Early prediction of very and extreme preterm births using a one-class classification framework on electronic health records in UAE Authors: Amir Ahmad, Wasif Khan, Md Mozakkir Ansari, Mohammad M Masud, Nadirah Ghenimi, Luai A Ahmed Published: 2025 ABSTRACT AbstractVery Preterm birth (vPTB) and extreme Preterm birth (xPTB) are the major concerns in maternal and child healthcare and are associated with increased morbidity and mortality. Machine learning methods have traditionally been used to predict preterm births (vPTB and xPTB). However, most medical datasets, including preterm births, are imbalanced in class distribution. Although data-balancing techniques can be employed, complications due to the limited sample size of the minority class are frequently encountered, leading to inconsistent results. This study adopted a novel approach by employing one-class classification (OCC) in conjunction with several strategies to predict instances of vPTB and xPTB within an Emirati pregnant population. We used a well-curated dataset acquired during the first trimester of pregnancy. We employed multiple OCC algorithms and their ensembles involving multiple aggregation strategies to predict vPTB and xPTB in both parous and nulliparous populations. Our approach effectively incorporated only majority class information during training. Our detailed experimental setup demonstrated that the proposed methodology achieved promising performance with a maximum AUC-ROC of 0.823 for the parous population without any explicit modeling of the minority class. Our approach demonstrated robustness and efficacy in identifying at-risk pregnancies within the Emirati population. Our results suggest that one-class classification framework which requires only normal data points for training can be used for early prediction of very preterm and extreme preterm births with reasonable accuracy. In this paper, we applied one-class classification framework only on the Emirati population. Generalizing the proposed approach in this domain requires experimentation on similar datasets from other countries.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-025-29219-4.Keywords: Very preterm birth, Extreme preterm birth, One-class classification, Outlier prediction, Anomaly prediction, Ensemble modelSubject terms: Health care, Medical research Introduction Preterm birth (PTB), which refers to the birth of a baby before completing 37 weeks of gestation, continues to be a serious global issue that contributes to elevated rates of infant mortality worldwide1–3. PTB can be categorized into very preterm birth (vPTB), where the infant is born before 32 weeks of gestation, and extreme preterm birth (xPTB), where the infant is born before 28 weeks of gestation4,5. This phenomenon affects a large number of newborns annually, with approximately 15 million babies being born prematurely, constituting ~ 10% of total global births1. Unfortunately, approximately 1 million premature infants do not survive, highlighting the significance of the problem6. vPTB and xPTB impose substantial emotional and financial burdens on families6. Effective strategies and interventions are urgently needed to determine the complex factors contributing to these adverse pregnancy outcomes. Detecting vPTB and xPTB at an early stage, especially during the first trimester, and implementing preventive measures can help mitigate the occurrence of such adverse outcomes. Machine learning (ML)-based algorithms have shown promising performance in predicting vPTB and xPTB delivery at an early stage. Different types of physiological measurements are used to predict vPTB and xPTB, including ultrasound, electrohysterography (EHG)6, tocodynamometry2, internal uterine pressure3, and fetal fibronectin level7. However, these measurements are either subjective or invasive8. In contrast, an analysis of electronic health records (EHR), which provides a comprehensive overview of a patient’s health, may be a reliable and noninvasive approach for predicting PTB (vPTB and xPTB)9–11. Despite considerable research into predicting adverse pregnancy outcomes such as vPTB and xPTB, data imbalance remains a problem due to the rarity of case-patients. Various data-balancing techniques have been employed to mitigate this challenge, such as ADASYN6,12–14, SMOTE14–21, bootstrapping, applying a steeper penalty on the model, and stratified k-fold cross-validation. However, these traditional techniques can introduce bias and may not yield accurate results20. Furthermore, these traditional oversampling approaches can be ineffective when dealing with mixed data types22,23. Moreover, obtaining enough PTB instances for accurate prediction is challenging and costly24. Conventionally, individual algorithms are used for vPTB and xPTB prediction, such as logistic regression (LR)25, support vector machines (SVM)26,27, random forest (RF)28, and neural networks29. However, combining the prediction of multiple models will lead to forming a collective decision-making mechanism that surpasses the capabilities of a single model. The combined prediction of such algorithms is termed ensemble modeling30. However, the dependence of ensemble models on individual algorithms considerably limits their efficacy. Despite their potential to provide a novel and effective framework, one-class classification (OCC) techniques are rarely used for vPTB and xPTB prediction. OCC modeling identifies instances that deviate significantly from the standard, making it an ideal candidate for handling the skewed distribution of vPTB and xPTB9. OCC can overcome the challenges of class imbalance, mitigate bias, and achieve accurate evaluation and prediction of vPTB and xPTB. Therefore, this study explored the potential of OCC methods for predicting vPTB and xPTB as an alternative to traditional ML algorithms. Early prediction of vPTB and xPTB is crucial; therefore, this study focuses on first-trimester data. Obtaining imaging modalities like ultrasounds is difficult, especially in remote areas; therefore, this study is focused on easy-to-collect tabular medical features (binary, numeric, and categorical). vPTB and xPTB data points are rare or may not be representative of the class during training. Therefore, we utilized OCC methods to train ML models with only one class (term patients) and predict the instances of vPTB and xPTB. The contribution of this study is to provide a comprehensive OCC framework that encompasses various OCC algorithms for predicting vPTB and xPTB in both parous (patients with a history of pregnancies) and nulliparous (first pregnancy) populations. Furthermore, we incorporated an ensemble of multiple OCC models using different aggregation and normalization strategies based on a novel dataset obtained during the first trimester of an Emirati pregnant population in the United Arab Emirates. Literature review Adverse pregnancy outcomes such as vPTB and xPTB are serious issues, and many studies have explored how to minimize the associated risks. Analyses on vPTB and xPTB from the medical perspective are abundant; however, few studies have attempted to predict PTB using ML algorithms. For instance, Mercer et al.31 performed LR analysis of PTB using data on 2,929 women to identify the risk factors and odds ratios for PTB. Lee et al.32 proposed ML methods to predict PTB and its significant risk factors using a dataset from 596 patients. The authors used six different classifiers (ANN, LR, DT, NB, RF, and SVM) for PTB classification and achieved accuracies of 0.911, 0.918, 0.832, 0.111, 0.891, and 0.914, respectively. The authors also identified BMI as the most significant risk factor followed by hypertension, diabetes mellitus, prior cone biopsy, and prior placenta previa for PTB prediction. Tran et al.33 proposed stabilized sparse LR for PTB prediction to identify important risk factors using a dataset from 15,814 women that achieved an AUC-ROC of 0.85 to predict PTB at 34 weeks’ of gestation and 0.79 at 37 weeks. Taha et al.34 identified the factors associated with low birth weight (LBW) and PTB using a dataset of 1,610 mothers in the United Arab Emirates. They revealed that the PTB and LBW ratios based on the sample size were 6.3% and 9.4%, respectively. Moreover, they reported that Arab women had a 2-fold higher risk of having PTB than non-Arab women. Sun et al.9 evaluated the performance of six ML algorithms (RF, SVM, ANN, K-Mean, and Naïve Bayesian) for PTB prediction. They used data from 9,550 pregnant women (PTB: 4775; control: 4775) and constructed an ML model to predict PTB at 20, 22, 24, and 26 weeks of gestation. They revealed that RF exhibited the best performance with an AUC, sensitivity, and specificity of 0.89, 0.75, and 0.88, respectively, at 26 weeks of gestation. Koivu et al.35 evaluated the performance of multiple classifiers such as LR, ANN, and lightgbm for PTB prediction. They used CDC and NYC for external validation, with 10% of the data used for testing. ANN and lightgbm models showed an AUC-ROC value of 0.67 with the CDC dataset and a maximum AUC-ROC value of 0.64 with the NYC dataset. Raja et al.36 evaluated the performance of ML classifiers SVM, RF, DT, and LR for PTB classification on a dataset of 1,300 samples with 35 different features. The authors found that SVM achieved the best performance when the dataset was balanced using the SMOTE technique with an accuracy, sensitivity, and specificity of 0.90, 0.89, and 0.78, respectively. Belaghi et al.24 used multiple ML models (RF, LR, DT, and ANN) to predict PTB in nulliparous women during the first and second trimesters. The authors created a feature set based on a dataset obtained from 112,963 nulliparous women. The sensitivity and specificity of LR in the first trimester were 0.50 and 0.64, respectively, whereas those of RF were 0.29 and 0.84, respectively. For the second trimester, the sensitivity and specificity obtained for RF and ANN were 0.45, 0.94 and 0.62, 0.84, respectively. Belaghi et al.24 performed multivariate statistical analysis for PTB and spontaneous PTB in multiparous and nulliparous women using data from 267,226 births. The AUCs obtained by LR analysis for the first trimester were 0.68 and 0.73 for nulliparous and multiparous women, respectively, whereas those for the second trimester were 0.72 and 0.78, respectively. Another study12 on PTB used EHR data from 25,689 deliveries and trained recurrent neural network (RNN) ensemble models to predict xPTB. The study also identified risk factors associated with xPTB, such as twin pregnancy, short cervical length, hypertensive disorder, systemic lupus erythematosus, and hydroxychloroquine sulfate level. Similarly, Ren et al.14 proposed a PTB screening method that adopted a long short-term memory model based on a time-series analysis of obstetric examination records. Electronic medical record (EMR) data from 5,187 Chinese pregnant women were collected, encompassing over 25,000 obstetric EMRs from the early trimester to 28 weeks of gestation. The proposed method achieved an AUC-ROC value of 0.65, and the important risk factors identified were blood pressure, blood glucose, lipids, uric acid, and other metabolic factors. Similarly, another study done by16 evaluated a PTB prediction model trained on data from 4,332 pregnancies in 3,611 patients. They identified gestational age, dilation, rupture of membranes, time since admission, and heart rate as important risk factors associated with PTB delivery16. Although extensive research has been conducted to predict vPTB and xPTB, no studies have focused on OCC algorithms. OCC algorithms could provide valuable tools for improving prenatal care and intervention strategies. Methods This study evaluated the effectiveness of the OCC and ensemble models in accurately identifying vPTB and xPTB outcomes as outliers. Incorporating OCC algorithms only requires normal data points during the training phase; hence, no minority samples are needed. During the testing phase, the maximum number of positive (minority) data points is used, which helps obtain more reliable results reflective of real-world scenarios. A flowchart of the proposed methodology is presented in Fig. 1. Furthermore, Algorithm 1 details the process for predicting vPTB and xPTB using outlier detection and ensemble methods. Our method consists of several modules. The first module involves obtaining a well-curated dataset of vPTB and xPTB cases, wherein the selection of risk factors was carefully based on medical recommendations. All risk factors considered in this study were observed in the first trimester. In the second module, we trained multiple outlier detection algorithms using different training–testing strategies, including ensemble techniques. We aimed to enhance the accuracy and reliability of outlier detection by using multiple algorithms and combining their results through multiple ensemble strategies. The third module involves an extensive assessment using various evaluation metrics to evaluate the effectiveness of the proposed methodology. Through rigorous evaluation, we may gain insights into the strengths and limitations of the methodology, enabling us to further refine and enhance its performance. Proposed methodology for predicting vPTB and xPTB outcomes using OCC and ensemble models. vPTB and xPTB prediction algorithm. Dataset Ethics statement This analysis is based on data from an ongoing study37 in Al Ain, UAE. The study was conducted in accordance with the Declaration of Helsinki and was approved by the United Arab Emirates University Human Research Ethics Committee (ERH-2017-5512) and the Abu Dhabi Health Research and Technology Ethics Committee (DOH/CVDC/2022/72). All the participants in the study signed the informed consent form. Data preparation and preprocessing The dataset used in this study involved an Emirati pregnant population. Risk factors were obtained from EHRs and questionnaires administered during pregnancy. These risk factors were carefully selected based on medical recommendations encompassing various aspects, including maternal health, lifestyle factors, medical history, and socioeconomic factors38,39. The details of the risk factors utilized in this work are shown in Table 1. The details of patient demographics and baseline characteristics are reported in60. By incorporating these risk factors, the dataset captured a comprehensive view of potential predictors for PTB. vPTB was defined as delivery before 32 weeks of gestation, whereas xPTB was defined as delivery before 28 weeks of gestation. The population included both parous and nulliparous women. The distribution of patients is shown in Table 2. Table 1The list of features used in this study. Bold features show feature which are applicable to Parous population only. Very preterm and extreme preterm births are the target variables.S. NoFeature nameTypeS. NoFeature nameType1AgeContinuous variable12Physical activity (PA) during pregnancyCategories (0 = Never, 1 = 1–2 times/week, 2 = 3–5 times/week, 3 = daily)2Inter-pregnancy interval (days)13Physical activity (PA) before pregnancy3GravidaCount (Number of pregnancies including this one)14Worry about upcoming childbirthBinary (0 = No, 1 = Yes)4ParityCount (Number of previous viable delivery)15History of previous preterm birth5Planning status for this pregnancyBinary (0 = No, 1 = Yes)16History of previous cesarian delivery (CS)6Previous infertility treatment17History of previous pregnancy loss7Consanguinity18Pre-existing hypertension8Employment19Rh antibodiesBinary (0 = Negative, 1 = Positive)9Gestational diabetes mellitus20HousingBinary (0 = Rent, 1 = Owned)10Passive Smoking21Education levelBinary (0= ≤ _grade 12, 1 = > grade 12)11Pre-existing diabetes mellitus The list of features used in this study. Bold features show feature which are applicable to Parous population only. Very preterm and extreme preterm births are the target variables. Table 2Datasets used in this study. Preterm samples include both very preterm and extreme preterm birth samples.Dataset nameTarget variableTerm samplesPreterm samplesTotal samplesPreterm samples (%)NulliparousvPTB779228012.75ParousvPTB26733527081.29NulliparousxPTB79568010.75ParousxPTB26961027060.37 Datasets used in this study. Preterm samples include both very preterm and extreme preterm birth samples. OCC algorithms Multiple OCC algorithms were used to build a classification model for predicting the likelihood of vPTB and xPTB, which are briefly explained in Table 3. Table 3One-class classifiers used in this study. All algorithms were evaluated using default hyperparameters.AlgorithmBrief detailsAdvantagesDisadvantagesGMM40Gaussian mixture models represent data as a combination of Gaussian distributions.Flexible modeling of complex data distributions. Can capture multimodal data.Sensitive to the choice of the number of components. Computationally expensive for high-dimensional data.KNN41The k-nearest neighbors algorithm determines the proximity to the k-nearest neighbors in feature space.Simple and intuitive approach. Robust against noise and outliers.Sensitivity to the choice of k. Computationally expensive for large datasets.IForest (IF)42Isolation forest divides data randomly to isolate anomalies efficiently.Effective for high-dimensional data. Can handle large datasets efficiently.May produce false positives for clustered data. Sensitivity to the subsampling size and contamination rate.Autoencoder (AE)43AE is a neural network that learns to reconstruct input data.Captures nonlinear relationships in the data. Can handle complex feature interactions.May overfit normal patterns and fail to generalize well. Requires careful selection of the network architecture.SOS44The stochastic outlier selection scores the data points based on the outlying properties.Rapid and scalable algorithm. Suitable for large datasets.May struggle with complex data distributions. Sensitivity to parameter settings.KDE45Kernel density estimation determines the probability density function of the data.Effective for capturing complex data distributions. Provides a continuous anomaly score.Sensitive to the choice of kernel bandwidth. Computationally expensive for large datasets.LUNAR46Learnable unified neighborhood-based anomaly ranking combines multiple anomaly detection methods using statistical and distance-based rules.Robust against different types of anomalies. Interpretable results using a rule-based approach.Requires domain knowledge for rule selection. Performance may vary depending on the dataset and ruleset.OCSVM47A one-class support vector machine that separates data points from the origin in a high-dimensional feature space.Effective for high-dimensional data. Robust against outliers.Sensitive to the choice of kernel and regularization parameters. Computationally expensive for large datasets.KPCA48Kernel principal component analysis applies kernel tricks to PCA for anomaly detection.Nonlinear mapping for capturing complex relationships. Effective for high-dimensional data.Sensitivity to the choice of kernel and hyperparameters. Requires careful interpretation of transformed features.DeepSVDD49Deep support vector data description learns a compact representation of normal instances using deep neural networks.Captures complex patterns and relationships in the data. Robust against noise and outliers.Requires careful selection of the network architecture and hyperparameters. Computationally expensive for large datasets.VAE50VAE is auto encoder model that learns the underlying distribution of the majority class.Captures complex data distributions.May overfit normal patterns and fail to generalize well. Computationally expensive for large datasets. One-class classifiers used in this study. All algorithms were evaluated using default hyperparameters. Ensemble strategies We implemented three different ensemble strategies, which are described in the subsequent sections. Min–Max and Z-score normalization-based ensemble models for outlier detection An OCC method generates an outlier score (OS) for a data point which represents how anomalous the point is, for a normal point OS score is low whereas for an outlier point OS is large. These outlier scores (OSs) are in different scales for different OCC models. In our ensemble-based outlier detection approach, we applied two different normalization techniques: min–max normalization-based and Z-score–based ensembles. In the normalization-based ensemble, we aimed to standardize and combine the outlier scores (OSs) from multiple models in the ensemble using the OS obtained from the training data. We employed three strategies: minimum, maximum, and median. For each strategy, we computed the corresponding OS (minimum, maximum, or median) among the ensemble models for each instance in the testing set. To normalize the OSs, we used the minimum and maximum values obtained from the training set. For each instance in the testing set, we computed the normalized OS by subtracting the minimum OS and dividing it by the difference between the maximum and minimum OSs from the training set. We repeated this process for the maximum and median strategies to obtain three sets of normalized OSs. In the Z-score–based ensemble, we employed the Z-score normalization technique. We calculated the mean and standard deviation of the OSs from the training set. For each instance in the testing set, we computed the Z-score by subtracting the mean of the training OSs and dividing it by the standard deviation of the training OSs. By incorporating these normalized and standardized OSs into our ensemble-based outlier detection approach, we achieved a common scale and consistent interpretation of the OSs across multiple models. The output for a sample in ensemble model predictions can be aggregated using the strategies shown in Table 4. The ensemble models are briefly described below. Table 4Aggregation strategies for the ensembles used in the study.Aggregation methodStrategyExplanationEquationMeanMeanThe mean outlier score (OS) from multiple models. MedianMedian (1)The median OS is directly calculated. Median (3)The median OS along with the two nearest values (one on each side) is determined, and the mean of the three values is computed. MaxMax (1)The largest OS is directly calculated. Max (3)The three largest OSs are identified, and their mean is calculated for a balanced and optimistic outcome. MinMin (1)The smallest OS is directly computed. Min (3)The three smallest OSs are identified, and their mean is calculated for a more balanced outcome and reduced uncertainty. Aggregation strategies for the ensembles used in the study. Ensemble with mean aggregation The ensemble with mean aggregation calculates the average of the OS from multiple models to obtain the aggregated OS for a given sample as follows: Ensemble with sorted median aggregation This model first sorts the OS obtained from multiple models in ascending order and then calculates the median of the sorted scores to obtain the aggregated OS for a given sample . We explored two distinct strategies for performing median aggregation: computing a single median value, referred to as “median (1),” and determining three median values or “median (3).” For median (1), the median value of the OS was directly calculated. In contrast, for median (3), we determined the median value of the OS along with the two nearest values (one on each side) and computed their average. Ensemble with max aggregation We explored two strategies for maximum aggregation: computing the maximum value or “max (1)” and determining the three maximum values or “max (3).” For max (1), the largest OS value was directly calculated. Conversely, for “max (3),” we identified the three largest OS values and calculated their average, offering a more balanced and optimistic outcome. Ensemble with min aggregation In this model, the minimum OS was selected from the OSs obtained from different models. We explored two strategies for minimum aggregation: calculating the minimum value or “min (1)” and determining the three minimum values “min (3).” For “min (1),” the smallest OS value was directly computed. Conversely, for “min (3),” we identified the three smallest OS values and calculated their average, which provided a more balanced outcome and reduced uncertainty. The ensemble predictions obtained through these aggregation strategies allow for a more robust and comprehensive data analysis, potentially improving the accuracy and reliability of the final prediction for PTB51. Locally selective combination of parallel outlier ensembles We also used the locally selective combination of parallel outlier ensembles (LSCP) method from PyOD53 for ensemble-based outlier detection. LSCP is a powerful algorithm that combines the OSs obtained from multiple base models by selectively considering the local characteristics of the data. We trained various base models using algorithms such as GMM, KNN, IF, AE, SOS, KDE, LUNAR, OCSVM, KPCA, DeepSVDD, and VAE and obtained their OSs for the testing instances. The LSCP algorithm then selectively combines these OSs and assigns weights based on performance in local regions to compute the ensemble OS for each instance. This adaptive approach allowed us to effectively handle variations in the data and achieve improved outlier detection performance. By leveraging the diversity of base models and considering local patterns, the LSCP ensemble method enhanced the accuracy and robustness of outlier detection. Stacked ensemble In the stacked ensemble model52 of OCC models, we use two layers of OCC models. One OCC model is used in the second layer and the rest in the first layer. The training data is fed to each OCC model of the first layer of the stacked ensemble. We train OCC models in the first layer. These trained OCC models are used to generate outlier scores of each data point of the training data. The outputs of the trained OCC models in the first layer are used as the input to train the OCC model in the second layer of the stacked ensemble. If there are m OCC models, the first layer has m-1 trained OCC models. For each data point in the training data, (m-1) outliers scores are created by m-1 trained OCC models. These scores are the new features for that data point. If we have n data points in a training dataset, the first layer generates a dataset of size nx(m-1). This dataset is used to train OCC model in the second layer. When a testing point arrives, it will be fed to m-1 trained OCC models of the first layer, resulting in m-1 dimension feature vector for this testing point. These new features will be fed to the trained OCC model of the second layer to get the final outlier score of the testing point. The equation for the stacked ensemble is as follows: where the OSs () represent the predictions made by each individual model. is the final OS of the data point. Performance metrics The performance of the methods was evaluated using AUC-ROC (Area Under Curve for the Receiver Operating Characteristic Curve) and AUC-PRC (Area Under Curve for the Precision Recall Curve)54. These metrics provide a focused and effective assessment, considering the class distribution of our dataset. The AUC-ROC quantifies the ability of the model to distinguish between positive and negative instances, whereas the AUC-PRC value emphasizes precision and recall trade-offs. The value of AUC-ROC by a random model is 0.50, whereas for AUC-PRC value, for a random model is the ratio of minority samples to total number of samples and minority samples. The baseline AUC-PRC value can be calculated using the equation below: The baseline (random) AUC-PRC values for each dataset are show in Table 5 below. Table 5Baseline AUC-PRC value. Values exceeding the baseline indicate improved algorithm performance.DatasetTarget variableBaseline AUC-PRC valueNulliparousvPTB0.0535Parous Table S.1vPTB0.0255NulliparousxPTB0.0149Parous Table S.1xPTB0.0074 Baseline AUC-PRC value. Values exceeding the baseline indicate improved algorithm performance. Results The experimental setup for vPTB and xPTB prediction is depicted in Fig. 2. Initially, we divided the dataset into positive (anomalies) and normal class samples. Overall, 50% of the normal class samples were used for training via various anomaly detection algorithms. To evaluate the performance of these algorithms, testing was conducted using all positive samples and the remaining 50% of the normal samples. By including a portion of the majority class samples in the testing set, the model’s performance was assessed in a real-world scenario where vPTB or xPTB is relatively rare. Schematic diagram of the experimental setup. Experiments using the two-fold cross-validation technique were repeated five times with only majority samples during training. The average results of these experiments provide a comprehensive overview of the performance of individual algorithms and their ensembles. Moreover, the differences in features between parous and nulliparous women were considered. For instance, features such as pregnancy history or pregnancy loss were not applicable to nulliparous women. We named the nulliparous women dataset and for predicting vPTB and xPTB, respectively. Furthermore, to comprehensively investigate the influence of these features on the parous group, we established three different feature sets for predicting vPTB and xPTB. The first dataset was a comprehensive feature set that encompassed all attributes (Table 1). We named this dataset and . The second dataset contained features similar to those of nulliparous individuals (Under “feature similar to,” Table S.2) and was named and . Finally, a feature subset unique to the parous cohort and thus absent in the nulliparous population was named and . Details regarding the dataset and experiments are shown in Table 6. To implement the anomaly detection algorithms, the experiments were performed using Python 3.8 via a Dell PC with 8 GB RAM and data from the PyOD library53. Table 6Details regarding the datasets used in the experiments.Dataset nameDescriptionTarget variableResult table Nulliparous population with all featuresvPTBTables 5, 6 and 7 Similar features in parous and nulliparous populationsvPTBTables 8, 9 and 10 Parous population with all featuresvPTBTables S.2–S.4 Parous population features that differ from those in parous populationvPTBTables S.5–S.7 Nulliparous population with all featuresxPTBTables 11, 12 and 13 Similar features in parous and nulliparous populationsxPTBTables 15 and 16 Parous population with all featuresxPTBTables S.8–S.10 Parous population features that differ from those in the parous populationxPTBTables S.11–S.13 Details regarding the datasets used in the experiments. vPTB prediction for the nulliparous population The results of the analyses of the data regarding nulliparous women with all features () are summarized in Table 7. The individual algorithms exhibited similar AUC-ROC scores with a maximum AUC-ROC value of 0.527 achieved with GMM. SOS-based outlier detection performed the worst with an AUC-ROC score of 0.483. The DeepSVD algorithm obtained the best AUC-PRC value at 0.073. However, its corresponding AUC-ROC score was relatively low at 0.504. Table 7Experimental results obtained using one-class classification models for the nulliparous population. Baseline AUC-ROC is 0.50 while AUC-PRC is 0.0535.Performance metricsGMMKNNIFAESOSKDELUNAROCSVMKPCADeepSVDDVAEAUC-ROC 0.527 0.5140.5080.5140.4830.5100.5150.4970.5040.5040.518AUC-PRC0.0620.0590.0630.0610.0570.0590.0650.0590.058 0.073 0.063Best values are in bold. Experimental results obtained using one-class classification models for the nulliparous population. Baseline AUC-ROC is 0.50 while AUC-PRC is 0.0535. Best values are in bold. The results of the stacking approach are shown in Table 8. The best AUC-ROC value of 0.541with a AUC-PRC value of 0.063 was obtained with the KPCA, OCSVM, and KNN models. The results of the ensemble strategies employed in the study are shown in Table 9. The best overall performance was achieved using the ensemble method with the min (1) OS strategy. By employing min-max normalization, this ensemble attained an AUC-ROC value of 0.540 and a AUC-PRC value of 0.070. As a sample, we show the ROC and PRC curves for Table 9 in Supplementary Figures (S.1-S.4) because generating these figures for all tables results could be difficult to interpret. Table 8Experimental results using stacked ensembles for the nulliparous population. Baseline AUC-ROC is 0.50 while AUC-PRC is 0.0535.Performance metricsGMMKNNIFAESOSKDELUNAROCSVMKPCADeepSVDDVAEAUC-ROC0.516 0.541 0.4860.4760.5230.5280.519 0.541 0.541 0.5160.477AUC-PRC0.0700.0630.0570.0650.0580.0620.0690.0630.063 0.073 0.066Best values are in bold. Experimental results using stacked ensembles for the nulliparous population. Baseline AUC-ROC is 0.50 while AUC-PRC is 0.0535. Best values are in bold. Table 9Results obtained using ensemble models via min–max and Z-score–based normalization for nulliparous population. Baseline AUC-ROC is 0.50 while AUC-PRC is 0.0535.MetricsMeanMedian (1)Median (3)Max (1)Max (3)Min (1)Min (3)LSCPNormalization AUC-ROC0.5190.5140.5140.5220.524 0.540 0.5300.515 AUC-PRC0.0680.0600.0600.070 0.071 0.0700.0680.064Z-score AUC-ROC0.5200.5150.5150.5220.523 0.537 0.5260.515 AUC-PRC 0.071 0.0610.0610.0690.0690.0660.0640.064Best values are in bold. Results obtained using ensemble models via min–max and Z-score–based normalization for nulliparous population. Baseline AUC-ROC is 0.50 while AUC-PRC is 0.0535. Best values are in bold. In min–max normalization, the best AUC-ROC value of 0.540 and AUC-PRC value of 0.070 were achieved using the min (1) ensemble strategy, thereby demonstrating its good precision in identifying outliers. In Z-score normalization, the best AUC-ROC value of 0.537 was obtained using the min (1) ensemble strategy. However, the mean ensemble strategy yielded the highest AUC-PRC value of 0.071. As shown in Table 9, min–max and Z-score–based normalization using the min (3) ensemble strategy produced the best AUC-ROC values of 0.530 and 0.525, respectively. The mean (3) and max (3) ensemble strategies under both normalization techniques achieved the highest AUC-PRC value of 0.071. Overall, the best performance was achieved via min–max normalization using the min (1) ensemble strategy. vPTB prediction in the Parous population The results of the analysis of the parous population (that included the features similar to those of the nulliparous population are shown in Table 10. Among the individual algorithms, the VAE achieved the best AUC-ROC value of 0.620; however, its AUC-PRC value was 0.037. The highest AUC-PRC value of 0.043 was achieved with OCSVM; however, its AUC-ROC was 0.593. The performance of models using the stacking approach was comparable, with the best AUC-ROC value of 0.590 and best AUC-PRC value of 0.041 achieved by LUNAR (Table 11). Table 10Experimental results obtained using one-class classification models for Parous population. Baseline AUC-ROC is 0.50 while AUC-PRC is 0.0255.MetricsGMMKNNIFAESOSKDELUNAROCSVMKPCADeepSVDDVAEAUC-ROC0.6180.5950.5940.6150.5530.6060.5730.5930.6110.527 0.620 AUC-PRC0.0380.0370.0330.0360.0310.0400.0330.0430.0410.0300.037Best values are in bold. Experimental results obtained using one-class classification models for Parous population. Baseline AUC-ROC is 0.50 while AUC-PRC is 0.0255. Best values are in bold. Table 11Experimental results using stacked ensembles for Parous population. Baseline AUC-ROC is 0.50 while AUC-PRC is 0.0255.MetricsGMMKNNIFAESOSKDELUNAROCSVMKPCADeepSVDDVAEAUC-ROC0.5660.5850.5630.5250.4550.585 0.590 0.5850.5850.5730.547AUC-PRC0.0360.0380.0350.0300.024 0.041 0.0390.0380.0380.0360.033Best values are in bold. Experimental results using stacked ensembles for Parous population. Baseline AUC-ROC is 0.50 while AUC-PRC is 0.0255. Best values are in bold. The results of the ensemble techniques via different normalization strategies are shown in Table 12. For the normalization-based ensemble, the AUC-ROC values ranged from 0.573 to 0.621, whereas the AUC-PRC values ranged from 0.032 to 0.038. The median (1) and median (3) ensemble strategies achieved the best performance with an AUC-ROC value of 0.621 and AUC-PRC value of 0.038. The Z-score–based normalization using the mean ensemble strategy achieved the best performance with an AUC-ROC value of 0.617 and AUC-PRC value of 0.038. Table 12Experimental results obtained using ensemble models for the nulliparous population. Baseline AUC-ROC is 0.50 while AUC-PRC is 0.0255.MetricsMeanMedian (1)Median (3)Max (1)Max (3)Min (1)Min (3)LSCPNormalization AUC-ROC0.613 0.621 0.6210.5770.5940.5730.6080.612 AUC-PRC0.037 0.038 0.0380.0320.0340.0350.0380.038Z-score AUC-ROC 0.617 0.6130.6130.6060.6130.5720.6050.612 AUC-PRC0.0380.0370.0370.037 0.040 0.0350.0360.038Best values are in bold. Experimental results obtained using ensemble models for the nulliparous population. Baseline AUC-ROC is 0.50 while AUC-PRC is 0.0255. Best values are in bold. The results of analysis of data regarding parous women with all features ( are shown in Tables S.2–S.4. Among the individual algorithms (Table S.2), KPCA achieved the highest AUC-ROC value of 0.630, with a relatively low AUC-PRC value of 0.047. The KNN algorithm performed well with an AUC-ROC value of 0.627 and a higher AUC-PRC value of 0.053. In contrast, DeepSVDD showed the lowest performance with an AUC-ROC value of 0.507 and a AUC-PRC value of 0.032. Among the stacking ensemble models (Table S.3), the AUC-ROC values ranged from 0.464 to 0.624. LUNAR achieved the highest AUC-ROC among the ensemble models; however, its corresponding AUC-PRC value was lower than that of other models. Interestingly, DeepSVDD achieved the best AUC-PRC value of 0.051 while maintaining a comparable AUC-ROC value of 0.620. The results obtained for the parous population using different ensemble strategies and normalization techniques are presented in Table S.4. For the min–max normalization-based ensemble, the AUC-ROC values ranged from 0.587 to 0.621. The mean ensemble strategy achieved the highest AUC-ROC value of 0.621, indicating better overall performance in outlier detection. Similarly, the AUC-PRC values ranged from 0.039 to 0.045, with the max (1) ensemble strategy achieving the highest AUC-PRC value of 0.048. For Z-score–based ensemble, the AUC-ROC values ranged from 0.564 to 0.624, whereas the AUC-PRC values ranged from 0.034 to 0.047. The max (3) ensemble strategy achieved the highest AUC-ROC value of 0.624 and AUC-PRC value of 0.047 using min–max normalization. The results obtained for the parous population using only non-matching features ( are presented in Table S.5. Among the individual algorithms, the Isolation Forest algorithm achieved the highest AUC-ROC value of 0.663 but a lower corresponding AUC-PRC value of 0.054. The DeepSVDD algorithm achieved the lowest AUC-ROC value of 0.483 and a AUC-PRC value of 0.033. For the stacking ensemble models (Table S.6), the AUC-ROC values ranged from 0.473 to 0.618. The Isolation Forest algorithm performed the best among the ensemble models, achieving an AUC-ROC value of 0.618 and AUC-PRC value of 0.049. Individual algorithms performed better than the stacking ensemble in analyzing non-matching features. For the min–max normalization-based ensemble (Table S.7), the AUC-ROC values ranged from 0.521 to 0.651. The median (1) ensemble strategy achieved the highest AUC-ROC value of 0.651, indicating better overall performance in outlier detection. However, the associated AUC-PRC value was 0.045, suggesting a trade-off between recall and precision. The min (1) ensemble strategy achieved the lowest AUC-ROC value of 0.521, indicating its relatively poorer performance in outlier detection. For the Z-score–based normalization, the AUC-ROC values ranged from 0.568 to 0.621. The median (1) ensemble strategy achieved the highest AUC-ROC value of 0.621, whereas its AUC-PRC value was 0.042. The min (1) ensemble strategy achieved the lowest AUC-ROC value of 0.568. xPTB prediction for the nulliparous population The results of xPTB prediction in nulliparous women ( using various individual OCC algorithms and ensemble techniques (Table 13) indicate that the best AUC-ROC value of 0.673 was achieved with OCSVM but with a relatively low AUC-PRC value of 0.027. The best AUC-PRC value of 0.037 was achieved with IF with an AUC-ROC value of 0.663. Analysis using the stacking ensemble (Table 14) was not better than the individual algorithms with AUC-ROC values that ranged from 0.416 to 0.669 and AUC-PRC values that ranged from 0.013 to 0.024. Table 13Results obtained using OCC models for the nulliparous population. Baseline AUC-ROC is 0.50 while AUC-PRC is 0.0149.MetricsGMMKNNIFAESOSKDELUNAROCSVMKPCADeepSVDDVAEAUC-ROC0.6600.6630.6630.6630.6340.6560.639 0.673 0.6450.6000.672AUC-PRC0.0280.0300.0370.0260.0280.029 0.039 0.0270.0270.0240.028Best values are in bold. Results obtained using OCC models for the nulliparous population. Baseline AUC-ROC is 0.50 while AUC-PRC is 0.0149. Best values are in bold. Table 14Results obtained using stacked ensembles for the nulliparous population. Baseline AUC-ROC is 0.50 while AUC-PRC is 0.0149.MetricsGMMKNNIFAESOSKDELUNAROCSVMKPCADeep SVDDVAEAUC-ROC0.6030.6550.6540.5940.4160.6690.5710.655 0.655 0.5260.639AUC-PRC0.0210.023 0.024 0.0210.0130.0220.0200.0230.0230.0230.024Best values are in bold. Results obtained using stacked ensembles for the nulliparous population. Baseline AUC-ROC is 0.50 while AUC-PRC is 0.0149. Best values are in bold. The results of xPTB prediction in parous women using different normalization techniques and ensemble strategies are shown in Table 14. For the ensemble strategies shown in Table 15, the AUC-ROC values ranged from 0.609 to 0.663, whereas the AUC-PRC values ranged from 0.023 to 0.041. The normalization ensemble with the max (1) aggregation strategy achieved the highest AUC-ROC, whereas the Z-score ensemble with the median (1) strategy achieved the best AUC-PRC value of 0.041. Table 15Results obtained using ensemble models for the nulliparous population. Baseline AUC-ROC is 0.50 while AUC-PRC is 0.0149.MetricsMeanMedian (1)Median (3)Max (1)Max (3)Min (1)Min (3)LSCPNormalization AUC-ROC0.6570.6530.653 0.662 0.6570.6240.6470.659 AUC-PRC0.0290.0270.0270.0300.031 0.038 0.0310.028Z-score AUC-ROC0.658 0.664 0.6640.6420.6630.6090.6620.659 AUC-PRC0.0280.0300.0300.0230.0250.038 0.041 0.028Best values are in bold. Results obtained using ensemble models for the nulliparous population. Baseline AUC-ROC is 0.50 while AUC-PRC is 0.0149. Best values are in bold. xPTB prediction for the Parous population The results of the analyses including the features similar to those of the nulliparous population are shown in Table 16. Among the individual algorithms, the KNN achieved the best AUC-ROC at 0.614 with a AUC-PRC value of 0.010. In the stacking approach represented in Table 17, DeepSVDD showed the best performance with AUC-ROC and AUC-PRC value of 0.569 and 0.010, respectively. 0.041. Table 16Results obtained using OCC models for the Parous population. Baseline AUC-ROC is 0.50 while AUC-PRC is 0.0074.Performance metricsGMMKNNIFAESOSKDELUNAROCSVMKPCADeepSVDDVAEAUC-ROC0.575 0.614 0.5870.5690.5470.5980.5970.4870.5560.4810.577AUC-PRC0.009 0.010 0.0100.0090.0090.0090.0100.0080.0080.0080.009Best values are in bold. Results obtained using OCC models for the Parous population. Baseline AUC-ROC is 0.50 while AUC-PRC is 0.0074. Best values are in bold. Table 17Results obtained using stacked ensembles for the Parous population. Baseline AUC-ROC is 0.50 while AUC-PRC is 0.0074.Performance metricsGMMKNNIFAESOSKDELUNAROCSVMKPCADeepSVDDVAEAUC-ROC0.4240.5260.4970.4480.4330.5210.5240.5260.526 0.569 0.467AUC-PRC0.0070.0080.0080.0070.0070.0080.0080.0080.008 0.010 0.007Best values are in bold. Results obtained using stacked ensembles for the Parous population. Baseline AUC-ROC is 0.50 while AUC-PRC is 0.0074. Best values are in bold. The results of the ensemble techniques with different normalization strategies are shown in Table 18. For the min–max normalization-based ensemble, the best AUC-ROC and AUC-PRC value of 0.586 and 0.009, respectively, was achieved with the mean and max (3) ensemble strategy. Z-score–based normalization also achieved a comparable performance. Table 18Results obtained using ensemble models for the Parous population. Baseline AUC-ROC is 0.50 while AUC-PRC is 0.0074.MetricsMeanMedian (1)Median (3)Max (1)Max (3)Min (1)Min (3)LSCPNormalization AUC-ROC0.5860.5780.5780.567 0.586 0.5620.5690.574 AUC-PRC 0.009 0.0090.0090.0090.0090.0090.0080.009Z-score AUC-ROC 0.582 0.5760.5760.5560.5770.5630.5820.574 AUC-PRC 0.009 0.0090.0090.0080.0090.0080.0090.009Best values are in bold. Results obtained using ensemble models for the Parous population. Baseline AUC-ROC is 0.50 while AUC-PRC is 0.0074. Best values are in bold. The results of xPTB prediction using individual anomaly detection algorithms and ensemble stacking in parous women with all features ( are shown in Table S.8. Among the individual algorithms, IF achieved the highest AUC-ROC value of 0.675, whereas DeepSVDD obtained the lowest AUC-ROC value of 0.496. KPCA showed the highest AUC-PRC value of 0.053; however, the associated AUC-ROC value was only 0.497. For the ensemble stacking approach (Table S.9), the AUC-ROC values ranged from 0.421 to 0.630, whereas the AUC-PRC values ranged from 0.006 to 0.021. The results of xPTB prediction in parous women using ensemble strategies via different normalization techniques are shown in Table S.10. Among the ensemble strategies, the min (3) approach using min–max normalization achieved the highest AUC-ROC value of 0.645 and a AUC-PRC value of 0.014. Overall, the individual algorithms performed better than the stacking and ensemble strategies when all features were included in the analyses. Tables S.11 and S.12 show the results of xPTB prediction for using different individual algorithms and their stacking ensembles when only non-matching features were considered. As shown in Table S.12, compared with individual algorithms (Table S11), the stacking approach using IF achieved the best performance with AUC-ROC and AUC-PRC values of 0.823 and 0.032, respectively. Among the normalization models with their various ensemble strategies (Table S.13), the best performance was achieved using the median (3) ensemble strategy via min–max normalization with AUC-ROC and AUC-PRC values of 0.747 and 0.022, respectively. Discussion The results of this study can be briefly summarized in the Tables 19, 20, 21 and 22 below. Table 19 shows that OCC methods performed better than random guessing, as most of the models surpassed the default random guessing threshold (AUC-ROC = 0.50). Furthermore, the best OCC models differed across the datasets. For example, the KNN model in the stacked ensemble category outperformed other models with an AUC-ROC value of 0.541, whereas VAE was the best model for the parous ( dataset with an AUC-ROC value of 0.620. Similarly, for , LUNAR and ensemble using the max (3) strategy achieved the best performance with an AUC-ROC value of 0.624. For , IF achieved the best performance with an AUC-ROC value of 0.663. Finally, the ensemble methods clearly demonstrate potential in enhancing OCC model performance, outperforming only in the nulliparous population. Ensemble models achieved higher AUC-ROC values than random guessing (AUC-ROC = 0.50) across all datasets. Table 19Results of AUC-ROC for the vPTB dataset.DataModel typeNumber of models/combinationNumber/% of models that performed better than random guessing (AUC = 0.50)Best model/strategyValueWhether any ensemble model is better than the best single modelNulliparousSingle119GMM0.527YesStacked ensembles118KNN0.541Ensemble with combinations1616Min (1) Min–max0.540Parous(D1 vPTB)Single1111VAE0.620NoStacked ensembles1110LUNAR0.590Ensemble with combinations1616Mean0.617Parous Single1111KPCA0.63NoStacked ensembles1110LUNAR0.624Ensemble with combinations1616Max (3)0.624Parous Single1110IF0.663NoStacked ensembles1110IF0.618Ensemble with combinations1616Median (1)0.651 Results of AUC-ROC for the vPTB dataset. Parous (D1 vPTB) Parous Parous Table 20Results of AUC-PRC for the vPTB.dataset.Data (AUC-PRC)Model typeNumber of models/combinationsDefault AUC-PRC valueNumber/% of models that performed better than random guessing (PR value)Best model/ strategyValueWhether any ensemble model is better than the best single modelNulliparousSingle110.053511DeepSVD0.073NoStacked ensembles1111DeepSVD0.073Ensemble with combinations1616Mean, max (3)0.071ParousSingle110.025511OCSVM0.043NoStacked ensembles1111KDE0.041Ensemble with combinations1616Max (3)0.040Parous Single110.025511KNN0.053NoStacked ensembles1111DeepSVD0.051Ensemble with combinations1616Max (1)Min–max0.48Parous Single110.025511KPCA0.061NoStacked ensembles1111IF0.049Ensemble with combinations1616Mean (1), (3)Min–max0.045 Results of AUC-PRC for the vPTB.dataset. Parous Max (1) Min–max Parous Mean (1), (3) Min–max Table 21Results of AUC-ROC for the xPTB dataset.Data (xPTB)Model typeNumber of models/combinationNumber/% of models that performed better than random guessing (AUC = 0.50)Best model/strategyValueWhether any ensemble model is better than the best single modelNulliparousSingle1111OCSVM0.673NoStacked ensembles1110KPCA0.655Ensemble with combinations1616Median (1)Z-score0.664Parous(D1 xPTB)Single119KNN0.614NoStacked ensembles116DeepSVD0.569Ensemble with combinations1616Mean, max (3)0.586Parous Single119IF0.675NoStacked ensembles1110IF0.630Ensemble with combinations1616Min (3) min–max0.645Parous Single1111IF0.787YesStacked ensembles1111IF0.823Ensemble with combinations1616Median (3)0.747 Results of AUC-ROC for the xPTB dataset. Median (1) Z-score Parous (D1 xPTB) Parous Parous Table 22Summary of AUC-PRC values for predicting xPTB.DataModel typeNumber of models/ combinationsDefault AUC-PRC valueNumber/% of models that performed better than random guessing (PR value)Best model/ strategyValueWhether any ensemble model is better than the best single modelNulliparousSingle110.014911LUNAR0.039YesStacked ensembles1110IF0.024Ensemble with combinations1616Min (3), Z-score0.041ParousSingle110.007411KNN, IF, LUNAR0.010NoStacked ensembles117DeepSVD0.010Ensemble with combinations1616All strategies0.009Parous Single110.007411kpca0.053NoStacked ensembles1110GMM, KNN, AE, KDE, LUNAE, ocsvm, kpca0.021Ensemble with combinations16Min (1), Z-score0.024Parous Single110.007411KPCA0.118NoStacked ensembles1111IF0.032Ensemble with combinations1616Mean (min–max, Z-score)0.024 Summary of AUC-PRC values for predicting xPTB. Parous Parous The AUC-PR values for vPTB prediction are shown in Table 20. All models achieved better AUC-PRC values than random guessing. Furthermore, none of the ensemble models, whether stacked ensembles or ensembles with combinations, could manage to surpass the AUC-PRC values achieved by the best individual model in any of the datasets. For xPTB prediction, the AUC-ROC value of all models (single, stacked ensembles, and ensembles with combination) shown in Table 21 performed better than the random guess (AUC = 0.50), demonstrating that the OCC framework has superior anomaly detection performance. Furthermore, no single OCC algorithm obtained superior results for all datasets. Finally, the ensemble models did not perform better than any single OCC algorithm. The results of AUC-PRC presented in Table 22 indicate that single models consistently outperform random guessing, with AUC-PRC values ranging from 0.0074 to 0.0149, whereas the best performing OCC models obtained varying results depending on the dataset characteristics. Furthermore, single models consistently performed better than ensemble models. In summary, the findings of this study provide a comprehensive exploration of the efficacy of OCC methods in predicting adverse pregnancy outcomes and addressed the research questions we sought to elucidate. The results consistently demonstrate that OCC models outperformed random guessing thresholds, underscoring their value in addressing the challenges posed by class imbalance and anomaly detection. However, the best OCC model varies across datasets, reflecting the importance of tailoring models to specific contexts. Interestingly, while ensemble methods exhibited potential in enhancing OCC model performance, single models consistently maintained a strong predictive performance, both in terms of AUC-ROC and AUC-PRC value. These results emphasize the robustness and versatility of single models across different datasets. Furthermore, the research underscores the significance of appropriate normalization methods and combination methods in enhancing model performance. As discussed in the Introduction section, it is very challenging to obtain high prediction accuracy for early prediction in this domain. We are getting reasonable accuracy which is better than random prediction. In other words, the models learned the problem. The models can be helpful to the medical practitioners to take better decisions. Overall, this study lays a sound foundation for the development of models for predicting adverse pregnancy outcomes for highly imbalanced data. Conclusion We conducted a comprehensive analysis of models for predicting vPTB and xPTB in an Emirati pregnant population with nulliparous and parous women. The analyses exclusively used EHR data obtained during the first trimester of pregnancy that included vital information on maternal health, lifestyle, medical history, and socioeconomic factors. Multiple experimental setups were explored, and different combinations of risk factors were used to assess their impact on prediction performance in both parous and nulliparous populations. We employed several OCC algorithms and ensemble techniques to combine the strengths of individual algorithms. Our analysis highlighted the significance of appropriate normalization methods and feature selection in enhancing the predictive capabilities of the models. To address the limitations of predicting vPTB and xPTB, we demonstrated that OCC algorithms and ensembles can be employed using majority samples alone (without the need of anomaly samples during training). This study utilizes a dataset collected from the United Arab Emirates (UAE), which may limit the generalizability of its findings to other countries. Variations in genetics, demographic profiles, socio-economic factors, healthcare infrastructure, and access to healthcare can significantly influence health outcome. These factors must be considered when applying the results beyond the UAE. OCC algorithms have shown better results across diverse domains and dataset types56,57. Therefore, the proposed OCC framework is expected to be applicable to similar datasets from other regions. We also aim to improve our pregnancy outcome prediction methods by exploring different algorithms. To provide a timely predictive model, this paper addresses the problem of early prediction of very preterm and extreme preterm births; hence, only first-trimester data were used. However, the risk factors may change over time (for instance, during the second trimester); therefore, such changes should be incorporated into future prediction models. Furthermore, we aim to perform extensive hyperparameter optimizations. Traditionally, the hyperparameter optimization is done using validation data. However, in OCC problems, training data has only normal data points. In the absence of outlier points, we cannot create validation data (consisting of normal and outlier points)55. Therefore, we aim to generate artificial outlier points to create validation data that can be used for hyperparameter optimization. Various methods have been proposed to create artificial outlier points from one class dataset57–59. Desir et al.58 propose an artificial outlier generation process using the distribution of the training one class data. Khan et al.59 suggest that ‘outliers’ from the normal class can serve as proxies for the outlier points. In future, we aim to use these methods to create validation data. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Acknowledgements This work was supported by a grant from the United Arab Emirates University (31R239 Grant code G00003303). Author Amir Ahmad also acknowledges financial support from United Arab Emirates University (Grant No. 12R313) through the Big Data Analytics Center, UAEU. Abbreviations United Arab Emirates One-Class Classification Preterm birth Low birthweight Electronic health records Area under curve Precision-recall curve Author contributions A.A. conceptualized the study, developed the methodology, performed the experiments, and drafted the original manuscript. W.K. contributed to drafting the original manuscript. M.M.A. performed the experiments and helped in analysing the results. M.M.M. supported the conceptualization and manuscript editing. N.G. contributed to the study design from a medical domain perspective. L.A.A. provided the data, supported budget acquisition, and contributed to conceptualization from the medical domain. A.A. serves as the corresponding author (amirahmad@uaeu.ac.ae). All authors reviewed and approved the final manuscript. Funding This work was supported by a grant from the United Arab Emirates University (31R239 Grant code G00003303). This work is also supported by a grant from the United Arab Emirates University (Grant No. 12R313). Data availability The data and code presented in this study can be made available on request from the Mutaba’ah Study. For data, approval from a research ethics committee may be required. Please get in touch with Luai Ahmed (luai.ahmed.uaeu.ac.ae), the author responsible for data requests. Declarations Competing interests The authors declare no competing interests. Ethics approval and consent to participate The study was conducted in accordance with the Declaration of Helsinki and approved by the United Arab Emirates University Human Research Ethics Committee (ERH-2017-5512) and the Abu Dhabi Health Research and Technology Ethics Committee (DOH/CVDC/2022/72). Informed consent was obtained from all subjects involved in the study. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement The data and code presented in this study can be made available on request from the Mutaba’ah Study. For data, approval from a research ethics committee may be required. Please get in touch with Luai Ahmed (luai.ahmed.uaeu.ac.ae), the author responsible for data requests.