Machine learning-based risk assessment of neonatal perinatal adverse outcomes of anemia during pregnancy: a modeling study

Abstract

Background

Gestational anemia significantly elevates the risk of adverse maternal and neonatal outcomes, necessitating early predictive tools for targeted intervention. This study aimed to develop and validate a robust machine learning (ML) framework to forecast perinatal complications and facilitate early risk identification.

Methods

Perinatal mortality, preterm birth, low birth weight and macrosomia are defined as adverse outcomes. Analyzing a retrospective cohort of 5,710 pregnant women, we identified 22 initial variables using Lasso regression integrated with seven ML-based screening algorithms. Subsequently, eight predictive models were constructed and benchmarked via internal and external validation. Model performance was rigorously evaluated using receiver operating characteristic (ROC), precision‑recall (PR), calibration, and decision curves.

Results

Seven key predictors were identified, including gestational hypertension, obstetric history, and hepatic markers (Albumin, Alanine Aminotransferase, Globulin). The XGBoost model consistently outperformed its counterparts, demonstrating superior discriminative power (area under the curve (AUC) and F1-score) and clinical utility, as confirmed by the DeLong test and Kolmogorov‑Smirnov (KS) statistics. Based on XGBoost probabilities, we established a three-tier risk stratification: low-risk (< 0.28), medium-risk (0.28–0.52), and high-risk (≥ 0.53).

Conclusions

Our ML-based framework offers a reliable tool for early risk assessment in gestational anemia, enabling clinicians to implement individualized management strategies through precise risk stratification.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12911-026-03669-w.

Background

Pregnancy-related anemia arises from inadequate availability of hematopoietic nutrients together with physiologic plasma volume expansion that outpaces the increase in red cell mass. Because pregnancy substantially increases nutritional demands, this condition remains common worldwide [1]. Recent estimates place the global prevalence of anemia during pregnancy at 40%, with a prevalence of 19.9% reported in China [2, 3]. Reducing maternal anemia is therefore a major public health and nutrition priority. Evidence has consistently linked anemia during pregnancy to adverse maternal and neonatal outcomes, including perinatal death, preterm birth, preeclampsia, low birth weight, and small-for-gestational-age infants [2, 4, 5]. Early identification of women at high risk for these outcomes may allow closer surveillance and timely intervention. However, to our knowledge, no study has specifically developed a model to predict adverse perinatal outcomes among women with gestational anemia.

Pregnancy is influenced by multiple interacting maternal and fetal factors, which can limit the performance of conventional prediction models, particularly when nonlinear associations and high-dimensional data are involved. In this context, machine learning (ML) offers clear methodological advantages and has increasingly been applied to clinical diagnosis, prognosis, and treatment evaluation, including in maternal-fetal medicine. Previous studies have shown its potential in obstetric prediction. For example, Mizuno et al. developed an early prediction model for preeclampsia with an AUC of 0.93 [6]. In twin pregnancies, Mustafa et al. reported ML-based prediction models for preeclampsia and gestational diabetes mellitus, with AUCs of 0.62 and 0.67, respectively [7]. Teng et al. retrospectively analyzed routine biomarkers from 2,600 pregnant women younger than 35 years and developed an ML model for preterm birth prediction, achieving AUCs of 0.89 in the training set and 0.91 in external validation [8]. These findings support the feasibility of applying ML to obstetric risk assessment and suggest that ML may also be useful for predicting adverse outcomes among women with gestational anemia.

Several ML algorithms, including logistic regression (LR), support vector machines (SVM), random forest (RF), k-nearest neighbors (KNN), decision trees (DT), and extreme gradient boosting (XGBoost), have been widely used in clinical prediction [9–12]. Because no single algorithm performs best across all datasets, comparing multiple methods may help identify the most suitable model for a given clinical task. To our knowledge, no prior studies have developed machine-learning–based, interpretable prediction models specifically for neonatal perinatal adverse outcomes in pregnancies complicated by anemia. Therefore, this study aimed to develop and compare interpretable ML models for this population. The resulting model may facilitate early identification of women at high risk and support timely intervention to reduce adverse maternal and neonatal outcomes.

Methods

This study follows the principles of the Declaration of Helsinki. The study has been approved by the Institutional Review Board of the First Affiliated Hospital of Anhui Medical University (PJ2022-08-46). The research process is shown in Fig. 1.

Fig. 1.

Research flowchart

Sample

This study consists of two independent and non-overlapping cohorts. Retrospective collection of pregnant women’s data produced in our hospital during the period from January 1, 2023–October 31, 2024, was used for the development of the prediction model and internal validation. Inclusion criteria: (1) Pregnant women diagnosed with anemia during pregnancy (2). Undergo prenatal examination, childbirth, or termination of pregnancy at our hospital. Exclusion criteria: (1) Termination of pregnancy due to fetal structural abnormalities (2). Individual cases with missing values for more than 20% of the 22 pre-collected research variables (≥ 5 variables missing) were excluded. Definition of anemia during pregnancy: Peripheral blood hemoglobin (Hb) less than 110 g/L and hematocrit less than 0.33 L/L‌.

Sample size calculation

Determined based on the statistical requirements for developing a prediction model for neonatal perinatal adverse outcomes in anemia during pregnancy, combined with the characteristics of this study design and the principles of sample size estimation for prediction models. First, referring to previous related studies [5] and preliminary survey data of this research, the incidence of neonatal perinatal adverse outcomes in anemia during pregnancy is approximately 25%~30%, (including perinatal death, preterm birth, low birth weight infants, or macrosomic infants). This study takes a conservative value of 25% as the expected incidence rate of outcome events; Set the allowable error (α) to 0.05 (two-sided test), the statistical power (1 - β) to 0.90 (β = 0.10), and the minimum detectable change in outcome incidence (Δ) to 0.05. Use the sample size calculation formula for research outcome event. The calculated basic sample size is 823 cases.

Data collection

Collect baseline demographic characteristics: Pre-pregnancy weight, age, and whether accompanied by gestational hypertension, gestational diabetes, or gestational thyroid disease, Any history of adverse pregnancy outcomes, whether it is the first delivery, and any cervical abnormalities, number of fetuses (single or twin (multiple)), whether it is natural pregnancy.

Laboratory indicators within one month before delivery: Albumin, Alanine Aminotransferase (ALT), Aspartate Aminotransferase (AST), Creatinine (Cr), Urea (Bun), Uric acid (UA), Globulin, White blood cell count (WBC), Hematocrit (HCT), Red blood cell count (RBC), Hemoglobin, Platelet count.

Neonatal perinatal adverse outcomes include perinatal death, preterm birth, low birth weight, or macrosomia. The standardized clinical definitions for the above outcomes are as follows: preterm birth was defined as live birth or stillbirth at a gestational age of 28 weeks or more but less than 37 completed weeks; low birth weight was defined as a neonatal birth weight of less than 2500 g regardless of gestational age; macrosomia was defined as a neonatal birth weight of 4000 g or more regardless of gestational age.

Data preprocessing

Categorical text variables were converted into numerical codes before analysis, whereas continuous variables were retained in their original scale. Binary categorical variables were numerically encoded to facilitate statistical analysis. The coding scheme is provided in Table 1. For example, the number of fetuses was coded as 0 for singleton pregnancies and 1 for twin or multiple pregnancies. Missing values in continuous variables were imputed using the expectation-maximization method, while missing categorical values were imputed using the mode.

Table 1.

Training set, internal validation set and external validation baseline data sheet

Variables	Total (n = 5710)	Training (n = 3436)	Internal (n = 1888)	External (n = 386)	Statistic	P
Age, M (Q₁, Q₃)	30.00 (28.00, 33.00)	30.00 (28.00, 33.00)	30.00 (28.00, 33.00)	30.00 (28.00, 33.00)	χ²=1.12#	0.572
Albumin, M (Q₁, Q₃)	34.70 (31.80, 36.80)	34.70 (31.70, 36.90)	34.70 (32.30, 36.50)	34.70 (31.72, 37.16)	χ²=1.14#	0.567
ALT, M (Q₁, Q₃)	18.00 (14.00, 24.00)	18.00 (13.00, 24.00)	18.00 (15.00, 23.00)	18.00 (16.00, 25.00)	χ²=11.98#	0.003
AST, M (Q₁, Q₃)	20.00 (18.00, 25.00)	20.00 (18.00, 25.00)	20.00 (18.00, 24.00)	21.00 (19.00, 30.75)	χ²=21.45#	< 0.001
Cr, M (Q₁, Q₃)	44.00 (40.05, 49.84)	44.00 (39.50, 49.92)	44.00 (41.20, 48.80)	44.00 (40.10, 54.00)	χ²=8.26#	0.016
Urea, M (Q₁, Q₃)	3.40 (3.00, 3.90)	3.40 (3.00, 3.90)	3.40 (3.00, 3.80)	3.40 (3.24, 3.94)	χ²=6.58#	0.037
UA, M (Q₁, Q₃)	300.00 (260.00, 356.00)	300.00 (258.00, 360.00)	300.00 (265.00, 337.25)	326.00 (254.00, 415.75)	χ²=24.33#	< 0.001
Globulin, M (Q₁, Q₃)	27.50 (25.50, 29.30)	27.50 (25.30, 29.30)	27.50 (25.80, 29.20)	27.50 (25.41, 29.60)	χ²=2.45#	0.294
WBC, M (Q₁, Q₃)	10.13 (8.43, 11.95)	10.13 (8.35, 12.08)	10.13 (8.61, 11.67)	10.13 (8.30, 12.00)	χ²=0.40#	0.820
HCT, M (Q₁, Q₃)	30.60 (28.40, 33.00)	30.60 (28.24, 33.20)	30.60 (28.60, 32.60)	30.60 (28.02, 33.60)	χ²=0.93#	0.628
RBC, M (Q₁, Q₃)	3.47 (3.18, 3.79)	3.47 (3.17, 3.82)	3.47 (3.21, 3.73)	3.47 (3.11, 3.87)	χ²=2.32#	0.313
Hemoglobin, M (Q₁, Q₃)	99.00 (92.00, 107.00)	99.00 (91.00, 107.00)	99.00 (93.00, 105.00)	99.00 (91.00, 109.00)	χ²=0.53#	0.765
Platelet count, M (Q₁, Q₃)	197.00 (163.25, 235.00)	197.00 (161.00, 237.00)	197.00 (170.00, 229.00)	197.00 (152.25, 244.00)	χ²=0.34#	0.842
Pre-pregnancy weight, M (Q₁, Q₃)	60.00 (60.00, 65.00)	60.00 (60.00, 65.00)	60.00 (59.00, 65.00)	60.00 (60.00, 64.00)	χ²=9.25#	0.010
Gestational hypertension no = 0, yes = 1, preeclampsia = 2, n (%)					χ²=5.39	0.250
0	5044 (88.34)	3025 (88.04)	1671 (88.51)	348 (90.16)
1	262 (4.59)	160 (4.66)	81 (4.29)	21 (5.44)
2	404 (7.08)	251 (7.31)	136 (7.20)	17 (4.40)
Gestational thyroid disease no = 0, yes = 1, n (%)					χ²=1.71	0.425
0	5212 (91.28)	3136 (91.27)	1717 (90.94)	359 (93.01)
1	498 (8.72)	300 (8.73)	171 (9.06)	27 (6.99)
Gestational diabetes no = 0, yes = 1, n (%)					χ²=5.36	0.069
0	4595 (80.47)	2732 (79.51)	1543 (81.73)	320 (82.90)
1	1115 (19.53)	704 (20.49)	345 (18.27)	66 (17.10)
Number of fetuses Single = 0, twin (multiple) = 1, n (%)					χ²=10.62	0.005
0	5139 (90.00)	3057 (88.97)	1725 (91.37)	357 (92.49)
1	571 (10.00)	379 (11.03)	163 (8.63)	29 (7.51)
History of adverse pregnancy outcomes No = 0, yes = 1, n(%)					χ²=5.81	0.055
0	3703 (64.85)	2233 (64.99)	1200 (63.56)	270 (69.95)
1	2007 (35.15)	1203 (35.01)	688 (36.44)	116 (30.05)
whether it is the first delivery Yes = 0, no = 1, n(%)					χ²=3.90	0.142
0	3423 (59.95)	2058 (59.90)	1116 (59.11)	249 (64.51)
1	2287 (40.05)	1378 (40.10)	772 (40.89)	137 (35.49)
Cervical abnormalities no = 0, yes = 1, n(%)					χ²=0.98	0.612
0	5648 (98.91)	3395 (98.81)	1871 (99.10)	382 (98.96)
1	62 (1.09)	41 (1.19)	17 (0.90)	4 (1.04)
whether it is natural pregnancy Yes = 0, no = 1, n(%)					χ²=8.74	0.013
0	5354 (93.77)	3220 (93.71)	1759 (93.17)	375 (97.15)
1	356 (6.23)	216 (6.29)	129 (6.83)	11 (2.85)

#: Kruskal-waills test, χ²: Chi-square test

M: Median, Q₁: 1st Quartile, Q₃: 3st Quartile

Class balancing methods

The retrospective cohort was randomly divided into a training set (70%) and an internal validation set (30%). An external dataset was also used for validation. In the original dataset of this study, the ratio of adverse outcomes (0) to adverse outcomes (1) is 3.11:1. To address the problem of dataset imbalance, we adopted ClusterCentroids (an undersampling method based on K-Means cluster centroids) to balance the training set, thereby improving the model’s prediction performance on minority classes and enhancing its generalization ability and practical application value. ClusterCentroids was prioritized over oversampling methods such as SMOTE given our large sample size and moderate class imbalance ratio, as this K-Means-based undersampling method avoids the artificial generation of synthetic clinical samples that may introduce non-biologically relevant noise and compromise model generalizability, while retaining representative centroid samples of the majority class to preserve the original data distribution and clinical interpretability of the cohort.

Methods for selecting effective features

(1) ML screening of features: Feature importance of all collected variables was calculated sequentially using seven ML models, including LR, RF, KNN, XGBoost, Adaboost, SVM, and Gaussian Naive Bayes (GNB), with automatic parameter search applied for all models (detailed in Additional file 1) (2). Remove features with low weights: Features with extremely low importance weights across multiple models and those containing extreme outliers in the dataset were excluded from the subsequent analysis (3). Final LASSO-selected features: Subsequent to the above screening, feature selection was performed again using a Lasso regression model with the binomial method to determine the final set of modeling variables.

Model development and validation

Eight independent ML techniques were implemented in Python (version 3.8), including XGBoost, LR, LightGBM, Adaboost, DT, GBDT, GNB, and SVM (Additional file 2). To reduce overfitting, 10-fold cross-validation and hyperparameter tuning were performed on the training dataset using grid search. Model performance was evaluated using Area Under the ROC Curve (AUC), the Average Precision (AP) under the PR curve (Precision-Recall Curve), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), accuracy, and F1 score. In addition, the Hosmer–Lemeshow test (P > 0.05) was used to confirm whether the model has a good goodness of fit. The accuracy of model calibration was evaluated by comparing the differences between predicted probabilities and observed probabilities along the 45-degree diagonal in the calibration curve. Clinical usefulness was evaluated using decision curve analysis (DCA), which quantifies the net benefit at different risk thresholds. The best model was selected through the above methods collectively.

Model explanation and visualization

To gain an in-depth understanding and verify the selected optimal model, we conducted a comprehensive interpretive analysis and visualization. First, the model’s ability to distinguish between positive and negative samples was intuitively evaluated using a Kolmogorov-Smirnov (KS) statistic plot, where the maximum value of the KS statistic represents the maximum separation degree of the model at the optimal threshold; Next, a detailed examination of classification performance was conducted on the training set and validation set using a Confusion Matrix. We quantify the number of true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN), which allows us to calculate key metrics such as accuracy, recall, and precision. Given the multi-dimensional nature of model input features, we used a parallel coordinates plot to visualize high-dimensional data, aiming to display the value distribution of different samples across various features and the potential relationship between features and prediction results. Finally, to gain an in-depth understanding of the model’s internal working mechanisms, we applied SHAP (SHapley Additive exPlanations) analysis. This method can reveal the importance ranking of features from a global perspective, clearly identifying which features play a decisive role in overall prediction. Additionally, it provides sample-level explanations from a local perspective, specifically illustrating how the prediction result of a single sample is formed by the cumulative contribution of its various features, thereby revealing the basis on which the model makes specific predictions.

Establishment of risk levels and threshold determination for the optimal model

To ensure the accuracy and clinical operability of risk stratification thresholds, we based them on maximizing the Youden index combined with actual clinical needs. The rationale for selecting these thresholds was to maximize the Youden index and ensure clinical practicability by validating adverse outcome rates across risk groups. The following steps were specifically implemented: First, use the pre-trained XGBoost model to calculate the predicted probabilities of adverse perinatal outcomes for all samples in the training set and validation set, thereby obtaining the prediction probability distribution (0–1) for the entire sample. Next, three candidate thresholds were selected from the 25th percentile (P25), median (P50), 75th percentile (P75) of the prediction probabilities, and the threshold corresponding to the maximum Youden index. Finally, by calculating the incidence of adverse outcomes in each risk group divided by different candidate threshold values, its clinical rationality is verified. This ensures that the incidence in the high-risk group is significantly higher than that in the medium and low-risk groups, and the incidence in the low-risk group is below the clinically acceptable level of 10%, thereby ultimately determining the optimal stratification threshold.

Statistical analysis

All statistical analyses were performed using SPSS version 29.0 and Python version 3.8. Normally distributed continuous variables were presented as mean ± standard deviation (SD), whereas non-normally distributed variables were summarized as median and interquartile range (IQR). Differences between the two groups were assessed using the independent-samples t-test for continuous variables and the chi-square test or Fisher’s exact test for categorical variables, as appropriate. Differences in AUCs were compared using the DeLong test. A two-sided P value < 0.05 was considered statistically significant.

Results

Baseline data

After class balancing, the final dataset comprised 5,710 cases: 3,436 (60.18%) in the training set, 1,888 (33.06%) in the internal validation set, and 386 (6.76%) in the external validation set. Adverse outcomes were observed in 1,049 training cases (30.53%), 481 internal validation cases (25.48%), and 146 external validation cases (37.82%). There was a statistically significant difference in the number of fetuses (Single = 0, Twin or Multiple Birth = 1) among groups (P < 0.05), ALT, AST, Cr, Urea and Pre-pregnancy weight. The others showed no statistically significant difference (P > 0.05)(Table 1). The actual sample size and the number of outcome events were both far greater than the estimated requirements. It can fully support the subsequent development of ML models, validation, and key feature analysis, avoiding model overfitting and ensuring the stability of prediction performance. The research process is shown in Fig. 1.

Selecting effective features

Feature importance scores derived from the seven ML models were presented in Additional file 1. Across the models, five variables showed zero or negligible importance: gestational thyroid disease, cervical abnormalities, whether it is natural pregnancy, RBC, and pre-pregnancy weight. These variables were subsequently excluded, and LASSO regression was repeated for further feature selection. The selected λ value was 0.02. The final variables retained were Gestational hypertension, Number of fetuses, History of adverse pregnancy outcomes, Whether it is the first delivery, Albumin, ALT, and Globulin (Fig. 2; Additional file 1).

Fig. 2.

Final feature selection through Lasso Regression. A: Lasso Regression Coefficient Profile Plot; B: Lasso Regression Cross-Validation Curve

Model development and screening

Eight independent ML models were constructed using XGBoost, LR, LightGBM, AdaBoost, DT, GBDT, GNB, and SVM, respectively. The results of each model were presented in Tables 2 and 3, Additional file 2, and Fig. 3. The results show that XGBoost had an AUC of 0.776 (training set), 0.753 (internal validation set) and 0.715 (external validation set). Sensitivity (Training set: 0.744, Internal Validation set: 0.72, External Validation: 0.704); Specificity (Training set: 0.646, Internal Validation set: 0.639, External Validation: 0.597), F1 score (Training set: 0.568, Internal Validation set: 0.55, External Validation: 0.441), all of which were higher than those of other models. Evaluated by PR curves and ROC curves, it is also the highest in external validation dataset (Additional file 3). On the calibration curve and decision curve, XGBoost also outperforms the remaining 7 models. This indicates that the model developed with XGBoost has high clinical practicality, with strong consistency between predicted and observed results, confirming its reliability. In the subsequent Delong test, it was confirmed that the difference in performance between XGBoost and the other 7 models was statistically significant. Except for the p-value with Adaboost which was greater than 0.05, considering that XGBoost has a higher PR value than Adaboost, XGBoost was selected as the optimal prediction model after comprehensive consideration (Figs. 3, 4 and 5). A sensitivity analysis was also performed by excluding macrosomia from the composite adverse perinatal outcome to assess model robustness, with results presented in Additional file 4 and 5.

Table 2.

Different machine learning methods build different models with different results based on training cohort

Model	AUC (95%CI)	Cutoff (95%CI)	Accuracy (95%CI)	Sensitivity (95%CI)	Specificity (95%CI)	PPV (95%CI)	NPV (95%CI)	F1 (95%CI)	Kappa (95%CI)
XGBoost	0.776 (0.763–0.790)	0.325 (0.314–0.337)	0.674 (0.658–0.690)	0.744 (0.710–0.778)	0.646 (0.610–0.682)	0.461 (0.448–0.475)	0.864 (0.854–0.874)	0.568 (0.566–0.569)	0.33 (0.318–0.341)
Logistic	0.670 (0.653–0.688)	0.271 (0.267–0.274)	0.671 (0.665–0.678)	0.531 (0.515–0.547)	0.728 (0.712–0.744)	0.441 (0.434–0.449)	0.794 (0.792–0.796)	0.482 (0.478–0.485)	0.244 (0.240–0.249)
LightGBM	0.585 (0.573–0.597)	0.289 (0.289–0.289)	0.739 (0.738–0.740)	0.223 (0.221–0.225)	0.947 (0.946–0.948)	0.629 (0.624–0.635)	0.751 (0.751–0.752)	0.329 (0.326–0.332)	0.210 (0.208–0.213)
AdaBoost	0.757 (0.743–0.771)	0.485 (0.484–0.486)	0.652 (0.639–0.665)	0.744 (0.714–0.774)	0.615 (0.584–0.645)	0.439 (0.429–0.449)	0.857 (0.849–0.866)	0.551 (0.550–0.553)	0.297 (0.289–0.306)
DecisionTree	0.585 (0.573–0.597)	0.629 (0.624–0.635)	0.739 (0.738–0.740)	0.223 (0.221–0.225)	0.947 (0.946–0.948)	0.629 (0.624–0.635)	0.751 (0.751–0.752)	0.329 (0.326–0.332)	0.210 (0.208–0.213)
GBDT	0.585 (0.573–0.597)	0.919 (0.915–0.923)	0.739 (0.738–0.740)	0.223 (0.221–0.225)	0.947 (0.946–0.948)	0.629 (0.624–0.635)	0.751 (0.751–0.752)	0.329 (0.326–0.332)	0.210 (0.208–0.213)
GNB	0.670 (0.652–0.687)	0.086 (0.082–0.091)	0.625 (0.617–0.632)	0.623 (0.604–0.641)	0.625 (0.607–0.644)	0.402 (0.396–0.407)	0.805 (0.802–0.808)	0.488 (0.485–0.491)	0.213 (0.209–0.218)
SVM	0.432 (0.413–0.451)	0.444 (0.386–0.502)	0.705 (0.699–0.711)	0.073 (-0.003-0.148)	0.96 (0.923–0.997)	0.457 (0.386–0.528)	0.721 (0.710–0.732)	0.099 (0.024–0.174)	0.038 (-0.001-0.076)

AUC: area under the curve; CI: confidence interval; PPV: positive predictive value; NPV: negative predictive value

Table 3.

Different machine learning methods build different models with different results based on validation cohort

Model	AUC (95%CI)	Cutoff (95%CI)	Accuracy (95%CI)	Sensitivity (95%CI)	Specificity (95%CI)	PPV (95%CI)	NPV (95%CI)	F1 (95%CI)	Kappa (95%CI)
XGBoost	0.753 (0.709–0.796)	0.325 (0.314–0.337)	0.662 (0.644–0.681)	0.72 0 (0.674–0.766)	0.639 (0.600-0.678)	0.448 (0.431–0.465)	0.852 (0.836–0.868)	0.550 (0.535–0.566)	0.304 (0.280–0.328)
logistic	0.668 (0.616–0.721)	0.271 (0.267–0.274)	0.668 (0.661–0.675)	0.528 (0.498–0.559)	0.724 (0.713–0.735)	0.435 (0.425–0.445)	0.792 (0.783–0.802)	0.477 (0.459–0.494)	0.237 (0.216–0.257)
LightGBM	0.585 (0.550–0.620)	0.289 (0.289–0.289)	0.739 (0.730–0.748)	0.223 (0.204–0.242)	0.947 (0.936–0.958)	0.636 (0.587–0.685)	0.751 (0.746–0.756)	0.329 (0.305–0.352)	0.210 (0.185–0.236)
AdaBoost	0.750 (0.706–0.793)	0.485 (0.484–0.486)	0.641 (0.624–0.659)	0.731 (0.696–0.766)	0.605 (0.570–0.640)	0.429 (0.415–0.444)	0.849 (0.838–0.860)	0.539 (0.529–0.550)	0.278 (0.259–0.297)
Decisi onTree	0.585 (0.550–0.620)	0.629 (0.624–0.635)	0.739 (0.730–0.748)	0.223 (0.204–0.242)	0.947 (0.936–0.958)	0.636 (0.587–0.685)	0.751 (0.746–0.756)	0.329 (0.305–0.352)	0.210 (0.185–0.236)
GBDT	0.585 (0.550–0.620)	0.919 (0.915–0.923)	0.739 (0.730–0.748)	0.223 (0.204–0.242)	0.947 (0.936–0.958)	0.636 (0.587–0.685)	0.751 (0.746–0.756)	0.329 (0.305–0.352)	0.211 (0.184–0.226)
GNB	0.669 (0.617–0.721)	0.086 (0.082–0.091)	0.621 (0.607–0.634)	0.617 (0.583–0.651)	0.622 (0.598–0.646)	0.398 (0.385–0.410)	0.802 (0.790–0.813)	0.483 (0.467–0.499)	0.205 (0.183–0.228)
SVM	0.420 (0.364–0.477)	0.444 (0.386–0.502)	0.700 (0.691–0.709)	0.061 (-0.003-0.124)	0.957 (0.921–0.994)	0.419 (0.253–0.585)	0.717 (0.709–0.725)	0.084 (0.017–0.151)	0.020 (-0.012-0.052)

AUC: area under the curve; CI: confidence interval; PPV: positive predictive value; NPV: negative predictive value

Fig. 3.

Forest plots of various machine learning. The error bars in the figure represent the mean and SD of ROC.The model’s ROC mean and SD are obtained through 10-fold cross-validation

Fig. 4.

Evaluation of independent prediction models developed with different machine learning methods. A: ROC Curves of Different Model Training Sets. B: ROC curves of different model validation sets. C: PR curves of different models Training Sets. D: PR curves of different models validation sets. E: Calibration curves of different models validation sets. F: Decision curves of different models validation sets

Fig. 5.

DeLong’s test of different models

Model explanation and visualization

An analysis of the XGBoost model is conducted using a Kolmogorov-Smirnov Statistic Plot to evaluate its ability to distinguish between positive and negative samples (Class 0 and Class 1 in the figure, corresponding to 0 and 1 for adverse outcomes) (Fig. 6A). The result shows the maximum vertical distance between the two curves (0.315 in the figure, corresponding to a threshold of 0.184). It is generally believed that when the KS value is ≥ 0.25, the model has a good discriminative ability. This result indicates that the model has a stronger ability to distinguish between positive and negative samples. Meanwhile, the confusion matrix results also confirm the advantages of the model (Fig. 6C-D).

Fig. 6.

A: Kolmogorov-Smirnov Statistic Plot of the XGBoost model. B: Parallel Coordinates Plot of the XGBoost model. C: Confusion Matrix of the XGBoost model training Sets. D: Confusion Matrix of the XGBoost model validation sets

Parallel coordinates plot is used to visualize and analyze the multi-feature data of XGBoost (Fig. 6B). As can be seen from the figure, on the first four categorical features (Whether accompanied by gestational hypertension, Number of fetuses, History of adverse pregnancy outcomes, Whether it is the first delivery), the lines for Class 0 and Class 1 are almost completely overlapping. This indicates that the two classes have highly consistent values on these categorical features (likely both taking the value 0, i.e., no gestational hypertension, singleton pregnancy, no history of adverse pregnancy outcomes, and first delivery). On the last three numerical features (Albumin, Globulin, ALT), there are obvious differences between the two categories: albumin and globulin: the lines of the two categories almost overlap, indicating that in these two indicators, the value distributions of category 0 and category 1 are basically consistent. ALT: The two lines are completely separated. The green line (Class 1) has much higher values on ALT than the blue line (Class 0). This indicates that the ALT index of Class 1 samples is significantly higher than that of Class 0, serving as the core feature to distinguish between the two classes. To address possible overfitting and residual class imbalance in the baseline XGBoost model, we compared three XGBoost variants in both internal and external validation cohorts: the original model, a class-weighted version, and a nested cross-validation version that incorporated feature selection within each training fold. Model performance was assessed using AUC, sensitivity, specificity, and F1-score. As shown in Additional file 6, the baseline XGBoost model achieved the best overall performance.

SHAP analysis was performed to quantify the contribution of each feature in the XGBoost model. Higher SHAP values indicated a greater contribution to the prediction of adverse outcomes. The features ranked from most to least influential were ALT, Albumin, Globulin, Number of Fetuses, Whether it is the first delivery, Gestational hypertension, and History of adverse pregnancy outcomes. In addition, sample-level SHAP explanation plots were generated for three representative cases. The sample explanation graph visually explains how a single prediction value is composed of the baseline value (the model’s average prediction for all samples, represented as the gray vertical line in the graph) and the cumulative contributions of each feature band. Each subgraph displays the prediction result (f(x)) of a sample, where the red strip indicates that the feature value drives the prediction above the baseline (‘higher’ direction), and the blue strip indicates that the feature value drives the prediction below the baseline (‘lower’ direction). The length of the strip represents the degree to which the feature affects the prediction result, the longer the strip, the greater the influence. Taking the first sample as an example, the baseline value is approximately 0.1, and the final predicted value is 0.25: The red strip for the feature ‘Whether it is the first delivery (Yes = 0, No = 1 = 1.0)’ pushes the predicted value higher than the baseline, and the subsequent red band for ‘ALT = 11.0’ further increases the predicted value; although the blue bands for ‘Globulin = 31.0’ and ‘Albumin = 34.4’ attempt to lower the predicted value, the overall cumulative effect results in a final predicted value that remains higher than the baseline (Fig. 7 and Additional file 7).

Fig. 7.

A: SHAP analysis of each feature in the XGBoost model. B༚Shap’s sample explanation plot (3 samples)

Establishment and validation of XGBoost risk stratification

Patients were stratified into low- (< 0.28), medium- (0.28–0.52), and high-risk (≥ 0.53) groups based on predefined predicted probability thresholds. In the training cohort (n = 3436), the distribution across low, medium, and high-risk categories was 30.01% (n = 1031), 50.00% (n = 1718), and 19.99% (n = 687), respectively. The incidence of adverse outcomes escalated significantly with higher risk strata: 2.72% (28/1031), 36.14% (621/1718), and 58.22% (400/687). Applying consistent thresholds to the internal validation cohort (n = 1888) yielded a highly similar distribution (30.03%, 50.00%, and 19.96%, respectively). The adverse outcome rates also mirrored the training set, increasing from 2.65% (15/567) in the low-risk group to 30.29% (286/944) and 47.75% (180/377) in the medium- and high-risk groups. In the external validation cohort (n = 386), although the patient distribution was slightly more uniform (37.05%, 30.57%, and 32.38% for low, medium, and high risk), the risk stratification remained robust. Adverse outcome rates exhibited a consistent upward trend (30.07%, 33.90%, and 50.40%, respectively), further confirming the generalizability of the model’s risk stratification strategy (Table 4).

Table 4.

Establishment and validation of XGBoost Risk Levels

Type	Risk Level	Predicted probability range	Number of people	Percentage of total number of people in the queue(%)	Number of adverse outcome cases	Percentage of adverse outcome cases
Training set (n = 3436)	Low	< 0.28	1031	30.01 (1031/3436)	28	2.72 (28/1031)
Training set (n = 3436)	Medium	≥ 0.28 and < 0.53	1718	50.00 (1718/3436)	621	36.14 (621/1718)
Training set (n = 3436)	High	≥ 0.53	687	19.99 (687/3436)	400	58.22 (400/687)
Internal validation set (n = 1888)	Low	< 0.28	567	30.03 (567/1888)	15	2.65 (15/567)
Internal validation set (n = 1888)	Medium	≥ 0.28 and < 0.53	944	50.00 (944/1888)	286	30.29 (286/944)
Internal validation set (n = 1888)	High	≥ 0.53	377	19.96 (377/1888)	180	47.75 (180/377)
External validation set (n = 386)	Low	< 0.28	143	37.05 (143/386)	43	30.07 (43/143)
External validation set (n = 386)	Medium	≥ 0.28 and < 0.53	118	30.57 (118/386)	40	33.90 (40/118)
External validation set (n = 386)	High	≥ 0.53	125	32.38 (125/386)	63	50.40 (63/125)

Discussion

To our knowledge, this is the first study to systematically develop and compare eight ML models for predicting adverse perinatal outcomes in pregnancies complicated by anemia. Among them, XGBoost showed the best overall performance and was used to establish a clinically applicable risk stratification framework. In addition, ALT, Albumin, Globulin, Number of fetuses, Whether it is the first delivery, Gestational hypertension, and History of adverse pregnancy outcomes emerged as the most important predictors.

Maternal anemia reduces placental oxygen and nutrient supply and may initially induce compensatory adaptations, including placental enlargement, increased angiogenesis, and enhanced transport capacity. As hemoglobin declines further, these adaptations may fail, resulting in placental dysfunction and a higher risk of adverse perinatal outcomes [13, 14]. Evidence from large cohort studies in China and Canada indicates that the risk of preterm birth rises with anemia severity, increasing by 6%, 16%, and 36% in mild, moderate, and severe anemia, respectively, relative to no anemia [2, 15]. Consistently, an analysis of 20 cohorts showed that hemoglobin concentrations below 110, 100, and 90 g/L were associated with 10%, 33%, and 72% higher risks of preterm birth, respectively [16]. Maternal anemia has also been associated with perinatal death and low birth weight [4, 5, 15, 17, 18]. Accordingly, we aimed to predict adverse perinatal outcomes among pregnant women with anemia, defined as preterm birth, perinatal death, and abnormal birth weight, including low birth weight and macrosomia.

Feature selection strongly affects model accuracy and clinical applicability, yet the optimal strategy remains uncertain. Tree-based importance measures, including those from random forests, may favor variables with many categories or greater dispersion, whereas logistic regression can be compromised by multicollinearity. To improve robustness, we assessed feature importance using multiple ML methods, removed variables with consistently minimal contributions or marked outlier problems, and then applied LASSO for further selection. Besides established predictors of adverse perinatal outcomes, such as multifetal pregnancy, primiparity, gestational hypertension, and a history of adverse pregnancy outcomes [19], ALT, albumin, and globulin also emerged as important predictors in our models. This finding may be partly explained by the close link between maternal anemia and hypertensive disorders of pregnancy; notably, severe anemia has been associated with an almost threefold increase in the risk of preeclampsia, especially in low-income populations [20]. However, gestational hypertension can lead to elevated levels of ALT. As the top-ranked key feature in SHAP analysis, elevated ALT in anemic pregnant women may not be a confounding factor, but a comprehensive reflection of hepatic stress and preeclampsia overlap, with only a minor association with anemia-related low-grade inflammation. Gestational anemia causes maternal hepatic hypoperfusion and nutrient deficiency, inducing mild hepatocellular metabolic impairment and direct hepatic stress that elevates ALT. Moreover, preeclampsia-induced hepatic microvascular ischemia further exacerbates ALT elevation, forming a synergistic effect of anemia and preeclampsia on liver function. While anemia-induced chronic hypoxemia may trigger mild systemic inflammation, this may not be the dominant driver of ALT changes in our cohort. Thus, ALT acts as an integrated biomarker of hepatic stress and preeclampsia comorbidity in gestational anemia, underpinning its high predictive value in the model. Secondly, we identified that elevated globulin levels were also a key risk prediction factor, and we speculate that this may be because anemia during pregnancy is closely related to infectious diseases, thereby leading to increased globulin levels in pregnant women [21, 22].

Class imbalance can severely compromise the performance and generalizability of predictive models [23, 24]. In our original dataset, the ratio of normal (0) to adverse outcomes [1] was 3.11:1, a skewness that often leads to biased predictions and overfitting. To address this, we applied the ClusterCentroids algorithm to balance the training set. As an undersampling technique, it reduces the majority class by synthesizing centroids based on clustering, thereby removing noisy data while preserving essential structural information. Consequently, this approach improves the model’s sensitivity to the minority class and enhances overall robustness.

Eight ML algorithms were developed and evaluated for the prediction of adverse perinatal outcomes in pregnant women with anemia. XGBoost achieved the best overall performance, with the highest ranking across all evaluation metrics. It also showed good calibration and favorable net benefit on decision curve analysis, supporting its potential clinical usefulness. DeLong testing indicated that XGBoost significantly outperformed all competing models except AdaBoost (p > 0.05 for the comparison with AdaBoost). However, given its superior precision–recall performance, XGBoost was selected as the optimal model. We subsequently analyzed and visualized the contribution of individual features within this model.

The established risk thresholds can guide stratified clinical management in practice. Patients classified as low risk may continue with standard antenatal care. Those at intermediate risk should receive increased surveillance, including more frequent fetal ultrasound, non-stress tests, and repeated liver function and hemoglobin evaluation. For high-risk patients, consideration may be given to referral to a tertiary maternal-fetal medicine center, intensified fetal monitoring, hospitalization, or individualized delivery planning to prevent and promptly manage adverse perinatal events. It is also important to consider the clinical consequences of model misclassification at different risk levels. False-positive predictions, especially in the medium and high-risk groups, may result in unnecessary maternal anxiety, increased antenatal testing, and potential over-medicalization such as unindicated hospital admission or intensified fetal monitoring. In contrast, false-negative results in the high-risk group are clinically more concerning, as they may delay appropriate surveillance and intervention, thereby increasing the likelihood of preventable adverse perinatal outcomes. These findings emphasize that this model should be used as an adjunctive clinical tool rather than a definitive decision-making instrument, with final management guided by comprehensive clinical assessment.

Limitations of the Study: (1) Although this study included data from 5,710 cases, compared to multi-center studies, there may be a certain degree of bias (2). Due to a retrospective analysis and missing data, subgroup analysis of the period when anemia occurred has not been performed. At that time, more neonatal outcomes were not included, only perinatal outcomes of newborns were included (3). The gestational age at the diagnosis of gestational anemia was not recorded in this study due to the limitations of the clinical data collection system, which prevented us from exploring the impact of the timing of anemia onset on neonatal perinatal adverse outcomes. Later, we will incorporate models optimized and validated through multi-center, large-sample development, and conduct subgroup analysis based on the timing of anemia occurrence to accurately predict adverse neonatal outcomes at different stages of anemia during pregnancy (4). Another potential limitation is that feature selection was conducted on the entire training cohort before cross-validation, rather than nested within each cross-validation fold. Future studies may employ nested cross-validation for improvement.

Conclusions

Among eight ML models developed for predicting adverse perinatal outcomes in pregnant women with anemia, XGBoost demonstrated the best predictive performance. Based on predicted probabilities, the model also enabled risk stratification. These findings suggest that the model could support early identification of high-risk pregnancies and inform preventive and clinical management strategies. Further validation in multicenter cohorts and translation into web-based applications will be important for future clinical use.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

ALT

Alanine Aminotransferase

AP

Average Precision

AST

Aspartate Aminotransferase

AUC

Area Under the ROC Curve

CI

Confidence Interval

Cr

Creatinine

DCA

Decision Curve Analysis

DT

Decision Tree

F1

F1 Score

GBDT

Gradient Boosting Decision Tree

GNB

Gaussian Naive Bayes

Hb

Hemoglobin

HCT

Hematocrit

IQR

Interquartile Range

KNN

K-Nearest Neighbors

KS

Kolmogorov-Smirnov

LR

Logistic Regression

ML

Machine Learning

NPV

Negative Predictive Value

PPV

Positive Predictive Value

PR

Precision-Recall

ROC

Receiver Operating Characteristic

RBC

Red Blood Cell

SD

Standard Deviation

SHAP

SHapley Additive exPlanations

SVM

Support Vector Machine

UA

Uric Acid

WBC

White Blood Cell

XGBoost

Extreme Gradient Boosting

RF

Random Forest

LightGBM

Light Gradient Boosting Machine

AdaBoost

Adaptive Boosting

SMOTE

Synthetic Minority Oversampling Technique

TP

True Positive

FP

False Positive

TN

True Negative

FN

False Negative

Author contributions

YD conceived the study, collected the data, and drafted the manuscript. FH organized the data and performed statistical analysis. MG collected clinical data and completed data registration. HZ conducted follow-up investigations and data verification. BH collected ultrasound images and clinical information. CG performed follow-up of study subjects. XY collected clinical data and managed the dataset. CZ supervised the study, revised the manuscript, and provided funding support. YZ designed the study, revised the manuscript, and provided administrative support. All authors read and approved the final manuscript.

Funding

This work was supported by the Clinical Natural Science Foundation of Anhui Medical University (2022xkj167). Scientific research project of Anhui Provincial Health Commission (AHWJ2023A30027). Science and Technology New Star Program of Chinese Physician in Ultrasound (KJXX2023002). Open Research Project of Shanghai Key Laboratory of Neuro-Ultrasound for Diagnosis and Treatment (NUS2026014). The funding bodies played no role in the design of the study and collection, analysis, and interpretation of data and in writing the manuscript.

Data availability

The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

This study was performed in accordance with the principles of the Declaration of Helsinki. Approval was granted by the First Affiliated Hospital of Anhui Medical University’s Institutional Review Board, and the patients signed written informed consent (PJ2022-08-46).

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.