A prospective institutional-based observational study design was employed in South Gondar zone public hospital, Northwest Ethiopia. Among eight public hospitals, this study was done in Addis- Zemen primary hospital, Wegeda primary hospital, Nefas Mewcha primary hospital, and Debretabor General Hospital. The study was conducted from December 2020 to May 2021.

All preterm neonates delivered and admitted in South Gondar Zone Public Hospitals were a source population. Preterm neonates who were delivered in the labor ward and admitted to NICU within 72 h of life were included in the study. Pregnant women without reliable last menstrual period or pregnant women without an early ultrasound report (first tri -minister), incomplete prenatal, delivery, and neonatal chart, and preterm neonates without parents were excluded during the data collection period.

Since there is no prior prediction study for preterm neonates’ death in a similar setting to calculate the sample size, as the rule of thumb, ten events/preterm neonate death/per covariates were considered. Thirteen covariates were recruited in the final model from a total of 36 prognostic maternal and neonatal determinants based on literature or clinical experience. From a total of 456 neonates prospectively followed, there were 132 preterm neonate deaths that occurred.

Preterm neonates delivered from December 1, 2020 to May 30, 2021, delivered and admitted in a neonatal intensive care unit (NICU) in South Gondar zone public hospitals, were included. Eligible participants came up in the labor ward and admitted neonatal intensive care unit with consecutive sampling techniques. Proportional allocation based on the monthly average number of preterm delivered and admitted has been determined for each public hospital. We followed 204 preterm neonates at Debre Tabor Hospital, 76 preterm neonates at MekaneEyesus Hospital, 112 preterm neonates at Wegeda Hospital, and 28 preterm neonates at AdisZemen Hospital for 6 months during the data collection period.

We have used a reliable last menstrual period or an early ultrasound (first trimester) to determine gestational age. Data were collected based on an adapted interview-administered questionnaire and reviewing medical records. The questionnaire was adapted by reviewing different pieces of literature, including maternal socio-demographic, maternal medical, obstetric, prenatal, and neonatal variables, including age of the mother, the residence of the mother (urban, rural), marital status (union/non-union), pre-pregnancy body mass index, mother occupation, mother education, gestational age (late, moderate, very, extremely very preterm), parity, current pregnancy complications [preeclampsia/eclampsia/pregnancy-induced hypertension, preterm premature rupture of membrane, antepartum hemorrhage (APH), cord prolapse, and Rh factor], birth interval, antenatal care (ANC) follow-up, the plurality of the child (single, multiple), pregnancy intension (wanted/unwanted), mode of delivery, presentation, magnesium given for this pregnancy, antenatal steroids given or not, sex of the neonate, birth weight, age of the neonate, any resuscitation during delivery, sepsis, perinatal asphyxia (PNA), feeding status, APGAR score, kangaroo mother care (KMC) given or not, type of preterm (spontaneous or induced), respiratory distress syndrome (RDS), hypothermia, jaundice, congenital anomaly, and hypoglycemia.

The data were collected by 12 BSc neonatal nurses and four assistant lecturers in nursing supervisors. Trained data collectors interviewed mothers who were at 28 to less than 37 weeks (36 + 6 days) gestation at the labor ward and were abstracted pregnancy, delivery, and infant data from the chart and followed during a 24-h period until discharge, transfer, death, or 42 postmenstrual ages after birth, and their outcome was recorded.

Infant death between birth and 42 postmenstrual ages was declared dead. Neonates beyond 42 postmenstrual ages, referred to, and left with medical advice, were declared alive.

Neonatal morbidity was immediate neonatal complications, including (APGAR score, RDS, hyperbilirubinemia, and sepsis of preterm infants within 42 PMAs after birth (7). Sepsis is defined as clinical sign symptoms with risk factors in lab tests (microscopic) greater than two hematologic criteria (25). “Neonatal respiratory distress was diagnosed with the presence of one or more symptoms of tachypnea, intercostal muscle retraction, grunting, nasal flaring, and cyanosis” (26). Birth asphyxia is defined as “the failure to initiate and sustain breathing at birth.”

Questionnaires were pretested with a 5% sample size at Mekane Eyesus primary hospital. The training was given for data collectors and supervisors for 1 day before and 1 day after the pretest regarding the study’s objective, data collection tool, and ways of data collection and checking the completeness of the data collection tool.

Epi-Data version 4.6 software was used to enter the data. R-programming version 4.0.5 was used to analyze the data. In the beginning, variable selection in the multivariable model was made using least absolute shrinkage and selection operator (LASSO) regression. Two-sided P-values of less than 0.05 were considered statistically significant.

Model calibration was assessed by plotting declines of the predicted probability of preterm death in each decline and fitting a smooth line. The model discrimination probability was checked by using the area under ROC (AUROC) curve. “givitiR” R-packages were used to check the calibration plot. Bi-normal smoothing bootstrapping technique was used to validate AUROC internally. After bootstrapping, the model’s predictive performance was considered the performance that can be expected when the model is applied to future similar populations. Decision curve analysis (DCA) was used to evaluate the model’s clinical and public health impact.

The Youden index value, which classified high or low risk, was used to predict the likelihood of preterm death. Sensitivity, specificity, and the positive and negative predictive values were computed. The probability of preterm death for each neonate was predicted using the linear predictor of estimated risk of death, which is: P (Risk score for each patient) = risk score X prognostic determinant + … + N (risk score) X N (prognostic determinant). Additionally, a nomogram graphical presentation was prepared to construct an easily clinically applicable individualized prediction of preterm neonate risk of death.

The study participants were informed about the purpose, risk, benefits, and confidentiality of the study, and informed and voluntarily signed written consent has been obtained during the interview previously. Ethical approval was obtained from Debre Tabor University, College of Health Sciences, and the Institutional Review Board (IRB).

A total of 456 preterm neonates were involved in this study. The median age of the respondents was 26 years. Most of the respondents (95.6%) were married, followed by single status (3.5%). Thirty-eight percent of the respondents were unable to read and write. Regarding the occupation of the respondents, 364 (79.8%) of them were housewives, and 68 (14.9%) of them were employees. Three hundred (65.8%) of the respondents were rural residents (Table 1).

Of the total respondents, one hundred seventy-two (37.7%) were Para-one. Most of the mothers (97.4%) have ANC follow-up during this pregnancy. Four hundred sixteen (93.7%) have four times and above ANC follow-up among those with ANC follow-up. About the delivery route, the majority (91.2%) of mothers had vaginal route delivery. Seventy-six (16.2%) of the mothers encountered complications during this pregnancy. APH was the most common one 24 (31.6%) among those who had complications during this pregnancy. The median afterbirth interval of the respondents was 36 months for those gravid two and above (Table 2).

One hundred eighty-four (40.4%) of neonates were with gestational age between 32 and 34 weeks. Two hundred forty (52.6%) of the neonates were male (Table 3).

One-fourth of the neonates had RDS. The majority (89.9%) of preterm neonates were hypothermic during admission. Thirty-six percent of preterm neonates were diagnosed with sepsis. Forty-four (9.6%) preterm neonates had PNA. Seven percent (27) of the preterm neonates had hypoglycemia (Table 4).

The maternal socio-demographic, obstetric, and neonatal-related prognostic determinants were considered to predict preterm mortality. In the final model, the plurality of the neonate (multiple) has no RDS, gestational age of less than 32 weeks, low birth weight, and KMC were the remaining significant predictors of preterm mortality (Table 5).

After the multivariable logistic model, a total of five prognostic determinants were left for the prediction of preterm mortality, and the relative contribution of each prognostic determinant to the probability of preterm mortality was calculated by dividing each beta coefficient by the lowest beta coefficient and rounding to the nearest integer (score chart rule formula) (Table 6).

Therefore, the probability of preterm mortality among preterm neonates using the linear prediction formula was:

Linear predictor of the model (l p) = −1.54 + 2.66 * gestational age less than 32 weeks + 0.77 *multiple pregnancy- 4.4; *has no RDS, + 2.1; *has not gotten KMC, + 1.52; *low birth weight.

The area under the receiver operating characteristics curve (AUROC) had a discrimination power of 92.7% (95% CI: 89.9–95.4) (Figure 1A). The final prediction model was good calibrated (p-value = 0.519) (Figure 1B). The prediction model was particular (specificity = 95%) in its ability to predict preterm death, with a true positive rate (sensitivity) of 77%. This prediction had a low false-positive rate of 4% among the actual death of the preterm neonates. The model prediction accuracy was 90% (Supplementary File 1). The prediction model was internally validated using 2,000 bootstrap replicates.

The cost-benefit analysis of the prediction model was compared between no follow-up for all the preterm neonates, followed by all the preterm neonates, and critically follow selectively based on these predictors. For example, if we critically follow all preterm neonates at a threshold risk of death probability of 30%, we will find 30% prevalence of preterm death. Whereas, if we follow the preterm neonates based on the prediction model, the preterm neonates need critical follow-up at the risk of preterm death threshold of more than 80%. Therefore, following the preterm neonates selected based on the model has the highest cost-benefit ratio (Figure 2).

To construct an easily clinically applicable individualized prediction of preterm neonate risk of death was prepared in nomogram graphical presentation (Figure 3). The best threshold (Youden index) cut point value was 0.3 (30%) to predict a high or low risk of preterm death (Figure 1A). The maximum sensitivity and specificity of the Youden index cut point value were 83.3 and 94.4%, respectively (Figure 1A). The positive predictive and negative predictive values at the Youden index threshold value were 85.4 and 93.3%, respectively. The linear predictor of five parameters is estimated in the probability of preterm death in the nomogram to predict the preterm neonate death. For example, if a preterm neonate was born with multiple births, low birth weight, and the preterm neonate was not under KMC, the predicted probability of preterm death is calculated as follows. The total point for the three predictors was the sum of each predictor point, which was 1.8 + 3.5 + 4.8 = 10.1. Therefore, the probability of preterm death with the corresponding total points in the nomogram was 0.2 (20%), indicating a low risk of preterm death. Whereas a preterm neonate had RDS and low birth weight, the total points were 10 + 4.8 = 14.8. Thus, the probability of preterm death, approximately 0.68 (68%), is declared a high risk of preterm death (Figure 3).