Development of machine learning models for early prediction of small for-gestational-age births using maternal sociodemographic and obstetric data

Abstract

Background

Small for gestational age (SGA) is a significant concern in obstetrics, with implications for stillbirth, neonatal mortality, and long-term health outcomes. The early detection of SGA is crucial for prevention and treatment, but current methods have limitations. This study aimed to develop a machine learning (ML) models-based algorithm to predict SGA using sociodemographic and obstetric features during the preconception period.

Methods

We retrospectively analyzed first-trimester attendees (1 Jan 2022–31 Dec 2023) and developed parity-stratified prediction models (nulliparous vs. primiparous) using routinely available sociodemographic and obstetric variables at the first prenatal visit. Five algorithms (logistic regression, random forest, XGBoost, LightGBM, and extra trees) were trained using an 80/20 stratified train–test split with 5-fold cross-validation. Model performance was assessed using AUC-ROC, accuracy, sensitivity, and specificity. Reporting was guided by TRIPOD + AI recommendations for prediction model development and validation.

Results

Among nulliparous women, logistic regression achieved accuracy 72.7% and AUC 0.733 (95% CI for accuracy 0.464–0.990). Among primiparous women, XGBoost achieved accuracy 80% and AUC 0.92 (95% CI for accuracy 0.552–1.000). Anthropometric variables (weight, BMI, height) and previous birth weight (primiparous) were most influential predictors.

Conclusion

An ML model constructed with basic maternal sociodemographic findings and obstetric history may serve as an early prediction tool for SGA during the preconception period, particularly in resource-constrained settings, although broader validation is required.

Introduction

Small for gestational age (SGA) refers to a condition where a fetus fails to reach its optimal growth potential due to underlying pathological factors, most commonly placental dysfunction, maternal hypertension, or inadequate nutrition. SGA is a significant contributor to adverse perinatal outcomes, including stillbirth, neonatal mortality, and both immediate and long-term health complications on a global scale [1, 2]. In medical literature, small for gestational age (SGA) describes abnormal growth, often conflated with fetal growth restriction (FGR). FGR is the failure to achieve expected intrauterine growth because of placental insufficiency as opposed to SGA, which is defined by a birth weight classification [3, 4]. Not every SGA fetus comes with some form of growth restriction, but those with Fetal Growth Restriction (FGR) are more likely to experience unfavorable outcomes during the perinatal period [5]. Being small for gestation age poses a heightened risk of severe perinatal morbidity and mortality, along with potential long-term impairments such as damage to the neurological and cognitive systems in addition to increased risks of cardiovascular and endocrine conditions later in life [6–9].

Although prediction of SGA during pregnancy is valuable, true diagnosis can only be confirmed at delivery by birth weight assessment [10]. Nevertheless, early identification would be transformative in mitigating risks such as intrauterine hypoxia, although it does not prevent outcomes like premature birth or neonatal infection [11]. To tackle the recurring concern, researchers and clinicians have challenged early detection methods of SGA by testing prenatal risk assessment, typically using the maternal history, such as fetal biometry, Doppler ultrasound and placental function markers [12]. These methods have led to FGR detection early. Despite their still limited accuracy in SGA, their community interest in SGA has led to some moderate success in the timely detection of SGA [13, 14]. Moreover, these approaches require specialized equipment and expertise, so it has low accessibility in resource limited settings.

Recent advances suggest that better sensitivities will be achieved through combining maternal medical history, maternal health indicators, biochemical, and advanced imaging with maternal health indicators such as nutrition, blood pressure with higher sensitivities [12, 15]. However, these methods still remain to be developed and may not be very easy to access to people belonging to a diverse background. Promoting early diagnosis is important in limiting the risks associated with SGA and the improvement of prenatal care. Yet, challenges persist under developing these predictive tools for widespread accessibility and making assurance that they are used successfully in different demographics and healthcare settings.

In response, the medical field now uses artificial intelligence (AI) and machine learning (ML) algorithms for predictive modeling in obstetrics in order to reduce these challenges in the field [16–19]. These medical tools have proved their effectiveness because they demonstrate an ability to identify patterns in clinical datasets. Logistic regression and similar traditional statistical models need specific pre-set assumptions about data sets while their ability to process complex nonlinear relations remains severely restricted [20].

This investigation utilized five machine learning algorithms namely Logistic Regression (LR), Random Forest (RF), extreme Gradient Boosting (XGB), Light Gradient Boosting Machine (LGBM), and Extra Trees Classifier to determine SGA birth risk outcomes. The logistic regression stands among the most frequently employed statistics for binary classification through its ability to generate event probability estimates (20). Random Forest functions as an collective learning technique that built multiple decision trees of which it is popular in identifying the significantly prominent outcome in order to boost accuracy and avoid overfitting problems [21]. XGB functions as a sequential boosting algorithm that enhances weak models by developing better versions which can surpass other models for medical data structures [22]. LGBM functions as a gradient boosting framework which offers high efficiency for big clinical data applications [23]. The Extra Trees Calculator relies on random selection of features to enhance robustness without losing clarity [24].

To the best of our knowledge, this study applies five different machine learning algorithms to predict small-for-gestational-age (SGA) births during the preconception period, using maternal sociodemographic and obstetric factors to enable early risk assessment and inform potential clinical interventions. Given the limited performance and accessibility of early SGA screening based on imaging or biomarkers, we explored whether routinely collected maternal sociodemographic and obstetric history at the first prenatal visit could support an inexpensive, parity-specific risk tool that clinicians can use to triage surveillance resources.

Methods

Participants were stratified by parity (nulliparas) because parity fundamentally affects the availability and nature of predictive variables. In primiparas, prior obstetric history such as previous birth weight is available and can contribute to risk prediction. In contrast, such information is absent in nulliparas, requiring reliance on sociodemographic and anthropometric variables. As a result, parity-specific grouping ensures that prediction models are developed on logically distinct data structures and prevents bias from unequal information availability.

Research design and participants

The study employed a retrospective cohort study was at the Eskisehir City Hospital Department of Obstetrics between 1 January 2022 and 31 December 2023. The term ‘nulliparous’ was used operationally to denote women who were in their first pregnancy at the time of enrollment. They were subsequently classified as SGA + or SGA– based on the outcome of that index pregnancy after delivery. Although, by strict obstetric definition, these women would no longer be nulliparous after delivery, this operational terminology was applied to reflect their pregnancy status at enrollment and to enable consistent stratification. Similarly, ‘primiparous’ referred to women who had one prior delivery before the index pregnancy and were classified as SGA + or SGA– depending on the outcome of the current pregnancy. Women were recruited at the first prenatal visit in the first trimester (≤ 13 + 6 weeks). Predictor variables were those available at or before this visit (age, height, weight/BMI, gravida, parity, prior birth weight, pre-existing hypertension/diabetes, smoking). Preeclampsia and gestational diabetes were not used as predictors (they are diagnosed later); they were retained only for descriptive reporting. To ensure a homogenous dataset, the sample population was selected using the following inclusion criteria.

• Aged ˃18 or ˂ 40 years’ old.

• Spontaneous, non-anomalous, and singleton pregnancy.

• Delivery of a non-malformed live born or stillborn neonate after ≥ 24 weeks of gestation.

• Availability of all data regarding socio-demographic findings related to the study (smoking, hypertension, diabetes mellitus, and systemic lupus erythematosus) during the pre-conception period, obstetric history (gravida, parity, previous birth weight), delivery time, and baby’s birth weight.

• Absence of severe fetal chromosomal or structural abnormalities.

• No history of neonatologic disorders in previous pregnancies.

• No pregnancy with aneuploidy or major fetal abnormality resulting in termination, miscarriage, or fetal death before 24 weeks gestation.

The final analysis included pregnant women with complete sociodemographic and clinical information. The variables captured included smoking, hypertension, diabetes mellitus (DM), systemic lupus erythematosus, obstetric history (gravida, parity, previous birth weight), and delivery results.

Study design and case selection

We conducted a retrospective cohort study of all eligible first-trimester attendees during 1 Jan 2022–31 Dec 2023. For model development we used a parity-stratified case–control sampling from within this cohort to balance classes and stabilize training (SGA vs. non-SGA in approximately 1:1 ratio within nulliparous and primiparous strata).

Data collection & analysis

Maternal age, body mass index, gravida, parity, previous birth weight, history of hypertension or DM before pregnancy period, preeclampsia and gestational DM were collected and studied. The pregnancies were then segregated into SGA birth and non SGA births. SGA was defined as birth of a neonate at BW of <10th percentile according to gender and birth week.

The normality of the variables in the analysis was tested by a Shapiro-Wilk test. Data were presented as mean ± standard deviation, categorical variables as frequencies (percentages), and median (minimum and maximum). The independent samples t-test was used to make the comparisons of continuous variables which were found to have normal distributions between groups. The Mann-Whitney U test was used to compare continuous variables of groups for which the distribution did not become normal. The Chi Square test was used to decide the relation between the categorical variables of different groups. Overall, two-sided p values < 0.05 were considered as statistically significant in all the analyses.

Data preprocessing and handling of missing values

Before model advancement, data preprocessing was carried out to improve the quality of the dataset and reduce the potential for bias. For missing data, an imputation approach was utilised in which variables with < 10% missing values were imputed using mean (for continuous variables) and mode for (categorical variables). In comparison, cases with > 10% of the missing data were excluded from the analysis. Outliers in maternal BMI, age, and birth weight were identified using the interquartile range (IQR) method, and extreme values were winsorised or removed to avoid skewed predictions. Continuous variables such as maternal weight, BMI, and birth weight were first z-score transformed, while the smoking status and other medical conditions were transformed into dummy variables for compatibility with ML algorithms.

Machine learning model development

The research operated in compliance with ML principles. The research employed the Extra Trees Classifier together with Light Gradient Boosting Machine (LGBM) Classifier and eXtreme Gradient Boosting (XGB) Classifier and Logistic Regression (LR) and Random Forest Classifier as its ML algorithms. The ML models were selected because of their common and effective use in healthcare globally, their robust performances and due to their popularity [17]. Multiple ensemble-type ML algorithms including LightGBM, XGBoost and Extra Trees as well as Random Forest function by merging collective predictions to improve prediction accuracy and system reliability. Two approaches in ensemble methods known as Bagging and Boosting exist. Bagging works through Extra Trees and Random Forest to decrease model variations by using data subsets for model training yet boosting helps LightGBM and XGBoost models build new predictions that eliminate mistakes from previous models. The LR serves as a supervised learning algorithm designed for binary classification which creates an S-curve logistic function to capture the data correlations between features and outcomes [17]. The researchers utilized Scikit-Learn and XGBoost libraries through the Python framework for model implementation to achieve reproducibility.

Model training, validation, and performance evaluation

To develop the ML model effectively, the overall data was divided into an 80% and 20% ratio, with 80% for training and 20% for testing the sets using the stratified sampling technique to avoid class imbalance and biased model training. The model’s generalisability to minimize overfitting was done using 5-fold cross-validation. Several overfitting control methods were applied, including L1/L2 regularisation for Logistic Regression, hyper parameter tuning using grid and random search optimisation, early stopping on boosting models like XGB and LGBM, and dropout on tree-based models to reduce dependence on specific predictors.

Model performance was analysed using multiple classification metrics. Overall effectiveness was evaluated using accuracy, sensitivity (recall), and specificity. To reflect the model ‘s ability to detect true SGA cases, sensitivity was prioritized, which is a crucial factor in clinical applications where early identification can improve neonatal outcomes. Specificity was considered to capture non-SGA cases to reduce the number of false positives. The model’s performance in distinguishing between SGA and non-SGA cases was measured using the area under the receiver Operating Characteristic curve (AUC-ROC) (Fig. 1). The confusion matrix played a major role in evaluating the degree of misclassification and the trade-off between false positives and false negatives, given the significant implications of this in clinical decisionmaking.

Fig. 1.

Receiver operating characteristic curve graphs of the nulliparous’ variables

The model is intended as an early triage tool at the first visit to (i) flag patients at higher risk of SGA for enhanced surveillance (e.g., growth monitoring) and (ii) identify candidates for early risk-mitigation strategies (e.g., lifestyle counseling, smoking cessation referral). The model is not a diagnostic test and does not replace ultrasound-based assessment.

Feature importance analysis

Significant independent variables that impact the gestational age of the newborn (dependent variables) were identified using the permutation feature importance method. This method assesses the decrease in the model score when the value of a single variable is randomly shuffled, identifying key predictors of the SGA risk. The Permutation Feature Importance Plot for the nulliparous and primiparous groups are presented in Figs. 2 and 3, respectively.

Fig. 2.

Feature importance chart of the group of the nulliparous’ variables. 1: Height 2: Age 3: Weight

Fig. 3.

Feature importance chart of the group of the primiparous’ variables. 1: Weight 2: Previous birth weight <2500 g 3: BMI 4: Height 5: Maternal Age 6: Gravida

Ethical consideration

This investigation was done according to the outlines mentioned in the Declaration of Helsinki and the data was collected after obtained ethical agreement from the Ankara Medipol University Ethical Committee (Approval No. 08.01.2024/8). Due to the study’s retrospective design, patient consent was not required, and the data underwent complete de-identification prior to analysis.

Results

One hundred two mothers were selected based on the pre-defined criteria of the study. These patients were equally divided into having or not having an SGA birth, followed with nulliparous or primiparous categorization. As shown in Tables 1, 26 and 27 nulliparous mothers were with and without an SGA birth, respectively. Similarly with primiparous as 25 and 24 primiparous were with and without an SGA birth, respectively.

Table 1.

The demographic and obstetric findings of the patients included in the study

	Nulliparous SGA+	Nulliparous SGA-	p (Nulliparous)	Primiparous SGA+	Primiparous SGA-	p (Primiparous)
Age (mean ± SD)	25.48 ± 5.04	27.03 ± 5.30	0.228	26.64 ± 3.68	28.92 ± 4.76	0.067
BMI (mean ± SD)	29.01 ± 3.09	28.65 ± 4.08	0.720	31.55 ± 4.44	29.28 ± 2.24	0.032
Smoking (n, %)	25 (92.6%)	26 (100%)	0.157	24 (96%)	24 (100%)	0.322
Gravida (mean ± SD)	1.07 ± 0.27	1.11 ± 0.33	0.610	2.48 ± 0.91	2.63 ± 0.77	0.207
Hypertension before pregnancy (n, %)	26 (96.3%)	25 (96.2%)	1.000	25 (100%)	22 (91.7%)	0.235
DM before pregnancy (n, %)	26 (96.3%)	25 (96.2%)	1.000	21 (84%)	24 (100%)	0.110

Statistical analysis of maternal characteristics

In categorical comparisons of Hypertension before pregnancy and Diabetes Mellitus (DM) before pregnancywere reanalyzed using Fisher’s Exact Test, due to small cell counts that violate the assumptions of the Chi-square test.Among nulliparous women, neither hypertension nor DM before pregnancy differed significantly between the SGA and non-SGA groups (Fisher’s Exact p = 1.000 for both).Among primiparous women, hypertension and DM before pregnancy also showed no significant associations with SGA status (Fisher’s Exact p = 0.235 and p = 0.110, respectively).Therefore, contrary to the initial analysis, no significant associations were found between pre-pregnancy hypertension or DM and SGA outcomes in either parity group (Table 1).

Notably, hypertension and diabetes did not emerge as top predictors in the ML models, despite showing differences in the initial statistical analysis. Instead, anthropometric variables such as height, weight, and BMI contributed more strongly to model performance.

Height, weight, and BMI were the most important parameters among the six (Fig. 2 above) to identify significant predictors of SGA. The ML algorithms were applied to test the data after the training. It should be noted that all candidate predictors were included during model training. Height, weight, and BMI emerged as the most influential variables after training, based on the permutation feature importance analysis, rather than being preselected as model inputs. The results suggested that logistic regression achieved an overall accuracy of 72.7% and an AUC-ROC of 73% in predicting SGA births using maternal sociodemographic and obstetric history (Table 2; Fig. 4). Sensitivity was 50%, indicating that half of the true SGA cases were correctly identified, while specificity was 100%, indicating that all non-SGA cases were correctly excluded (Table 3).

Table 2.

Prognosis prediction results of different machine learning algorithms

Model Name	Model Training Results					Interval
	Nulliparous			Primiparous
	ROC- AUC	Accuracy	Interval	ROC- AUC	Accuracy
Random Forest Classifier	0.467	0.545	0.251–0.84	0.74	0.8	0.552-1.0
LGBM Classifier	0.5	0.455	0.16–0.749	0.5	0.5	0.19–0.81
Logistic Regression	0.733	0.727	0.464–0.99	0.8	0.7	0.416–0.984
XGB Classifier	0.683	0.727	0.464–0.99	0.92	0.8	0.552-1.0
Extra Trees Classifier	0.567	0.545	0.251–0.84	0.4	0.4	0.096–0.704

Fig. 4.

Receiver operating characteristic curve graphs of the primiparous variables

Table 3.

Classifier confusion matrix of the models

	Nulliparous			Primiparous
	Logistic Regression			XGB Classifier
SGA	Yes	No	%	Yes	No	%
Yes	3	3	50.0	4	1	60.0
No	0	5	100.0	1	4	80

In the second part of the study the analysis of statistically pregnant births of primiparas with and without SGA was carried out. As tabulated in Table 1, only the first visit venous glucose had statistically significant difference between the two groups (p = 0.01) and among the six parameters, weight, previous birth weight, BMI, height, maternal age, and gravida was the highest feature importance parameter (Fig. 3). First the training dataset was applied to the ML algorithms and it was observed that the XGB Classifier model was the best algorithm with 80% accuracy rate and AUC-ROC of 92.2% for predicting SGAs with the use of maternal sociodemographic variables and the obstetric history (Table 2; Fig. 4) with sensitivity and specificity of 80 and 80 (Table 3) respectively. In screening for SGA, sensitivity is of greater clinical importance because undetected cases miss the opportunity for early surveillance or intervention. This imbalance reduces the clinical applicability of logistic regression despite its acceptable overall accuracy. Optimizing sensitivity, even at the cost of some specificity, should be prioritized in future model development to ensure that more at-risk pregnancies are correctly identified.

Discussion

The ability to predict intrauterine growth restriction has the potential to change the course of the disease. It will, allow for early identification of the condition, help improve monitoring, and prevent or reduce fetal growth retardation. Therefore, it is extremely important to identify maternal risk factors that can predict intrauterine growth restriction in the early stages of pregnancy. This study aimed to create a machine-learning model to identify SGA by forecasting the probability of pregnant women developing the condition. While hypertension and diabetes showed differences in the initial statistical analysis, they did not rank highly in the ML feature importance results. This reflects the fact that p-values capture associations in isolation, whereas ML models evaluate predictors in a multivariable context where anthropometric factors (BMI, weight, previous birth weight) carried stronger predictive weight. Their lower importance here should therefore be interpreted as dataset- and model-specific rather than a dismissal of their clinical relevance. Although logistic regression demonstrated acceptable discriminatory performance in nulliparous women, its low sensitivity highlights an important limitation for screening applications. From a clinical standpoint, sensitivity is particularly critical in early SGA risk stratification, as undetected cases cannot benefit from enhanced surveillance or preventive interventions. Therefore, despite reasonable overall accuracy, the limited sensitivity of the model may reduce its clinical utility unless classification thresholds or modeling strategies are optimized. It is important to emphasize that although logistic regression achieved perfect specificity, its sensitivity remained low (50%).

Artificial intelligence (AI) techniques are currently being used to improve risk assessment accuracy and predict adverse perinatal outcomes. These methods are utilized in various areas, including general pregnancy risk assessment, prenatal diagnosis, pregnancy diagnosis, hypertensive disorders of pregnancy, fetal growth, stillbirth, gestational diabetes, preterm delivery, and route of delivery. A systematic review revealed that artificial neural network (ANN) methods are the most effective AI applications for aiding the prediction of pregnancyrelated medical conditions [25]. The continuous development of artificial intelligence in obstetrics encounters a significant obstacle because of the need for external validation of standardized models operating on diverse patient groups. The findings from research on singlecenter special populations cannot easily translate into universal knowledge for other demographic groups. The clinical adoption of these models by practitioners demands their validation through multi-ethnic datasets to enhance their practical value [15, 16].

The AI techniques are based on the notion that they have high potential to guide and assist obstetricians and other health professionals in making accurate decisions that rely on extensive data learning. These techniques can provide valuable insights and predictive capabilities that can be utilized in multiple ways with the aim of improving health outcomes.For example, they can be employed in prenatal screening and diagnosis, route of delivery prediction, postpartum care, and personalized care management. They can also be used to predict preterm delivery. By identifying potential risk factors and using trained and validated software, it is possible to detect complications at an early stage that can then be prevented. This has been demonstrated in the case of patients with preterm delivery [26]. In addition, the automation of risk assessment with AI models can decrease the burden on healthcare practitioners and enhance the productivity of clinical processes. Incorporating these models into electronic health record EHR systems may provide integrated real-time decision support to obstetricians regarding high-risk pregnancies almost instantaneously and without any extra effort [27].

Recent studies have suggested that clinical models that include maternal background and Doppler ultrasound in the first or second trimester have shown only moderate success [7]. Biometry of the fetus, placental logistic regression, and placental biomarkers have predictive accuracies ranging from 40% to 70% using traditional methods, and they often struggle with reliable early-stage detection. On the other hand, this study implied that machine learning models demonstrated further improvements in predictive accuracy, attaining sensitivity values of up to 80%, thus positioning them as viable alternatives for early-stage SGA screening. Based on these findings, there is the potential for improved success using artificial intelligence (AI) and machine learning techniques in prenatal risk assessment [18, 19, 28). Several studies have been conducted using different artificial intelligence models to detect fetal growth retardation and artificial intelligence and machine learning have demonstrated the potential to enhance predictions, particularly with cardiotocography (CTG) readings [29]. When additional imaging methods and parameters, such as CTG, are included in the clinical data, the detection rates for SGA can increase up to 93%. However, there are limitations to the meta-analysis, as not all studies utilized the same input variables and samples. Moreover, a significant limitation in the literature is the absence of a direct comparison of AI-based models with conventional risk assessment methods. Developing an evaluation assessment model for AI systems in obstetrics is crucial to enable their adoption into clinical practice. To establish a consensus, clinical prospective studies, such as randomized controlled trials, are necessary. Moreover, ML algorithms need to be evaluated for their cost-effectiveness compared to traditional approaches using Doppler ultrasound and biochemical markers. If ML-based tools can deliver similar or better prediction capabilities at a lower expense, they would be a valuable alternative, especially in healthcare systems with limited resources [19].

There are very limited reports about the prediction of SGA births during the preconception period using ML methods. Previous studies have published two studies about using ML to predict the SGA neonates of pregnant women exposed to agents (radiation and pesticides) that may aHect fetal development [30]. They showed that ML methods could have a role in predicting SGA neonates during the preconception period with a high accuracy rate, sensitivity, and specificity. In another study by Saw et al., the predictive ML methods were constructed with second-trimester fetal ultra-sonographic measurements [31]. The RF and Support Vector Machine (SVM) models were employed. They found that the highest performance rate for both SGA and severe SGA prediction was by the SVM model, achieving an accuracy of 78% and 83%, respectively.

In primiparas, inclusion of previous birth weight (absent in nulliparas) and stronger signal in anthropometrics likely drove higher AUC with tree-based models (XGB). In nulliparas, where historical obstetric features are unavailable, simpler linear structure (LR) generalized best, albeit with lower sensitivity. These findings support parity-specific modeling and feature sets.

Prior regression-based screens that use maternal history ± first-trimester markers typically report AUCs ~ 0.60–0.75, improving with Doppler and combined biomarker panels. Our history-only approach yields AUC ~ 0.73 (nulliparas) and ~  0.92 (primiparas) in internal testing, suggesting parity-informed modeling can match or exceed history-based regression tools and approach combined models in selected contexts. Our study is the first in the literature to have some characteristics compared to other studies. Unlike earlier research, which used specific risk exposures as ultrasound measurements (e.g., radiation and pesticides), we created a predictive maternal sociodemographic and obstetric characteristic model. This improvement expands the model’s practicality, allowing it to be used for preliminary screening in low-resource settings where advanced imaging or laboratory tests are not sufficiently available. It is also noteworthy that we utilised five different ML algorithms for performance predictive benchmarking, which improves the scope of assessing model performance. This comparative analysis provides valuable insights into the selection of optimal ML models for different maternal profiles, improving real-world clinical applicability. Consequently, health professionals can effortlessly assess the likelihood of SGA birth risks during the preconception period without incurring additional expenses. This not only allows for targeted interventions but also enables early lifestyle modifications that may reduce the likelihood of adverse pregnancy outcomes.

This study has common limitations, including its retrospective, single-center design, which may limit generalizability [32]. The study is limited by its small sample size, its retrospective nature at a single center, and the fact that it only involved Turkish patients. Women in their first pregnancy were categorized as ‘nulliparous’ for analysis purposes, even though after delivery they technically became primiparous. Their outcomes were then recorded as SGA + or SGA– based on the delivery of that pregnancy. The study’s strength is compromised by the lack of additional imaging methods, the absence of laboratory parameters, and the exclusion of additional calculation methods. We plan to validate the model in a larger, unselected regional cohort to confirm performance, recalibrate predicted risks to actual prevalence, and assess generalizability across sites.

Conclusions

Machine learning algorithms show potential for improving the prediction and treatment of SGA pregnancies. This approach is potentially more aHordable and accessible, particularly because it does not rely on advanced imaging or laboratory infrastructure and can oHer useful information for healthcare providers to intervene sooner to enhance outcomes for mothers and babies. The study’s results emphasize the significance of utilizing AI-driven methods in obstetrics to improve decision-making and patient care in high-risk pregnancy situations. More research and confirmation of these predictive models could help create individualized approaches for identifying and treating SGA pregnancies. Our research shows that AI/ML techniques can potentially improve the screening for SGA more precisely and cost-eHectively, ultimately enhancing pregnancy outcomes. However, further algorithm improvements and refinements are needed before these methods can be used in clinical settings. Additionally, it is crucial to ensure the consistency of diagnostic criteria and quality assessment before the widespread adoption of this approach.

Authors' contributions

YK, EAS, and ZB contributed to study conceptualization. TT, ÖÇ, and EAS contributed to data collection. ÖÇ contributed to data analysis. YK, TT, and ZB contributed to manuscript preparation. All authors (YK, EAS, ZB, TT) contributed to proofreading and approved the final version of the manuscript.

Funding

This research is based on the medical records of Ankara Medipol University but did not receive any funding from an external organization.

Data availability

The data that support the findings of this study are available from the corresponding author upon request.

Declarations

Ethics approval and consent to participate

This investigation was conducted in accordance with the Declaration of Helsinki and received ethical approval from the Ankara Medipol University Ethical Committee (Approval No. 08.01.2024/8). Due to the study’s retrospective nature, patient consent was not required, and data were de-identified prior to analysis.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.