	99%	p=0.001
GA	17	24 239	0.75	(0.71 to 0.79)	94%	p=0.14
SNAP	6	2919	0.82	(0.78 to 0.86)	44%	–
SNAP II	9	10 776	0.83	(0.76 to 0.90)	99%	–
SNAPPE	6	4073	0.88	(0.83 to 0.89)	87%	–
SNAPPE II	21	21 724	0.89	(0.86 to 0.92)	94%	p=0.006

AUCarea under the curveBWbirth weightCRIBEClinical Risk Index for BabiesGAgestational ageSNAPScore for Neonatal Acute PhysiologySNAPPE IIScore for Neonatal Acute Physiology Perinatal Extension II

Quality of evidence analysis

The studies were analysed using the Grading of Recommendations, Assessment, Development and Evaluations (GRADE) tool. Only one of the five domains was identified as high risk. Therefore, the certainty of the study’s evidence was classified as moderate. The graphs relating to the risk of bias and concerns about applicability are illustrated in table 4.

Table 4. Quality of evidence analysis.

GRADE domains	Judgement
Risk of bias	Low
Imprecision	Low
Inconsistency	High
Indirectness	Low
Publication bias	Low
Overall	Moderate

GRADEGrading of Recommendations, Assessment, Development and Evaluations

Discussion

The study showed that the most accurate neonatal death prediction score was SNAPPE II (0.89 (95% CI 0.86 to 0.92)),1424 25 29 35,37 39 40 44 47 48 50 54 56 and GA was the score with the lowest accuracy (0.75 (95% CI 0.71 to 0.79)).¹²¹⁷,20 22 23 26 There were 55 studies analysed, 8 scores being compared in a total of 193 849 newborns included.

As the most up to date of the main scores, SNAPPE II would tend to be the most accurate. Built on the positive and negative points of its predecessors, SNAPPE II is an extension of SNAP II.⁶⁶ In addition to measurements of blood pressure, temperature, PO2/FiO2 ratio, serum pH, amount of fainting and urinary output, there are three data points that aggregate SNAPPE II: BW≤749 g, Apgar<7 in the 5th minute and small for GA.⁵⁰ It is because of these refined data that SNAPPE II was probably the most accurate score in this study.

In our study, SNAPPE (0.88 (95% CI 0.83 to 0.93))^{13 20 21 46 49 55} and CRIB (0.87 (95% CI 0.85 to 0.89))¹³¹⁷,20 23 25 27 28 31 33 36 ^{43 46 52 53 64} were the second and third most accurate scores. This fact was different from the recently published systematic review and meta-analysis that compared the main scores in premature infants. In the study in question, CRIB was the most accurate score (0.98), followed by SNAPPE (0.71), the reason being that it is a simple-to-apply score.⁸ In fact, the CRIB has only six parameters to be collected in a hospital environment in the first 12 hours of a newborn’s life.^{31 66} SNAPPE, an extension of SNAP, has 31 parameters to be collected, becoming a barrier to effective collection.⁴⁹ This difficulty, years later, led to the emergence of the SNAP II and SNAPPE II updates, the most accurate score in our study.

BW (0.76 (95% CI 0.73 to 0.79))^{1213 17},^{23 26 50 52 53 63} was expected e o GA (0.75 (95% CI 0.71 to 0.79))¹²¹⁷,20 22 23 26 ^{38 41 46 52 53} were the worst scores, as they only analysed one parameter in isolation. Furthermore, such parameters are present in all other scores.¹² The inclusion of BW and GA was necessary to provide a base standard for the other scores in our study.

Given the existence of so many neonatal death prediction scores for the most diverse situations,^{8 66} choosing the best score becomes a herculean task. However, the guiding factor for this choice is precisely the situation in which the newborn is inserted.⁶⁶ Although SNAPPE II is the most accurate score in this study, if the newborn is in a hospital environment where the parameters necessary to predict death are only sufficient to compose the CRIB,³¹ the researcher can use this score without prejudice. Likewise, if it is necessary to collect data for more than 12 hours, as requested by the CRIB, the researcher can use SNAP,⁴⁹ for example, which also has high accuracy.

Despite the complexity of these scores, they are widely used in Neonatal Intensive Care Units (NICUs), mainly due to their degree of accuracy.⁶⁶ With data collection starting at birth and ending in the first 12 or 24 hours, the data to be collected involves components of prenatal, delivery and postpartum care, more precisely, the hospital environment, especially the NICU environment.⁶⁶ The elements collected from blood gas analysis and ventilation parameters make these scores almost impractical in places where access to healthcare is difficult.⁷

The closest solution would be a less complex score, with single or combined elements, such as GA and BW, however, as seen in this study, the accuracy is low, and it is not the best score to predict death. However, in the absence of access to quality healthcare, these scores become an alternative in the attempt to reduce neonatal deaths.

Heterogeneity was observed in 7 of the 8 scores analysed. The score with the greatest heterogeneity was CRIB II. When exploring it, it was seen that the countries with the greatest contribution were India, South Korea and China. Regarding the Indian studies,^{14 16 42 45 60} despite having similar eligibility criteria, there was a sample variation between the studies, which is the possible reason for the heterogeneity between these studies. As for the Korean studies,^{34 41} the possible factor of heterogeneity was the methodological difference of the studies, one being prospective and the other retrospective. Finally, the Chinese studies^{64 65} differ both in the eligibility criteria, one being more restrictive than the other, and in the size of the sample collected.

Regarding the lowest and highest accuracy scores, respectively, GA and SNAPPE II, both had the same heterogeneity value (I^²=94%). The first was influenced by German studies and the latter by Indian studies. The German studies^{18 19 27} were prospective, with similar eligibility criteria, but the collection time differed between them, with 5 years being the shortest collection time and 15 years being the longest. In the Indian studies,^{1447 59},⁶¹ not all were multicentre, despite the prospective nature of all of them. In addition, the collection time varied from 1 to 2 years.

In general, one of the main reasons for this difference between the studies, in addition to the characteristics of each country, was the longitudinal design of the studies, since many were prospective, and others were retrospective. Another point to be highlighted was the publication bias found in 2 of the 8 scores. An interesting detail to be highlighted was that these two scores are the most current (CRIB II and SNAPPE II). Despite the facts mentioned above, the heterogeneity of our study does not invalidate the individual clinical results of the included studies, since, individually, the studies maintained the expected methodological accuracy, given the complexity of the main scores.

In the analysis of the risk of bias, high-risk studies were found in three of the four domains. The main reasons for this occurrence were the retrospective nature of the studies, the knowledge of the reference standard, which was neonatal death, and the absence of all participants in the final analysis. Despite the finding of risk in the individual analysis, when using GRADE, we classified the degree of evidence of the study as moderate since heterogeneity was found in most of the scores analysed in the meta-analysis.

There are some limitations in our study. The first, inherent to review studies, arises from the fact that secondary data were collected to compose our qualitative and quantitative analysis. Another limitation is the use of retrospective cohorts which, despite being good accuracy studies, cannot completely nullify the fact of knowledge of the participant’s outcome. Finally, when using the AUC instead of data referring to true positives and negatives and false positives and negatives, there is a risk of overvaluation or undervaluation of the analysed data, but the largest evaluated sample reduces these bias’s chances to be significant.

Given the data presented and the limitations outlined, steps should be taken to verify and apply the data, as well as to eliminate the limitations presented. Prospective studies using the main available scores are interesting for analysing the impact of these scores on the same population and even on different populations. Furthermore, selecting the most appropriate scores, aiming to create a new score for the most diverse economic and social situations, to ensure simple and practical application, is an innovative and bold proposal that deserves the attention of both the academic community and public authorities, highlighting the issue of neonatal death as an important public health problem to be solved.

Conclusion

SNAPPE II was the most accurate score found in this study. Despite this, the choice of score depends on the newborn’s situation, and it is up to the researcher to analyse and decide which one to use. In view of this, it is interesting to conduct prospective studies, using the main available scores, in the same population and even in different populations, aiming to find the most appropriate score, as well as the creation of a new simple, practical and comprehensive score to improve the prediction of neonatal deaths around the world.

Data availability statement

All data relevant to the study are included in the article or uploaded as supplementary information.