Neonatal mortality (NM) refers to the death of a neonate within the first 4 weeks of life or during the neonatal period and is expressed per 1,000 neonates (1). It is classified into two categories: early neonatal mortality (ENM), which refers to the death of neonates within the first 7 days of life, and late neonatal mortality(LNM), which refers to the death of neonates between 7 and 28 days of life (2).

The World Health Organization (WHO) reported 2.6 million neonatal deaths in 2016, accounting for nearly 46% of all under-five deaths worldwide. Mortality remains the most direct and significant predictor of population-level health; however, 62% of global deaths are unreported, mainly in Africa. With numerous intervention initiatives and policies, the WHO and other stakeholders have worked to minimize maternal and infant mortality (3). Every day, approximately 6,700 neonatal deaths occur, with nearly a third of all neonatal deaths happening on the first day after birth and nearly three-quarters occurring in the first week of life (4, 5).

Globally, NMR was decreased by 51% between 1990 and 2017. However, it remains highest in West and Central Africa and South Asia (6). The annual neonatal mortality rate in the regions (sub-Saharan Africa and South Asia) is more than nine times higher than the average NMR in high-income countries. South Asia and sub-Saharan Africa together accounted for 79% of the overall neonatal death burden, and South Asia alone accounted for 38% of neonatal deaths, and approximately 18% of neonatal deaths occurred in East and South Africa (6). According to the WHO estimates, a significant proportion of all under-five deaths occur in the neonatal or perinatal period (7). Ethiopia is among the six countries which account for half of the global under-five mortality (8–10). It 's one of the top 10 countries with the highest level of worldwide neonatal mortality figures, with an estimated 122,000 newborn deaths per year, which accounts for two-thirds of global neonatal deaths (8–10).

Findings from various studies revealed that prematurity is the major cause of death during the early neonatal period, while asphyxia is the major cause of death during the late neonatal period (11, 12). Additionally, respiratory distress (RD) syndrome is a common cause of neonatal moratlity (11, 13). Other contributing factors associated with both ENM and LNM include birth injury, hyperthermia, neonatal infections, and birth asphyxia (14–16).

Other factors associated with NM include maternal age, parity, pregnancy interval, multiple pregnancies, maternal health problems during pregnancy, malpresentation, problems during delivery, and infant sex (17, 18).

Efforts have been made globally and nationally to reduce this devastating public health problem, which includes Millennium Development Goals (MDGs) later extended to Sustainable Development Goals (SDGs) (19). Ending preventable child deaths is a critical part of the Global Strategy for Women’s, Children’s, and Adolescent’s Health (2016–2030) and the third Sustainable Development Goal (SDG) to ensure safe lives and encourage prosperity for all citizens of all ages (20).

Over the past two decades, the Government of Ethiopia has implemented a health sector development program aimed at improving the accessibility and quality of health services for all segments of the population. As a result, there have been significant increases in institutional deliveries, antenatal care (ANC) coverage, and postnatal care provided within 48 h of delivery (21). Ethiopia has adopted and implemented various strategies to reduce maternal and neonatal mortality, such as emergency obstetric and newborn care, adapted to enhance neonatal and maternal outcomes, resulting in substantial success in reducing under-five mortality (22, 23). However, reduction in neonatal mortality has been the least significant (24).

A clinical prediction model combines various risk factors to estimate the probability of mortality in newborns In neonatal care units. This facilitates personalized diagnostic and therapeutic decision-making within healthcare settings and enables the early identification of high-risk neonates, ensuring timely and appropriate interventions (25–27).

A risk score model derived from clinical variables can support evidence generation and facilitate timely decision-making. Moreover, up to our knowledge in the study area, a study conducted regarding NM particularly did n't consider a prediction model. Validation of such model analysis plays a great role in identifying problems and allowing accurate risk estimation. Therefore, this study aimed to identify important predictors that can forecast which infants are at high risk of neonatal death and which are not. This helps policymakers ad governmental and non-governmental organizations to take appropriate interventions.

This study focused on the development and validation of a simplified risk score to predict neonatal mortality among neonates admitted to intensive care units in Southern Ethiopia using various predictors.

The current study shows the incidence of neonatal mortality at 15.4% among neonates admitted to the NICU. This result is higher than the studies conducted in the Wolaita Sodo referral hospital (37), at Nekemte Referral Hospital (38), Tigray region (11), Bangladesh (39), and rural surveillance centers of India, Bangladesh, and Nepal (40). On the other hand, it is lower than those in studies conducted at the University of Gondar (UOG) comprehensive specialized hospital, Amhara region referral hospitals, and Debre Markos referral hospitals (11, 13, 15).

This difference might be due to hospital-to-hospital service delivery, infrastructure differences, and some of the studies conducted in community surveillance centers. This result is almost consistent with the findings of studies conducted in Southern Ethiopia (41), Nigeria, and Nepal (42, 43). This may be due to the treatment and classification of disease based on WHO guideline recommendations and both countries are underdeveloping countries.

The study demonstrated optimal performance by integrating maternal, obstetrical, and neonatal characteristics to predict mortality among neonates admitted to the intensive care units in Southern Ethiopia. The prognostic factors included residence, perinatal asphyxia, breastfeeding status, primigravida, poor 5th-minute Apgar score, amniotic fluid status, and low birth weight.

In our study, PNA had a significant association with neonatal mortality. This finding is consistent with the result of a study conducted in the Wolaita Sodo referral hospital, the University of Gondar referral hospital, and seven hospitals of the Tigray region (11, 15, 37). This may be due to oxygen deficit at delivery can lead to severe hypoxic–ischemic organ damage in newborns followed by a fatal outcome or increases the probability of death.

In this study, failure to initiate breastfeeding was associated with increased neonatal mortality among admitted neonates. This is consistent with a study conducted in Southern Ethiopia (37). The finding is also supported by a systematic review of initiation of breastfeeding and neonatal mortality (44). The possible explanations might be immunological deficits, infection susceptibility, metabolic instability, and nutritional deficiencies. These physiological disruptions collectively predispose neonates to severe illness and death.

In this research, amniotic fluid status was also associated with neonatal mortality among admitted neonates. This finding is supported by the study conducted in eastern and northwestern Ethiopia (45, 46). The possible explanation might be that stained amniotic fluid, particularly when stained with meconium, is an indicator of potential fetal distress. This condition can lead to serious neonatal complications, including meconium aspiration syndrome (MAS), respiratory distress, sepsis, and other infections. These adverse outcomes can contribute to an increased risk of neonatal mortality.

Our study demonstrated that rural residence was a significant predictor of neonatal mortality, consistent with the findings of the EDHS 2019 and the National Newborn and Child Survival Strategy of Ethiopia (21, 24). This may be due to limited access to healthcare facilities, low utilization of maternal healthcare services, socioeconomic challenges, delays in receiving specialized care, disparities in healthcare resources, weak referral systems, and inadequate health infrastructure in rural settings.

Incorporating birth weight in forecasting neonatal mortality is the highest discrimination power. In this study, birth weight is one of the predictors to result in 78% of discrimination power. This is also supported by a study conducted in Gondar (47) and Bangladesh (39). Other studies have also shown the association between birth weight and neonatal mortality (14). This might be due to physiological vulnerabilities, including immature organ development, metabolic instability, and higher susceptibility to birth complications. Additionally, in resource-limited settings, the lack of specialized neonatal care worsens these risks.

Birth weight and 5th-minute Apgar score are also among the variables that affect the performance of our model. Some studies have also reported that these factors influence the predictive performance of the model (39, 47, 48). Fifth-minute Apgar score is one of the main predictors of neonatal mortality among neonates admitted to NICU, which has also been reported by a study conducted at the University of California (49).

This may be due to low Apgar scores in neonates reflecting physiologic immaturity rather than poor condition at birth and may also be caused by maternal drugs, infections, neurological diseases, and congenital anomalies.

The AUC of the combined predicting model is 0.796 (95% CI: 0.753, 0.840), and the final reduced and internally valid model yielded an AUC of 0.781 (95% CI: 0.737, 0.825), indicating that the model has a good predictive ability. This finding is comparable to a study conducted in Bangladesh where a simplified model (birth weight, GA, lethargy, cyanosis, non-cephalic presentation, and trouble suckling) predicted death with good discrimination [validation area under the receiver operating characteristic curve (AUC) 0.80 (95% CI: 0.73, 0.87)]. However, our model showed better performance than the additional simplified model from the same study (gestational age, non-cephalic presentation, lethargy, trouble suckling) that predicted death with moderate discrimination [validation AUC 0.74, 95% CI (0.66, 0.81)] (39).

Our model’s performance is lower than that reported in a study conducted in Mexico, which achieved an AUC of 0.92. This discrepancy may be due to the socioeconomic status of the countries (48). The performance is also lower in comparison with a prognostic study conducted in Nepal (43). This is mainly due to the study focusing on neonatal factors using Score for Neonatal Acute Physiology with Perinatal Extension (SNAPPE) scoring systems.

The performance score of our model is also higher than those reported in studies conducted in low- and middle-income countries, which analyzed data from population surveillance sites in India, Nepal, and Bangladesh. In these studies, the AUC was 0.59 (95% CI 0.58, 0.61) at the start of pregnancy and 0.73 (95% CI 0.70, 0.76) at the start of delivery. However, our model’s performance score was lower compared with the AUC measured at 5 min post-partum [AUC: 0.85, 95% CI (0.80, 0.89)] (40). This is maybe due to the inclusion of participants from different countries and substantial sociodemographic differences. In addition, the study incorporates a huge sample size.

After bootstrap internal validation, optimism corrected AUC was 0.781 (0.737, 0.825). Model optimism was estimated at 0.015 indicating minimal overfitting of the model to the data. This is lower than that reported in a study conducted at the University of Gondar (47). This discrepancy may be due to differences in predictors.

In our prediction score, using a cutoff point of 3 determined by the Youden index yielded an acceptable level of specificity (66%), sensitivity (80%), PPV (0.33), NPV (0.93), LR+ (2.3), and LR− (0.3) to predict neonatal mortality among neonates admitted to the NICU. The sensitivity observed in our study was higher compared with those reported in a study conducted in Bangladesh (39) and two studies in Mexico (48, 49) (80% vs. 65.3%, 66.7%, and 40%, respectively) but nearly similar to those reported in a study conducted in Gondar which is 81%. Specificity was higher in studies conducted in Bangladesh and Mexico (39, 49). On the other hand, it is lower than those reported in studies conducted in Gondar and Mexico (47, 48). The difference may be due to different cutoff values, different prognostic factors, the quality of intensive care, and differences in socioeconomic status.

This prognostic prediction model provides a significant potential for implementation in neonatal health services to predict mortality among neonates admitted to intensive care units. It enables healthcare providers to assess mortality risk early, prioritize high-risk neonates, and implement timely interventions such as enhanced monitoring and critical treatments. The model also supports standardized evaluations and evidence-based care, reducing variability in decision-making and improving neonatal outcomes.

In resource-limited settings, the model serves as a valuable tool for guiding the efficient allocation of medical resources, ensuring they are directed toward neonates at the highest risk. Additionally, it provides training opportunities for healthcare workers, enhancing their decision-making capabilities. On a broader scale, the model's predictive data can inform public health policies and optimize resource distribution.

A key strength of the study is that our prediction model was internally validated using a bootstrapping technique by simulating a huge sample to replace itself, resulting in a small optimism coefficient. We used an adequate number of outcomes, which enabled us to construct the prediction model using an adequate number of predictors. We developed the model using risk scores which enabled health professionals to calculate risk scores easily without advanced calculations. Lastly, our prediction model was constructed from easily available and measurable predictors that make it relevant in healthcare settings.

The limitation of the study was this study was conducted on admitted neonates who were at greater risk of death. Consequently, the NM reported from this study could be overestimated. In addition, it was conducted in a teaching referral hospital. We suggest external validation of this risk prediction model before adoption in diverse clinical and public health contexts to confirm its generalizability. Moreover, since it was a retrospective study, some important variables with strong predictive ability may not have been included.

The overall incidence of neonatal mortality among neonates admitted to the NICU of HUCSH was high. We constructed a prediction model that has good performance for neonatal mortality, based on maternal characteristics and obstetrical and neonatal factors in the NICU of teaching hospitals in southern Ethiopia. The prediction score can help risk stratification during ANC, delivery, and postnatal period and in the identification of neonates at higher risk of mortality. High-risk groups can take precautions and early appropriate interventions. This is a feasible prediction score that can be readily applied by health professionals in the unit, contributing to the reduction of neonatal mortality and improvement of overall maternal and child healthcare.