Easy-to-use nomogram to predict neonatal hyperbilirubinemia

Patients

The study included neonates born between January 2021 and January 2024 at the Second Hospital of Dalian Medical University with a GA ≥ 35 weeks. The diagnosis of hyperbilirubinemia was based on both clinical observations and serological evidence, following the recommendations of the 2022 American Academy of Pediatrics (AAP) clinical practice guideline revision (Kemper et al., 2022). Infants with congenital biliary atresia and those with severe malformations or chromosomal abnormalities were excluded. The study was conducted in accordance with the ethical principles of the Declaration of Helsinki and was authorized by the Ethics Committee of Second Hospital of Dalian Medical University (No. KY2024-325-01). The definitions of the stages of labor used in this study are as follows: the first stage of labor prolongation refers to more than 20 h for primiparous women and more than 14 h for multiparous women. The second stage of labor prolongation refers to more than 3 h for primiparous women and more than 2 h for multiparous women. Additionally, newborns were monitored for a duration of 2 weeks.

Data acquisition

We performed a retrospective analysis of the medical records of all participating patients. The initial step involved contacting the guardian of each patient by phone to secure their consent for participation. Upon receiving agreement from the guardian, we sent the informed consent form for their consideration, signature, and subsequent return. A comprehensive literature review and analyses were conducted, and consultations took place with neonatal specialists using the case files of the infants to identify and summarize the risk factors linked to neonatal hyperbilirubinemia. Furthermore, we accessed the child’s medical records and referred to the AAP clinical practice guideline of neonatal hyperbilirubinemia. This process led to the creation of the final version of our “Survey on Risk Factors for Neonatal Hyperbilirubinemia”. Data collection was performed as previously outlined by Zhang et al. (2023), which served as a research instrument to gather pertinent clinical information and establish a standardized data framework.

Statistical analyses

The handling and analyses of data were conducted using R version 4.4.0, the developed by the R Foundation in Vienna, Austria, accessible at https://www.r-project.org/. Data were tested for normal distribution, and those that met normal distribution were expressed as mean ± standard deviation ($\bar{x}$ ± SD) and compared between groups using the independent samples t-test, and those that did not meet normal distribution were expressed as median (M) as well as 25th and 75th percentiles (P25, P75). Categorical variables are articulated in terms of frequency and percentage (n (%)), employing the chi-square test, with the statistic denoted as the χ² value. Continuous variables included GA, birth weight (BW), and maternal age. Categorical variables variables included fetal sex, umbilical cord blood gas pH, Apgar score, blood group type with positive orthogonal resistance, cerebral hematoma, breastfeeding, probiotic supplementation, weight loss >9% within 3 days, cesarean section, gestational hypertension, gestational diabetes, amniotic fluid contamination, PROM ≥ 18 h or combined with intrapartum fever, placental abruption, placenta previa, prolonged labor, and umbilical cord abnormalities. Least absolute shrinkage with selection operator (LASSO) regression analysis was used to screen risk factors for neonatal hyperbilirubinemia. Random sequences were set and the study population was randomly divided into training and validation sets in the ratio of 7:3. Based on the training set data, the risk factors screened by LASSO regression analysis were subjected to multifactorial logistic regression analysis to determine the independent risk factors for neonatal hyperbilirubinemia, which in turn led to the construction of a clinical prediction model column-line diagram. Based on the training set and validation set data, the differentiation ability of the model was examined and calculating the area under the curves (AUCs), and the model was considered to have good accuracy if the AUC > 0.7; the calibration curves were plotted to evaluate the consistency between the actual results and those of the prediction model; decision curve analysis (DCA) assesses the clinical benefits of the model. P < 0.05 was considered a statistically significant difference.

The nomogram scores were derived from the logistic regression coefficients. Each predictor’s score reflects its relative weight in the model, scaled to a 0–100 range for the smallest to largest effect. The total score maps to the predicted probability via a logistic function (Iasonos et al., 2008).

Clinical characteristics of the patients

This study included 646 newborns, and the consort diagram is illustrated in Fig. S1. The subjects were randomly divided into a training set (454 newborns) and a validation set (192 newborns) in a ratio of 7:3. The baseline characteristics are presented in Table 1. The mean GA of all newborns was 38.42 ± 1.38 weeks, and the mean BW was 3,264.09 ± 490.74 g. Based on the relevant diagnostic criteria for hyperbilirubinemia, newborns were categorized into two groups: those with and those without neonatal hyperbilirubinemia. A total of 164 (25.3%) newborns were diagnosed with neonatal hyperbilirubinemia, while 482 (74.7%) newborns were not.

Comparison of clinical characteristics between newborns with and without hyperbilirubinemia

In the hyperbilirubinemia group, there were 92 males (56.1%) and 72 females (43.9%), with a mean GA of 37.54 ± 1.66 weeks and a mean BW of 3,049.85 ± 620.12 g. In contrast, the non-hyperbilirubinemia group was comprised of 258 males (53.53%) and 224 females (46.47%), with a mean GA of 38.72 ± 1.12 weeks and a mean BW of 3,336.98 ± 414.31 g. GA and BW were significantly different between the two groups (p < 0.001), while no significant difference was observed in sex distribution (p > 0.05) (Table S1).

In addition, significant differences were found in maternal-infant blood type incompatibility with direct Coombs test positivity (4.36 vs. 21.34%, p < 0.001), weight loss exceeding 9% within the first three days (0.41 vs. 4.88%, p < 0.001), and probiotic supplementation (62.03 vs. 29.27%, p < 0.001). No notable differences were found between the two cohorts regarding umbilical blood pH < 7.2, Apgar scores ≤ 7, instances of cephalohematoma, and breastfeeding (p > 0.05) (Table S2).

Comparison of maternal clinical characteristics between newborns with and without hyperbilirubinemia

A comparative analysis of maternal clinical characteristics between the hyperbilirubinemia and non-hyperbilirubinemia groups revealed statistically significant differences between rates of: cesarean section (25.31 vs. 38.41%), gestational hypertension (9.75 vs. 18.90%), gestational diabetes (55.39 vs. 45.12%), meconium-stained amniotic fluid (12.86 vs. 6.10%), or prolonged rupture of membranes (PROM) ≥ 18 h or concomitant intrapartum fever (9.13 vs. 26.83%) (P < 0.05). There were no statistically significant differences between the two groups regarding maternal age, gestational hypothyroidism, placental abruption, placenta previa, prolonged labor, and cord abnormalities (p > 0.05) (Table S3).

Development of a nomogram for predicting neonatal hyperbilirubinemia

Six predictor variables with non-zero coefficients were identified using LASSO regression analysis with vertical curves plotted at the λ-minimum (λ = 0.046), including GA, BW, PROM ≥ 18 h or concurrent maternal fever, maternal-infant blood type incompatibility with positive direct Coombs test, supplementation with probiotics, and weight loss >9% within 3 days (Fig. 1). Based on the results of the LASSO regression analysis, the logistic regression equation (Table S4) was established using the “stepwise regression method”, and a user-friendly nomogram for neonatal hyperbilirubinemia was drawn to visualize the model, as shown in Fig. 2.

The nomogram assigns a value to each predictor variable based on the logistic regression model, and an example is shown in Table S5. In the training set (n = 454), the median score was 98 (IQR: 73–135) with a score range of 20–282, while in the validation set (n = 192), the median score was 99 (IQR: 74–125) with a score range of 27–267. Neonates who developed hyperbilirubinemia had significantly higher scores than those without hyperbilirubinemia.

Evaluation and validation of the nomogram

The discriminatory power of the nomogram was assessed using receiver operating characteristic (ROC) curve construction. The training set had an area under the ROC curve (AUROC) of 0.825 (95% confidence interval (CI) [0.777–0.874]). The Youden index was calculated as sensitivity + specificity −1. The maximum value corresponds to a threshold probability of 0.262. This threshold yielded a sensitivity of 73.7% (95% CI [0.658–0.817]) and a specificity of 82.1% (95% CI [0.780–0.862]). Further, according to the column line graph model calculations, when the predicted probability was equal to this optimal cut-off value (0.262), the score corresponding to the total score scale of the column line graph was 118.34 points (Fig. 3A). Additionally, internal validation was performed using data from the validation set (Table 1), where the predictive model exhibited satisfactory performance, with an AUROC of 0.829 (95% CI [0.757–0.901]), sensitivity of 65.2% (95% CI [0.515–0.790]), and specificity of 90.4% (95% CI [0.856–0.952]), as shown in Fig. 3B. In both training and validation sets, the model demonstrated strong discriminatory ability.

The predictive model was evaluated using calibration curves (Fig. 4). The results indicated that the predictions made using the nomogram only slightly differed from the actual observed outcomes, demonstrating good consistency between the two datasets. Moreover, the Hosmer-Lemeshow goodness-of-fit test demonstrated that the model exhibited a satisfactory fit in the training set (P = 0.766), while also showing good accuracy in the validation set (P = 0.443). The nomogram is shown in Fig. 5. DCA revealed that the threshold probabilities for predicting neonatal hyperbilirubinemia were 8%–78% in the training set and 7%–76% in the validation set.