Machine learning to improve the prediction of Large for Gestational Age (LGA) neonates: a cohort study Authors: Gabriel Cubillos, Max Monckeberg, Alejandra Plaza, Maria Morgan, Pablo A Estevez, Mahesh Choolani, Matthew W Kemp, Sarah W L Li, Sebastian E Illanes, Claudio A Perez Published: 2026 ABSTRACT AbstractBackground Prediction of Large for Gestational Age (LGA) risk is important as it can enable earlier, more effective interventions, and avoid or mitigate cumulative injury to both mother and baby at the time of delivery. The goal of this research is to improve the prediction of LGA using machine learning (ML) models and variables that are widely available as that allow for broad intervention in late pregnancy and during delivery.MethodsAn improved prediction of LGA was achieved using twelve ML models with hyperparameter optimization. Also, to improve the LGA prediction a data augmentation method was employed in the training set. Additionally, improvement in LGA prediction was obtained with four variable selection methods employed to identify the most significant variables. To rank the best models on the validation set after training, the Area under the Receiver Operating Characteristic Curve (AUROC) was used. Finally, to assess the generalization performance, the best models were evaluated on the test set.ResultsOur method enabled us to identify several models with high sensitivity and specificity. The best models included those that achieved a sensitivity of 0.84, a specificity of 0.84, an accuracy of 0.84, and AUCROC 0.83, requiring 14 variables. Another model reached an accuracy of 0.87, a sensitivity of 0.71, and a specificity of 0.90 with an AUCROC of 0.83 (14 variables). Additionally, a model with a sensitivity of 0.58, the same as that described by Hadlock et al. [1] for ultrasound, required 10 variables, and reached an accuracy of 0.91, a specificity of 0.97, and an AUCROC of 0.86. Both models included maternal BMI (body mass index), First Control, Maternal Weight, BMI Last Control, and EFW (estimated fetal weight) as the most important variables.ConclusionsThe main contributions of our study include the prediction of LGA using data obtained from standard clinical evaluations during prenatal care, as well as the development ML models, tuning the hyperparameters to improve prediction results, thus achieving high levels of sensitivity and specificity. To achieve optimal LGA prediction outcomes across different models, a data augmentation approach is introduced, an important improvement in LGA prediction over using only Hadlock’s ultrasound formula.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12884-026-09411-8.Keywords: Large for Gestational Age (LGA), LGA prediction, Ultrasound, Data augmentation techniques, Machine learning algorithms, Commonly accessible variables Introduction Large for gestational age (LGA) infants are infants born with a birth weight above the 90th percentile for gestational age. The overall prevalence of LGA is about 10% across European countries when national charts are used for classification [1], similar to rates seen in Asia [2, 3]. In the United States, the estimated prevalence is 11.2%, but the highest risk of macrosomia (26.0%) was observed among infants born to mothers with pre-gestational diabetes mellitus (GDM) [4]. While diabetes in pregnancy is not the only risk factor for LGA and macrosomic birth, it is one of the main contributors to this condition [5]. Unfortunately, since the prevalence of diabetes in pregnancy is rising globally [6], we expect rates of LGA births to rise accordingly. LGA infants have an increased risk of perinatal complications, such as shoulder dystocia, hypoglycemia and a prolonged stay in the neonatal special care unit. They also contribute to increased maternal risks, such as operative delivery and postpartum hemorrhage [7]. In the long term, LGA infants are more prone to develop obesity, as well as metabolic disease in adulthood [8]. Therefore, prenatal screening of pregnancies at high risk of ending in an LGA birth is crucial, so that timely antenatal interventions may be implemented. However, current screening strategies are suboptimal and fail to adequately identify LGA infants at risk of adverse perinatal outcomes. In clinical practice, estimated fetal weight (EFW) obtained from fetal biometry measurements performed between 32- and 36-weeks of gestation is often used as the sole predictor for LGA births, with data then compared against growth charts to determine if the fetus has an increased risk of LGA. Some popular charts are Hadlock 1991 [9], Fetal Medicine Foundation (FMF) [10], INTERGROWTH-21st [11], and National Institute of Child Health and Human Development (NICHD) [12]. However, a chart-based estimation approach is notoriously inaccurate, with sensitivities of only 60% [13–15]. Even with the use of growth charts that are specific to individual populations or ethnicities, the detection rate of LGA births has only improved marginally [15]. Various groups have attempted to develop predictive models for LGA using several variables apart from fetal biometry with modest performances. However, an easily implementable model that can reach high sensitivity and specificity for predicting LGA with the fewest number of readily collectable variables is still lacking. Machine Learning (ML) approaches, which offer the advantage of being able to analyze large datasets may allow for improved LGA screening accuracy, delivering better outcomes for mother and baby as well as saving time and expenses. When applied to the context of LGA prediction, we hypothesized that, ML would help identify risk factors for LGA that are not typically included in conventional predictive models, and thereafter, allow for the development of models that can improve our predictive capability of LGA infants most at risk for an adverse outcome. A few studies have deployed ML approaches, however with limited performance, and predominantly among women who were already identified through clinical screening to be at risk for an LGA birth [16, 17]. The main objective of this research was to develop ML models that can better predict the diagnosis of LGA at birth, compared to the current standard of using a single measure of fetal biometry via estimated fetal weight or Hadlock percentile ≥ 90th centile by ultrasound. ML consists of a set of computational methods that enable statistical algorithms to learn from data and generalize the results to unseen data [18]. ML models improve their performance as they are exposed to a large number of examples, identifying subtle, but complex signals or patterns to solve a problem [19]. State of the art Various models have been developed to diagnose Large for Gestational Age (LGA) before the birth of the baby [2, 5, 9–12, 14–16, 20–28]. The most common way to diagnose LGA is using the Estimated Fetal Weight (EFW) calculated in the 3rd trimester using ultrasound (US), and various curves and charts, such as Hadlock [9], Fetal Medicine Foundation (FMF) [10], INTERGROWTH-21st [11], National Institutes of Child Health and Human Development (NICHD) [12], and World Health Organization (WHO) [20]. LGA is diagnosed when the birth weight is > 90th percentile based on the standard growth chart of the population [5]. In [14, 21], the performances of some of the previously mentioned growth curves are compared. Along the same line, the z-score of the Abdominal Circumference (AC) measured according to the INTERGROWTH-21st standard is used to predict LGA [22]. In [2], for patients with Gestational Diabetes Mellitus (GDM), a Logistic Regression (LR) model was used to predict LGA having the Head Circumference (HC) measured between 24 and 30 weeks of gestation during the GDM diagnosis, the AC within 2 weeks of delivery, after 34 weeks of gestation, and the Pre-pregnancy Body Mass Index (BMI) was used to predict LGA. In [15], a combination of the FMF fetal growth chart and hemoglobin A1c (HbA1c) was used to predict LGA. This work [15] is limited to patients with pregestational diabetes. In [23], LGA prediction was obtained using maternal characteristics, such as age, BMI, ethnicity, parity, smoking and fetal sex, the change of EFW, the Uterine artery resistance index (UtA-RI), and the Umbilical artery pulsatility index (UA-PI) between measurements performed in the 2nd and 3rd trimesters. In [24], 4 metabolites, measured at 36 weeks of gestation using ultrahigh performance liquid chromatography-tandem mass spectroscopy, were used for LGA prediction. Other models perform LGA prediction using data collected in the early stages of pregnancy. In [5] a Random Forest (RF) model was used with input data collected before the 28th gestational week. The input variables were maternal age, family history of type-2 diabetes mellitus (T2DM), gestational diabetes mellitus (GDM), history of macrosomia, total cholesterol, triglyceride, low-density lipoprotein cholesterol (LBL-C), and fasting blood glucose (FBG). In [16] a Logistic Regression (LR) model was used, having as input variables the maternal height, age, BMI, gestational weight gain, parity, ethnicity, smoking status, family history of diabetes, mean arterial blood pressure, and oral glucose tolerance test (OGTT). The data was obtained between 24 and 32 weeks of gestation. In [25] a LR model was used to predict LGA with data collected before 21 weeks of gestation using the following input variables: age, ethnicity, parity, BMI, smoking, education level, household income, vegetable consumption, multivitamin supplementation, presence of chronic disease, systolic and diastolic blood pressure, hemoglobin, glucose, triglycerides, and ferritin concentrations. In [26] two approaches for LGA prediction were developed, one with data from pre-pregnancy, and another that added data obtained at 26 weeks of gestation. Also, the model separates the LGA prediction for primiparae and multiparae. The best AUCROC performance was reached with an LR model using 26 week data, and with multiparae mothers, having the following input variables: Age, common-law/married status, area-level income quintile, area of residence (urban/rural), smoking before pregnancy, BMI, hypertension, diabetes pre-pregnancy, gravidity, parity, GDM in a previous pregnancy, caesareans, preterm pregnancy, neonatal death, previous macrosomia, and low birth weight, fetal sex, weight gain in pregnancy at 26 weeks, smoking, substance use during pregnancy, GDM, pregnancy-induced hypertension and psychiatric disorder. In [27] an LR model was used to predict LGA using only BMI, plasma protein A (PAPP-A), and multiparity information in the first trimester. In [28] the LR model employed to predict LGA used the following variables: age, pre-pregnancy BMI, weight gain during pregnancy, Apolipoprotein B, and A1 in the first trimester (6–14 weeks). As previously stated, although several studies have addressed LGA prediction, there remains a clinical need for a model that can achieve high sensitivity and specificity using a minimal set of variables. Employing easily accessible variables from routine prenatal screenings will enhance the usability of the LGA prediction model, particularly in low-resource settings where advanced testing may be unavailable or challenging to obtain. The main objective of our work is to develop ML models for predicting the risk of LGA in neonates, with better performance than approaches based on US findings. Variables used for LGA detection include age, maternal weight, BMI, and OGTT. In this study, we present the development of 12 distinct machine learning models, along with the optimization of their hyperparameters to enhance LGA prediction accuracy. Additionally, a variable selection process was implemented, with redundant data being removed to boost model performance. Methods Study design and population This was a retrospective analysis of a dataset obtained from patients who attended the Obstetrics and Fetal Medicine Unit at the Hospital Parroquial de San Bernardo, Santiago, Chile, from 2016 to 2022. All patients whose first antenatal visit occurred before the 15th week and last visit after the 24th week of gestation were included. Patient records with missing data were excluded. A total of 59 input variables were available to develop the models including maternal demographics and medical history, past and current obstetric history, BMI at the start and end of pregnancy, oral glucose tolerance test results, estimated fetal weight (EFW) as well as corresponding Hadlock percentile data (Table 1) obtained by ultrasound in the third trimester. GDM was diagnosed through universal screening according to the IADPSG guidelines [29, 30]. Universal screening for GDM was conducted in the study population in accordance with standard national and international obstetric guidelines. All participants underwent routine screening within the recommended gestational window, and cases of GDM were identified based on established diagnostic criteria. Gestational age was dated according to first-trimester crown-rump length (CRL) measurement. The diagnosis of LGA was made after birth, using local growth charts as recommended by the Chilean Society of Pediatrics (SOCHIPE) [31]. The database was divided into three random partitions: training set (70%), validation set (10%), and test set (20%). The data for the input to the model of each non-binary variable, e.g., age, weight, height, BMI, was normalized by subtracting the mean and dividing it by the standard deviation. Table 1Clinical variables of the patients collected in the datasetVariable/FeatureNon-LGA fetus(n = 1514) Mean IQRLGA fetus(n = 288) Mean IQRAcquisition(GW)Age28.01 (23–32)29.53 (25–34)4–15Maternal Weight (first control) [kg]71.50 (60–81)81.63 (69.38-93)4–15Maternal Weight (last control) [kg]82.18 (71.5–91)93.85 (81.30–103)24–41Height [m]1.59 (1.55–1.63)1.61 (1.56–1.65)4–15BMI (first control) [kg/m^2]28.29 (24.12–31.90)31.65 (26.99–35.20)4–15BMI (last control) [kg/m^2]32.52 (28.62–35.81)36.39 (32.42–39.79)24–41Weight Gain/Loss10.68 (7–14)12.22 (7–16)24–41Gravidity1.25 (0–2)1.63 (1–2)4–15Parity1.00 (0–2)1.29 (0–2)4–15Abortions0.25 (0–0)0.31 (0–0)4–15Vaginal deliveries0.78 (0–1)0.91 (0–2)4–15Caesarean deliveries0.22 (0–0)0.39 (0–1)4–15Stillbirths0.02 (0–0)0.03 (0–0)4–15First trimester fasting glycemia [mg/dL] (1TFG)78.75 (72–84)81.57 (74–87)4–12OGTT (fasting) [mg/dL]77.02 (70–83)82.56 (73–89)24–28OGTT (2 h) [mg/dL]105.70 (86-119.75)122.31 (95.75–141)24–28(%)(%)GDM11.6925.3524–28Pre-pregnancy Use of Tobacco7.4611.464–15Pre-pregnancy Use of Alcohol3.304.514–15Pre-pregnancy Use of Illicit Drugs2.561.744–15Cardiac Disease0.790.354–15Liver Disease0.200.004–15Biliary Disease1.191.744–15Recurrent Urinal Tract Disease2.972.084–15Thromboembolic Disease0.130.004–15Systemic Lupus Erythematosus /Antiphospholipid Antibody Syndrome0.200.004–15Psychiatric Disorder1.984.514–15Endocrine Disorder0.920.004–15Chronic Kidney Disease0.330.004–15Epilepsy0.790.694–15Inflammatory Bowel Disease0.070.004–15Chronic Lung Disease2.444.174–15Gynecological Disorder5.284.514–15Pre-Gestation Diabetes Mellitus0.591.394–15Diabetes Mellitus Treatment**4–28Insulin Resistance3.575.904–15Hypothyroidism4.825.554–15Pregnancy Hypertensive Disorder12.3516.324–41Chronic Hypertension5.557.994–15Antihypertensive Drugs4.164.514–41Use of Aspirin5.554.514–15Preeclampsia4.435.5620–41Preeclampsia Moderate2.644.8620–41Preeclampsia Severe1.850.3520–41HELLP Syndrome0.330.0020–41Eclampsia0.130.0020–41Gestational Hypertension3.304.1720–41Polyhydramnios0.734.8622–41Oligohydramnios3.831.0422–41Malformations1.060.3522–41Fetal Presentation**28–41EFW2458.13(2053-2945.75)3003.42(2442.25–3633.00)28–41Hadlock < p3%1.390.3528–41Hadlock < p5%2.710.3528–41Hadlock < p10%4.360.3528–41Hadlock < p25%14.400.6928–41Hadlock > p75%26.2975.0028–41Hadlock > p90%10.2453.4728–41Hadlock > p95%5.0943.0528–41SGA**11.820.00N/ASGA Severe**3.500.00N/AAGA**88.180.00N/AMacrosomia**1.1260.76N/AFirst Visit (Weeks)**9.99 (8–12)9.88 (8–12)N/ALast Visit (Weeks)**36.70 (36–38)36.68 (36–38)N/AUS Date (Weeks)33.99 (32–35)34.39 (33–36)N/AHadlock percentiles were calculated using Hadlock 1991 curves [9] with the EFW and the gestational age at the time of the US in the third trimesterGW Gestational Weeks*Categorical Variables**Statistical Variables, not used for prediction Clinical variables of the patients collected in the dataset Systemic Lupus Erythematosus / Antiphospholipid Antibody Syndrome 2458.13 (2053-2945.75) 3003.42 (2442.25–3633.00) Hadlock percentiles were calculated using Hadlock 1991 curves [9] with the EFW and the gestational age at the time of the US in the third trimester GW Gestational Weeks *Categorical Variables **Statistical Variables, not used for prediction Data augmentation Data Augmentation (DA) is a commonly used technique in ML studies to significantly expand the training dataset by applying transformations or modifications to the original data, such as adding noise, performing geometric transformations, or altering values. This process helps prevent overfitting, improves model generalization and performance, and increases model robustness by creating variation in the training data that the models may encounter in real-world scenarios. DA can also help in balancing classes with fewer examples [32–35]. We developed a DA method by adding a range of values that the newly generated data could take, based on advice given by specialists in maternal fetal medicine, resulting in creation of ‘new patients’ for training by slightly changing the input values of some variables as follows: (I) Age: must be in a range of ± 2 years of the original patient; (II) Glycemia Tests: must be in a range of ± 5 mg/dL. For both OGTTs, the diagnosis of GDM must not change by itself, i.e., if the original patient had a value of 90 mg/dL in OGTT Fasting, a new patient with 94 mg/dL will not be created; (III) Height: must be in a range of ± 3 centimeters (IV) Weight: Both weight variables, at the first and last visits, can vary their values in a range of ± 5 kg and (V) BMI: The BMI, at both visits, was adapted to the changes of height and weight in the newly created patients, however, a new patient would not be created if the new BMI classification were different from that of the original patient. Additionally, the variable of weight gain/loss, which indicates the difference in weight between the first and the final follow up visits, followed the same logic as that of the BMI; that is, it also adapted to the change of the weight variables, but it did not have any restriction. In the experiments performed, a limited range for the DA range of values provided by the medical specialists was also considered. This limited range, and the original range of values, are shown in Table 2. DA was used in one specific variable or in multiple variables [34]. Table 2The data augmentation (DA) values, along with a limited range of values, provided by the obstetrician-gynecologist. Both ranges were utilized in this studyDA\VariablesAge (Years)Height (cm)Weight Values (kg)BMIWeight Gain/LossGlucose TestsExpert original range± 2± 3± 5***± 5Limited expert range± 1± 1± 2***± 1Weight Values include Maternal Weight (first visit) and Maternal Weight (final follow up). The Glucose Tests include the First Trimester Fasting Glycemia, and Oral Glucose Tolerance Test, both fasting and 2-hour* The BMI was calculated according to the new values in height and weight. However, the new patient was created if it conserved the same BMI classification as the original patient. We used the classification proposed by the WHO [36]** Formula: Maternal Weight (final follow up) – Maternal Weight (first visit) The data augmentation (DA) values, along with a limited range of values, provided by the obstetrician-gynecologist. Both ranges were utilized in this study Weight Values include Maternal Weight (first visit) and Maternal Weight (final follow up). The Glucose Tests include the First Trimester Fasting Glycemia, and Oral Glucose Tolerance Test, both fasting and 2-hour * The BMI was calculated according to the new values in height and weight. However, the new patient was created if it conserved the same BMI classification as the original patient. We used the classification proposed by the WHO [36] ** Formula: Maternal Weight (final follow up) – Maternal Weight (first visit) Development of ML models and evaluation We used twelve ML models and optimized their hyperparameters to reach the best levels of predictive performance. The models are the following: Gaussian Naïve Bayes (GNB), Bernoulli Naïve Bayes (BNB), Decision Trees (DT), Support Vector Machines (SVMs), Multi-Layer Perceptron (MLP), K-Nearest Neighbors (KNN), Logistic Regression (LR), Random Forest (RF), Extra Trees (ET) [37, 38], Balanced Random Forest (BRF) [39], and Gradient Boosting Machines (GB) implemented in Extreme Gradient Boosting Machines (XGB) [40], and Light Gradient Boosting Machines (LGBM) [41]. Each model was trained using the training set for exploring more than three thousand different hyperparameter combinations. The models were implemented in Python 3.11.2 using Scikit-Learn [37], Imbalanced-Learn [39], XGBoost [40], and LightGBM [41] libraries. The main hyperparameters used for each model, and the Grid search used to determine the best ones, including a range of values analyzed [34] can be found in Supplementary Table 1. Model evaluation The performance of all the models was evaluated using a 5-fold cross-validation (CV) [42] on the training set, and a grid search was conducted to find the best hyperparameter values. The grid search method allows for the identification of nearly optimal hyperparameter values through the evaluation of various combinations. Additionally, a variable selection process [43] was carried out to improve model performance by reducing redundant variables. This selection process involved four methods: F-test of ANOVA (Analysis of Variance), Chi-Square Test, Mutual Information (also known as Information Gain) [37], and feature importance from the Balanced Random Forest [39]. The models were trained and tested with different combinations of input variables selected by the four methods mentioned above. Following training, the top 15% of the models with the highest area under the Receiver Operating Characteristic (ROC) curve (AUCROC) [44] were chosen and evaluated on the validation set. The models that performed best on the validation set were selected to find a balance between high Sensitivity and good Specificity [44], and a decision threshold was determined based on the validation set results. Finally, the selected models were assessed using the test set to measure their generalization performance. The test set was not utilized on any previous steps, including training, model selection, and decision threshold determination. The best-performing models were further trained using DA on the combined training and validation sets, and early-stopping was used to determine the number of training epochs. Following training, these models were tested using the testing set, with sensitivity and specificity serving as the primary performance metrics [34]. The following improvements to the methods of [34] were made as well as the required developments for the prediction of LGA. (i) The database was expanded to 1802 patients, (ii) new variables were included: US findings (EFW, Hadlock < p3%, Hadlock < p5%, Hadlock < p10%, Hadlock < p25%, Hadlock > p75%, Hadlock > p90%, Hadlock > p95%), OGTTs (fasting and 2-hour) and last visit data (weight and BMI), (iii) BRF was added to the variable selection methods to improve the prediction performance of the models, with a total of 4 different variable selection methods. BRF was used to select model 33 on Table 7, reaching a high specificity (97%), and 54% sensitivity, (iv) The variable selection procedure was improved by performing an initial macro search (every 5 variables), followed by a second more detailed exploration to determine the optimum number, considering every variable within the range, (v) More values of hyper-parameters in the same range included in the analysis (i.e., ‘n_estimators’, LGA: [10, 50, 100, 200, 400, 500, 800, 1000, 1500, 1600, 2000]) (affected hyperparameters: ‘hidden_layers_size’, ‘learning_rate_init’, ‘learning_rate’, ‘n_estimators’, ‘var_smoothing’, ‘alpha’). (vi) A new Data Augmentation method was implemented for the variables Weight and BMI at the Last Visit, Weight Gain/Loss and OGTTs (fasting and 2-hour). Table 7Comparison between Hadlock > p90% and models with same level of Sensitivity (13, 33) and same level of Specificity (34, 35) and other models with lower number of errors (29, 30)ModelModelTypeNumber of input variablesDataAugmentationAccuracySensitivitySpecificityRecall MacroAUC ROCFPFNFP + FNN/AHadlock > p90%1N/A0.84210.58180.88890.73540.735434235716MLP6DA EO0.89200.58180.94770.76480.838916233929BRF8DA LE0.85040.74550.86930.80740.840340145430LR14DA EO0.86700.70910.89540.80230.829432164833MLP10DA EO0.90860.58180.96730.77460.864910233334XGB14DA EO0.86150.70910.88890.79900.826434165035LR10DA LE0.86150.70910.88890.79900.8552341650Abbreviations: w/o DA No data augmentation, DA LE Data Augmentation w/Limited Expert range, DA EO Data Augmentation w/Expert original range Comparison between Hadlock > p90% and models with same level of Sensitivity (13, 33) and same level of Specificity (34, 35) and other models with lower number of errors (29, 30) Abbreviations: w/o DA No data augmentation, DA LE Data Augmentation w/Limited Expert range, DA EO Data Augmentation w/Expert original range Accuracy, sensitivity, specificity, and recall macro were measured using a specific decision threshold, as determined by the validation set. The ROC curve was plotted by modifying the decision thresholds to alter sensitivity (the True Positive Rate) as a function of the False Positive Rate. The following formulas were used: Accuracy = (TP + TN)/(TP + FP+TN + FN), Sensitivity = TP/(TP + FN), Specificity = TN/(TN + FP), and Recall Macro = (Sensitivity+Specificity)/2 [34]. Hadlock replacement Models obtained were also tested replacing Hadlock variables with a combination of Ultrasound Date (US Date) during gestational week, and EFW, with the objective of verifying the utility of Hadlock percentiles, and to check if it was possible to replace it with a curve made by the models, that could fit better with the population used in this study. Results Population characteristics This study included a total of 1,802 pregnant women. The dataset was divided into a training set with 1,261 patients, a validation set with 180, and a test set with 361 patients. Within the test set, 55 neonates were diagnosed with LGA. The overall prevalence of LGA was 16.0% (288 out of 1,802). Table 1 shows the variables collected in this study. Variable selection A total of 59 variables are available in our dataset for predicting LGA. Since some of them may be irrelevant or redundant, variables were selected using the 4 methods mentioned in Sect. 2.5: F-Test ANOVA, Chi-Square, Mutual Information, and BRF. The variable selection was performed with groups of variables, starting with 5, and then increasing the number in steps of 5 variables each up to 59. The performance of each model was assessed with each set of variables. The model with the best balance between Sensitivity and Specificity (Recall Macro) required 10 variables, and therefore, a new search was performed, using a number of variables between 6 and 14, in a step of 1 variable. The best results were reached using the Mutual Information method for variable selection, resulting in models that had 11 input variables- Hadlock > p95%, Hadlock > p90%, Hadlock > p75%, BMI last control, Maternal weight last control, BMI First control, EFW, Hadlock < p25%, Age, Gynecological disorders and polyhydramnios. Table 3 shows the 11 variables that reached the best results by the 4 selection methods. Table 3List of eleven most significant variables for predicting LGA, identified through four methods: F-Test ANOVA, Chi-Square, Mutual Information, and Balanced Random ForestRankingF-Test ANOVAChi-SquareMutual InformationBRF1Hadlock > p95%EFWHadlock > p95%Hadlock > p75%2Hadlock > p90%OGTT 2-hoursHadlock > p90%EFW3Hadlock > p75%Maternal Weight Last ControlHadlock > p75%Maternal Weight Last Control4Maternal Weight Last ControlHadlock > p95%BMI Last ControlBMI Last Control5EFWMaternal Weight First ControlMaternal Weight Last ControlWeight Gain/Loss6BMI Last ControlHadlock > p90%BMI First ControlBMI First Control7Maternal Weight First ControlHadlock > p75%EFWMaternal Weight First Control8BMI First ControlBMI Last ControlHadlock < p25%OGTT 2-hours9OGTT 2-hoursOGTT FastingAgeOGTT Fasting10OGTT FastingBMI First ControlGynecological DisordersHeight11Hadlock < p25%DM TreatmentPolyhydramniosHadlock > p95% List of eleven most significant variables for predicting LGA, identified through four methods: F-Test ANOVA, Chi-Square, Mutual Information, and Balanced Random Forest Model performance Table 4 shows the prediction of LGA at birth when using only the Hadlock definition of EFW ≥ 90th centile. Table 5 shows the prediction results with our ML models, as well as the model type, number of input variables, whether or not DA was used, and the type of DA, if used. Both Tables 4 and 5 show the results obtained for Accuracy, Sensitivity, Specificity, Recall Macro, AUCROC, False Positive (FP), False Negative (FN), and FP + FN (total number of errors). Results of all previous metrics were computed using the test set As mentioned in the Methods section, the test set was used solely to assess the generalization performance of the models, and was not involved in training or hyperparameter selection. Table 5 shows the best model for each sensitivity level. Models with high sensitivity result in minimal false negatives when screening patients. This is important since timely detection and diagnosis of LGA can prevent unfavorable consequences. Table 4Performance using only Hadlock > p90% as the prediction variableVariableAccuracySensitivitySpecificityRecall MacroAUC ROCFPFNFP + FNHadlock > p90%0.84210.58180.88890.73540.7354342357 Performance using only Hadlock > p90% as the prediction variable Table 5Top models for different sensitivity levelsModelModelTypeNumber of input variablesDataAugmentationAccuracySensitivitySpecificityRecall MacroAUCROCFPFNFP + FN1MLP25DA LE0.515210.42810.71410.865217501752MLP30DA LE0.52630.98180.44440.71310.830517011713MLP15DA LE0.59000.96360.52290.74330.812514621484MLP15DA EO0.64820.94550.59480.77010.848212431275MLP15DA LE0.67870.92730.63400.78060.822411241166MLP15DA LE0.69810.90910.66010.78460.840410451097MLP10DA LE0.73960.89090.71240.80170.839788694 8 BRF 10 DA EO 0.8033 0.8727 0.7909 0.8318 0.8404 64 7 71 9BRF10DA EO0.80610.85450.79740.82600.840462870 10 BRF 10 DA EO 0.8255 0.8364 0.8235 0.8299 0.8404 54 9 63 11 BRF 15 DA EO 0.8338 0.8182 0.8366 0.8274 0.8412 50 10 60 12 BRF 15 DA EO 0.8393 0.8000 0.8464 0.8232 0.8409 47 11 58 13 BRF 15 DA EO 0.8504 0.7455 0.8693 0.8074 0.8351 40 14 54 14MLP15w/o DA0.85870.70910.88560.79740.846135165115MLP30w/o DA0.88640.65450.92810.79130.841922194116MLP15DA EO0.89200.58180.94770.76480.8389162339In boldface, the models with the best Recall Macro Abbreviations: w/o DA No data augmentation, DA LE Data Augmentation w/Limited Expert range, DA EO Data Augmentation w/Expert original range Top models for different sensitivity levels In boldface, the models with the best Recall Macro Abbreviations: w/o DA No data augmentation, DA LE Data Augmentation w/Limited Expert range, DA EO Data Augmentation w/Expert original range The best models, based on the performance of Recall Macro, are models number 8 and 10 (boldface on Table 5); both models are of the same type (BRF with equal hyperparameters and variables), but with different decision thresholds. Model 8 shows a sensitivity of 0.87, and a specificity of 0.79, with a Recall Macro of 0.8318, when reaching an accuracy of 0.80. Model 10 presents a more balanced performance, with a sensitivity of 0.84, a specificity of 0.82, a Recall Macro of 0.83, and an accuracy of 0.83. Both models, 8 and 10, reached an AUCROC of 0.84. Also, as can be observed on Table 5, and in Fig. 2, several models can be chosen according to the desired sensitivity and specificity. Table 6 shows the best results when selecting inputs among 6–14 variables. Models 18, 20, 21, and 23–30 improved their performance while maintaining the same sensitivity level using fewer variables. In that table, the best models were models number 24 and 26 (boldface on Table 6); both models are of the same type (BRF). Model 24 had a sensitivity of 87%, and a specificity of 80%, with a Recall Macro of 83.8%, when reaching an accuracy of 81%. Model 26 had more balanced performance, with a sensitivity of 84%, a specificity of 84%, a recall of 84%, and an accuracy of 84%. Both models, 24 and 26, reached an AUC over 0.83. Among the models in which the Hadlock variables were replaced by US Date, and EFW, an MLP model (model 33) with 10 variables, reached an accuracy of 0.9086 (the highest among all the models), a sensitivity of 58.2%, and a specificity of 96.7%, with a recall macro and AUC of 0.77 and 0.86, respectively. The numbers of FP, FN, and FP + FN (error) were 10, 23, and 33, in that order. Surface displaying all available models with the number of variables ranging from 6 to 14. The yellow dots highlight the top-performing models, indicated in bold on Table 6 (model numbers 24, 25, 26, and 27) Table 6Top models with 6 —14 input variables for different sensitivity levelsModelModelTypeNumber of input variablesDataAugmentationAccuracySensitivitySpecificityRecall MacroAUC ROCFPFNFP + FN17MLP8DA LE0.501410.41180.70590.8419180018018BRF8DA EO0.53190.98180.45100.71640.8325168116919MLP9DA EO0.58170.96360.51310.73840.8172149215120MLP14DA EO0.65650.94550.60460.77500.8277121312421MLP14DA EO0.67870.92730.63400.78060.8277112411622BRF11DA EO0.68980.90910.65030.77970.8308107511223BRF11DA EO0.77290.89090.75160.82130.837376682 24 BRF 11 DA EO 0.8144 0.8727 0.8039 0.8383 0.8423 60 7 67 25 BRF 11 DA EO 0.8199 0.8545 0.8137 0.8341 0.8423 57 8 65 26 BRF 14 DA EO 0.8366 0.8364 0.8366 0.8365 0.8321 50 9 59 27 BRF 11 DA EO 0.8449 0.8182 0.8497 0.8339 0.8450 46 10 56 28BRF7DA EO0.83930.80000.84640.82320.829747115829BRF8DA LE0.85040.74550.86930.80740.840340145430LR14DA EO0.86700.70910.89540.80230.829432164831MLP12DA EO0.88370.65450.92480.78970.836123194232MLP11DA EO0.88920.58180.94440.76310.8346172340In boldface, models with the best Recall Macro Abbreviations: w/o DA No data augmentation, DA LE Data Augmentation w/Limited Expert range, DA EO Data Augmentation w/Expert Original range Top models with 6 —14 input variables for different sensitivity levels In boldface, models with the best Recall Macro Abbreviations: w/o DA No data augmentation, DA LE Data Augmentation w/Limited Expert range, DA EO Data Augmentation w/Expert Original range Figures 1 and 2 provide two alternative views of the same surface that depict the model performance for different values of total errors (FP + FN), True Positives, and the number of input variables used in each model. Figure 1 shows the results for the first variable analysis, including a number of variables from 5–59 with multiples of 5, while Fig. 2 shows the results for a number of variables between 6 and 14. It is possible to choose different models, based on the requirements of sensitivity and specificity desired, since high sensitivity and high specificity models are available. On the surface displayed in Fig. 1 the red dots represent the best models in boldface on Table 5 (model numbers 8, 10, 11, 12, and 13), while in Fig. 2 the yellow dots represent the best models in bold type from Table 6 (model numbers 24, 25, 26, and 27). Surface showing all available models, along with different hyperparameter values, corresponding error levels (FP + FN), True Positives, and number of variables. The best models, indicated by red dots and listed in bold on Table 5 (model numbers 8, 10, 11, 12, and 13), are represented Figure 3 presents the ROC curves for the models highlighted in bold in Table 5 (a), and Table 6 (b). In Fig. 3(a), the ROC curves correspond to the models in Table 5 with sensitivities of 0.8727, 0.8545, 0.8364, 0.8182, 0.8000, and 0.7455. Figure 3(b) displays the models in Table 6 with sensitivities of 0.8727, 0.8545, 0.8364, and 0.8182. In the Supplementary Material, Supplementary Fig. 1 presents the ROC curves for all models listed in Tables 5 and 6, including Model 33. ROC curves corresponding to the models listed in Table 5, 0.8727, 0.8545, and 0.8364 (BRF 10 Variables, same curve for the three models 8, 9, and 10), 0.8182 (BRF 15 Variables), 0.8000 (BRF 15 Variables) and 0.7455 (BRF 15 Variables). (b) ROC curves of the models listed in Table 6, showing sensitivities of 0.8727 and 0.8545 (BRF 11 Variables, same curve for both models 24 and 25), 0.8364 (BRF 14 Variables) and 0.8182 (BRF 11 Variables) Supplementary Table 2 of the Supplementary Material presents the mean AUCROC, the 95% confidence interval, and the standard deviation of the models, calculated using five different random seeds for initialization. Comparison with Hadlock Table 7 shows a comparison between Hadlock > p90% and models with the same level of Sensitivity (16, 33), and the same level of Specificity (34, 35), and other models with a lower number of errors (29, 30). Comparison of models with/without data augmentation Supplementary Table 3 shows the performance comparison among models presented on Tables 5 and 6, with Data Augmentation (w/DA), and without Data Augmentation (w/o DA). The comparison includes the same models with the same hyperparameters w/DA, and w/o DA. The best models included DA. Discussion The main findings of this research are: (i) LGA prediction based on routine examinations; (ii) The creation and optimization of twelve distinct machine learning models to maximize prediction performance, along with their hyperparameters; and (iii) The proposed data augmentation method with a positive effect enhancing LGA prediction outcomes across most models; (iv) A significant improvement over Hadlock methodology was obtained in all metrics (accuracy, sensitivity, specificity, AUC); and (v) Improved sensitivity while maintaining specificity which helps in reducing the number of ‘missed’ LGA diagnoses (false negatives), and therefore, reducing complications during the delivery. Instead, an increment in specificity reduces possible unnecessary cesarean sections that may cause several additional complications to the mother and newborn (e.g., infection, increased risk of respiratory distress in the newborn, difficulties in future pregnancies, etc.). Incorrect prediction of fetus weight before the delivery may cause either of both FN and FP, with their respective complications. Table 6 shows the results of variable selection between 6 and 14 variables. It can be seen that model numbers 18, 20, 21, and 23–30 reach better results than those of models with the same sensitivity level presented on Table 5. Better results are achieved either when metrics provide improved results, or when the results reach the same performance with fewer required numbers of variables. Comparison of ML models with Hadlock Comparing the performance of one of our best models (Model 24, BRF 11 Variables; Table 6) to that of the ultrasound based Hadlock definition (Baseline Case, Hadlock > p90%; Table 4), it could be seen that the use of additional variables made a substantial improvement of 50% in sensitivity while reducing specificity by only 9.56%. A direct comparison is also available on Table 7, where models with the same sensitivity and the same specificity are compared to the Hadlock base case. For example, model 33 has an additional improvement of 8.82% in specificity with the same level of sensitivity as Hadlock. Also, models 34 and 35 achieve an improvement of 21.88% in sensitivity with the same level of specificity (88.89%). Models presented in Tables 5 and 6 use a combination of Hadlock and other variables as input. Model 33 does not use Hadlock, but a combination of EFW, US Data and the other variables. Comparison of models with and without data augmentation The use of DA had a positive effect in most of the models that use Hadlock as input variables (models 1–13, 16–33), as can be seen on Supplementary Table 3. Just two models, 14 and 15, had better performance without using DA. The maximum improvement in specificity by using DA is 201.91% (model 3, Supplementary Table 3), and the maximum improvement in AUCROC is 3.34% (model 1, Supplementary Table 3). It is especially interesting to note that the average improvement in specificity by using DA is 33.05%, with the same sensitivity. Analysis of the models at fixed false positive rate (FPR) Performance at predefined FPR thresholds is critical to ensure unbiased clinical interpretation, improve clinical decision-making, and reduce unnecessary interventions, particularly in the context of high cesarean delivery rates. Table 8 presents the performance of the model at fixed false positive rates of 5% and 10%, corresponding to specificities of 95% and 90%, respectively. Table 8Performance of the model at fixed false-positive rates of 5% and 10%, corresponding to specificities of 95% and 90%, respectively. Models presented are 1, 16, 24, 26 and 30 with fixed FPR 5% and 10%. Additional models 36, 37, 38, and 39, not shown previously, were includedModelModelTypeNumber of input variablesDataAugmentationAccuracySensitivitySpecificityRecall MacroAUC ROCFPFNFP + FN1–5% FPR*MLP25DA LE0.87530.45450.95100.70280.86521530451–10% FPR**MLP25DA LE0.85320.60000.89870.74930.865231225316 − 5% FPR*MLP15DA EO0.89750.61820.94770.78290.838916213716 − 10% FPR**MLP15DA EO0.86150.63640.90200.76920.838930205024 − 5% FPR*BRF11DA EO0.83660.23640.94440.59040.842317425924 − 10% FPR**BRF11DA EO0.84760.52730.90520.71630.842329265526 − 5% FPR*BRF14DA EO0.85040.29100.95100.62090.832115395426 − 10% FPR**BRF14DA EO0.86660.47270.90200.68730.832130295930 − 5% FPR*LR14DA EO0.86430.43640.94120.68880.829418314930 − 10% FPR**LR14DA EO0.86700.70910.89540.80230.829432164836 − 5% FPR*MLP13DA LE0.88640.52730.95100.73910.823715264137 − 10% FPR**MLP14No DA0.86700.67270.90200.73910.835930184838 − 5% FPR*MLP8DA LE0.88370.50910.95100.73000.832915274239 − 10% FPR**MLP12DA EO0.86700.67270.90200.78730.8309301848*: Closest value to 5% FPR/95% of specificity reached by the model**: Closest value to 10% FPR/90% of specificity reached by the model Performance of the model at fixed false-positive rates of 5% and 10%, corresponding to specificities of 95% and 90%, respectively. Models presented are 1, 16, 24, 26 and 30 with fixed FPR 5% and 10%. Additional models 36, 37, 38, and 39, not shown previously, were included *: Closest value to 5% FPR/95% of specificity reached by the model **: Closest value to 10% FPR/90% of specificity reached by the model High specificity is important in the prediction of LGA fetuses, as false-positive classification may influence clinical decision-making, leading to increased surveillance, heightened patient anxiety, and potentially unnecessary obstetric interventions, such as elective induction or cesarean delivery, which may increase maternal and fetal morbidity. However, identification of LGA fetuses is also important because they are at increased risk of adverse outcomes, including shoulder dystocia, birth trauma, and intrapartum complications. Therefore, a clinically effective screening tool must balance high specificity with sufficient sensitivity to avoid missing truly affected cases. Thresholds that prioritize specificity alone may reduce sensitivity to a level that limits the clinical utility of the tool. For this reason, in addition to reporting performance at fixed FPR thresholds, we also present operating points that provide a more balanced trade-off between sensitivity and specificity (highlighted in Tables 5 and 6). These thresholds may be particularly relevant in clinical contexts where LGA risk assessment is used to guide individualized monitoring rather than to directly trigger intervention. Together, the presentation of results at fixed FPRs (5% and 10%) and at balanced operating points allows clinicians and researchers to interpret model performance across different clinical use-cases, from conservative decision-making requiring high specificity to broader risk identification strategies where sensitivity is also important. Comparison with the current state of the art Datasets from previously published papers are not widely available to the public, which presents challenges when attempting to compare model performances directly [2, 5, 14–16, 21–28]. Moreover, previous studies used different curves, according to the population studied, to generate the models that predict LGA at birth, and therefore, model results are not directly comparable. The performance of LGA predictions from previous studies that use similar input variables are shown on Table 9. Table 9Results of models from the published literature using similar input variablesModelsAccuracySensitivitySpecificityRecall MacroAUC ROC1 Variable**, NC MSUH [14]-0.7380.800.769*0.831 Variable**, FMF [21]0.8444***0.7720.8540.813*0.811 Variable** [22]----0.783 Variables, LR, GDM [3]0.8670.8110.8740.8425*0.9162 Variables, LR, PGDM [15]-0.7860.8000.793*0.857 Variables, LR [23]-0.430.900.665*0.76Datasets used in some previous studies are different and not publicly available*Values calculated by us from the results displayed, using the formula of recall macro: (Sensitivity + Specificity)/2**Deterministic Model***Inferred Results of models from the published literature using similar input variables Datasets used in some previous studies are different and not publicly available *Values calculated by us from the results displayed, using the formula of recall macro: (Sensitivity + Specificity)/2 **Deterministic Model ***Inferred Supplementary Table 4 shows a list of input variables used in each of the models, including models of the SOTA and those developed in our study. This study has some limitations, which we summarize as follows. The retrospective nature of the data and with a single center design may be considered as a constraint to the generalization of our findings to populations with different demographic characteristics or LGA prevalences. Although the use of internal validation and data-augmentation strategies to reduce overfitting and enhance the robustness of the model, we recognize that external validation in independent and prospective cohorts may be needed before a routine clinical adoption can be performed. From a clinical perspective, although the models in our study demonstrated improved predictive performance compared with conventional ultrasound-based approaches, their impact on clinical decision-making to improve neonatal outcomes should be assessed in clinical practice. However, our proposed method implementation appears feasible after a proper prospective validation, because the models rely exclusively on routinely collected clinical and ultrasound variables, and it could be easily integrated into ultrasound reporting software and automated risk calculation in a clinician-friendly decision-support interfaces, as it is done nowadays for preeclampsia [45]. Conclusions The main contributions of our study include achieving high-sensitivity LGA prediction using routine examinations, developing and optimizing twelve machine learning models and their hyperparameters to maximize LGA prediction performance, and proposing a data augmentation method to enhance LGA prediction outcomes across most models, that is an important improvement over using Hadlock; and an increment in sensitivity, while keeping the same specificity, that may help in reducing non-detected LGA fetuses and the higher risk of complications during delivery. Also, for the same level of sensitivity, an increment in specificity reduces possible unnecessary cesarean sections that may cause several additional complications to both the mother and newborn (e.g., infection, increased risk of respiratory distress in the newborn, difficulties in future pregnancies, etc.). Our models provide a large range of options for sensitivity and specificity that can be chosen by the specialist as an operating point. Our results are relevant since the incorrect prediction of fetus weight before delivery may cause an increase in associated complications. We selected a model that offers a good balance between specificity and sensitivity, which could be valuable in managing patients with a probable LGA diagnosis (Supplementary Table 5). Our proposed Model 26 (Table 6) demonstrates a strong balance, with both sensitivity and specificity above 0.836. However, the medical specialist may opt for alternative thresholds depending on the clinical decision to be made, for example, whether to proceed with a vaginal trial or to perform an elective cesarean section. Model 30, also presented in Table 6, has a lower sensitivity (0.71) but a high specificity of 0.895, which may be more useful in guiding the clinical decision for an elective cesarean section. As part of future work, we plan to increase the number of patients in our dataset, allowing us to test models specifically for subgroups of patients affected by diabetes. Supplementary Information Supplementary Material 1. Acknowledgements Not applicable. Disclosure statement The authors declare that they have no conflict of interest in relation to this study. Abbreviations First trimester fasting glucose test Abdominal circumference Analysis of variance Area under of curve of receiver operating characteristic Body mass index Bernoulli Naïve Bayes Balanced random forest Crown-rump length Cross validation Data augmentation Deep neural network Diabetes Mellitus Decision tree Estimated fetal weight Extra trees Fetal medicine foundation Fasting blood glucose FN False negative Fasting plasma glucose False positive False positive rate Gradient boosting Gestational diabetes mellitus Gaussian Naïve Bayes Gestational weeks Hyperglycemia and adverse pregnancy outcome Glycated hemoglobin Head circumference International association of diabetes and pregnancy study groups K-nearest neighbors Low-density lipoprotein cholesterol Large for gestational age Light gradient boosting Logistic regression Machine learning Multi-layer perceptron National institutes of child health and human development Oral glucose tolerance test Plasma protein A Random Forest Sociedad Chilena de pediatría State of the art Support vector machine Type 2 diabetes mellitus True positive rate Umbilical artery pulsatility index Ultrasound Ultrasound date Uterine artery resistance index World health organization Extreme gradient boosting Authors' contributions GC contributed with the machine learning design of the study, statistical analysis, data-cleaning, machine learning models implementation, draft and revisions, MM performed data collection, data analysis, results interpretation and paper edition, AP, MPM and SWL helped with data analysis, clinical interpretation and final draft revision, PE contributed with the machine learning and statistical methods, results and paper draft revisions, MC, SL and MWK performed clinical interpretation and conclusions as well as helped writing the final version of the paper, SEI and CAP contributed with the conception and design of the study, designed the methods, interpreted the results and conclusions from the clinical as well as de machine learning perspectives, wrote and edited the final version of the paper. Funding This work was supported in part by Agencia Nacional de Investigacion y Desarrollo (ANID) under Basal funding for Scientific and Technological Center of Excellence, IMPACT, #FB210024, and FONDECYT 1231675, also by the Dept. of Electrical Engineering, Universidad de Chile, and by the Dept. of Obstetrics and Gynecology, Faculty of Medicine, Universidad de los Andes. Data availability The datasets used in this study are not publicly available due to privacy reasons. The dataset is provided by Hospital Parroquial de San Bernando, access to this data may be provided to qualified researchers upon request and permission of this institution (sillanes@uandes.cl).The code used for the analysis, data cleaning, and model implementation are not available due to proprietary reasons and requires the data to be used. The models were implemented with Python using libraries which are publicly available for anyone that wants to replicate the experiments. Declarations Ethics approval and consent to participate The data usage was approved by the by the institutional review board (IRB) of Hospital Parroquial de San Bernardo, Santiago, Chile, who determined that the research does not involve human subjects because the study uses previously collected, de-identified data. The “Comité de Etica del Hospital Parroquial de San Bernardo” waived the need for informed consent. All methods were carried out in accordance with relevant guidelines and regulations. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Sebastian E. Illanes, Email: sillanes@uandes.cl Claudio A. Perez, Email: clperez@ing.uchile.cl References Associated Data Data Citations Supplementary Materials Supplementary Material 1. Data Availability Statement The datasets used in this study are not publicly available due to privacy reasons. The dataset is provided by Hospital Parroquial de San Bernando, access to this data may be provided to qualified researchers upon request and permission of this institution (sillanes@uandes.cl).The code used for the analysis, data cleaning, and model implementation are not available due to proprietary reasons and requires the data to be used. The models were implemented with Python using libraries which are publicly available for anyone that wants to replicate the experiments.