The American Diabetes Association defines gestational diabetes mellitus (GDM) as a “diabetes diagnosed in the second or third trimester of pregnancy that was not clearly overt diabetes prior to gestation” (16). This disease has been related to several negative outcomes on maternal and fetal health. In the short-term, it increases the risk of pre-eclampsia, preterm delivery, macrosomia and clinical neonatal hypoglycemia; and in the long-term, of maternal prediabetes, maternal diabetes, offspring obesity and offspring impaired fasting glucose (17).

ML has been applied in GDM research, for diverse purposes ( Table 1 ).

Preeclampsia (PE) is a pregnancy syndrome that presents two different clinical scenarios, both characterized by the development of maternal hypertension from the 20th week of gestation, an alteration that persists throughout pregnancy. The first of the scenarios is characterized by a moderate form of PE, which symptoms become evident late, from 34 weeks of gestation. It is characterized by blood pressure ≥140/90 mmHg, and other symptoms that indicate liver or renal damage, thrombocytopenia or proteinuria ≥3g/24h, and by not inducing alterations on fetal growth. The second of the scenarios correspond to a severe form of PE, which symptoms become evident before 34 weeks of gestation. It is characterized by blood pressure ≥160/110 mmHg, multisystemic damage and/or proteinuria ≥5g/24h, and for being generally associated with intrauterine growth retardation (IUGR) (33). These conditions can also lead to more serious situations than PE alone, such as HELLP syndrome and eclampsia, which is a severe form of PE accompanied by seizures (34). Severe forms of PE are associated with at least two times the risk of IUGR, and fetal and neonatal death (35). The origin of PE is still unknown, however, the most accepted hypothesis indicates that the placenta does not form properly. The latter would not allow a correct flow of maternal blood towards the placenta, triggering a compensatory response that would increase blood pressure to meet the metabolic requirements of the fetus in gestation. This process would begin during the first trimester of pregnancy, producing serious effects on the mother, and affecting the fetus during the second and third trimesters of pregnancy (36). Thus, the early detection of PE, i.e. before the appearance of adverse symptoms in the mother, is necessary.

The early detection of PE has been assessed through the determination of the levels of human chorionic gonadotropin (37), anti-Müllerian hormone (38), sFlt-1 (39), the soluble form of Endoglin (40), among others, with sensitivities between 20 and 80%, and specificities between 40 and 90%. Interestingly, algorithms mediated by ML have been proposed as new strategies to predict this pathology earlier ( Table 2 ). Various ML models have been developed for PE prediction using different types of variables, such as metabolites (41), proteins (42), plasma DNA (26) and circular RNA (43), but by far the most common approaches are based on maternal medical data (44–51).

Some ML-based studies have aimed to predict PE before 20 weeks of pregnancy. For example, Marić et al. used clinical and biochemical maternal data and different ML techniques to predict this pregnancy complication before 16 gestational weeks. Their elastic net (EN) model achieved an AUC of 0.79 for all cases of PE, and an AUC of 0.89 for early-onset PE, showing that ML approaches can become a powerful early prediction tool for this obstetric disorder (46). Sandstrom et al. also used clinical and biochemical maternal variables and different ML techniques to predict PE, but before 15 weeks of gestation. Their LR model with 12 pre-specified variables yielded AUCs of 0.68, 0.68 and 0.67 for PE with delivery <34, <37 and ≥37 weeks of pregnancy, respectively (48). A different example is the study of Gupta et al., who aimed to predict hypertensive disorders of pregnancy, including PE, with placenta ultrasound images from the first trimester of gestation. The analysis of abnormal placental image texture with deep convolutional neural networks (CNN) achieved a sensitivity of 70.6% and a specificity of 76.6% (51). In contrast to the aforementioned articles, other ML-based studies have intended to predict PE from 20 weeks of gestation onwards. For instance, Han et al. measured 25 parameters of maternal clinical chemistry before PE clinical diagnosis, and combined them to predict this pregnancy disorder. Their back-propagation neural networks (BPNN) model, which strongest predictors were ALB, MPV, BUN, LDH and TG, displayed an accuracy of 79.8% (44). Likewise, Jhee et al. retrieved maternal data (collected between 14 and 34 weeks of pregnancy) from EHR and tested them to predict late-onset PE. Their ML models, based on decision trees (DT), naïve Bayes (NB), SVM, RF, stochastic GB, and LR reached AUCs of 0.857, 0.776, 0.573, 0.894, 0.924 and 0.806, respectively (45).

Other PE-related studies have applied ML in additional contexts, such as biomarker identification, risk estimation and drug screening. For example, Liu et al. analyzed microarray data to identify hub genes as diagnostic biomarkers of PE. Their bioinformatics approach revealed 17 differentially expressed hub genes between PE and control subjects: IL7R, IL18, CCL2, HLA-DRA, CD247, ITK, CD2, IRF8, CD48, GZMK, CCR7, HLA-DPA1, LEP, IL1B, CD8A, CD3D and GZMA. Those hub genes were combined and assessed for classification by SVM. Their model reached an AUC of 0.958 in the training set, and an AUC of 0.834 in the test set (52). Similarly, Guo et al. screened placental mRNA data to identify PE biomarkers. Their ML-based approach allowed them to select a subset of 13 mRNA features: HTRA4, PROCR, MYCN, ERO1A, EAF1, PPP1R16B, CRH, FLNB, PIK3CB, PLAAT3, FBN2, RFLNB, and TKT, which were combined and tested for PE and control subjects classification by ML. Their model, which fused three ML classifiers, C4.5, AdaBoost and multilayer perceptron, yielded an accuracy of 82.2% (53). A different case is the study of Bodnar et al., who aimed to assess the effect of fruit and vegetable intake and dietary synergy on the risk of various adverse pregnancy outcomes. Their SL with TMLE ML model revealed that high fruit and vegetable densities were associated with 3.2 and 4.0 fewer cases of PE per 100 births, respectively (54). A final example is the article of Tejera et al., who developed a ML-based strategy to identify currently existing drugs that could be repurposed for PE management. Their approach was built on pharmacological targets of drugs under clinical trial for PE, and was designed to exclude those that have shown negative effects in pregnancy. Their ML-based virtual screening identified estradiol, estriol, vitamins E and D, lynestrenol, mifrepristone, simvastatin, ambroxol, and some antibiotics and antiparasitics as potential drugs for PE treatment (55).

The World Health Organization (WHO) defines perinatal deaths as those that occur from late stillbirth, i.e. after 28 weeks of gestation, up to 28 days of extra-uterine life, including late neonatal deaths (56). Worldwide, more than 5 million perinatal deaths happen every year (57). Progress in reducing the high numbers of stillbirths and neonatal deaths has been slow. Even though the rate of perinatal deaths has been lowered in developed countries, its reduction in low- and middle-income countries has been insufficient. Indeed, low- and middle-income countries present the highest rates and the slowest reduction (58, 59). The Sustainable Development Goals set by the United Nations General Assembly include to put an end to the avoidable deaths of newborns by 2030 (60), however, during 2019 there were approximately 7000 newborns deaths each day (61). These numbers highlight the necessity to implement new methods and techniques to identify high-risk pregnancies, early enough to be able to provide them personalized attention so as to improve prevention, or reduce risk and perinatal death.

Spontaneous abortion (SA) is defined as the loss of pregnancy before the 20th week of gestation (73). It is often referred also as miscarriage, but according to literature, miscarriage is considered to occur before the 24th week of gestation (74). Both situations imply a common and serious pregnancy complication that has a significant psychological impact on the mother and the family. For this reason, and due to its complicated etiology (75), SA has become a hot topic in scientific research and gynecology.

Recent advances in technology, particularly in the artificial intelligence field, have allowed the use of the increasing amount of data that can be obtained in biomedical studies to improve patients’ outcomes. This is consistent with the notion of precision medicine, that is, the need of a more personalized medicine to improve or predict the medical outcome (76), in this case, of a pregnant woman.

Interestingly, ML has been applied in the context of SA and miscarriage ( Table 4 ). In 2013, Bottomley at al., developed a score based on demographic data, symptom variables and ultrasound data to predict the likelihood of a woman to have a successful pregnancy by performing a retrospective study (74). The ML method used was LR. Interestingly, the authors found that the combination of all the factors was able to provide a more accurate prediction of pregnancy viability than the obtained by analyzing the factors in a separated way, with an AUC of 0.924. This score model worked, but at that time it was not proven if it would be able to prevent miscarriage and, as the authors pointed out, the psychological morbidity associated with pregnancy loss should be integrated to the analyses. A distinct approach was made in 2019 using next generation sequencing to analyze 200 DNA samples of 100 couples presenting recurrent miscarriages (RM) (77). This work aimed to develop an algorithm based on the genetic analysis of the HLA protein codifying genes, considering the relationship of the HLA antigen sharing between couples and SA (81) in the context of immune interactions as a possible cause of SA and RM. It has been described that when the mother and the father share HLA antigens, the mother and the fetus will be homozygous for several of these loci. This issue alters the mother immunologic protection to the fetus inducing immunologic rejection and consequently SA (82). The SVM-based algorithm was able to correctly classify 67% of the total subjects, with an AUC of 0.71 and a false positive rate of 57%, which negatively affected the algorithm performance. Interestingly, this study is one of the first to predict RM probabilities in a case-by case basis, having a potential use in couple genetic counseling before conception. A different example is the study of Wu et al., who aimed to predict recurrent SA with prethrombotic state (PTS) serum biomarkers. PTS is known as one of the possible causes of SA. Wu et al. work was based on the analysis of different PTS-related proteins using multiplex array technology (78). They were able to distinguish control and affected individuals with high accuracy and precision using IL-24, exotoxin-3 and epidermal growth factor. Indeed, their DT model got an AUC of 1.000. Despite this excellent result, the cohort used for this study needs to be incremented to evaluate the real diagnostic power of this promising model.

In vitro fertilization embryo transfer (IVF-ET) is nowadays an alternative for couples with difficulties to conceive. This procedure implies high risks of miscarriage, being psychologically stressful for couples. Therefore, it becomes necessary to find a way or system that allows the prediction of the transfer outcome, and the early detection of possible problems (83). Recently, Liu et al. developed a ML-based model with historic data obtained by transvaginal ultrasonography from females that underwent IVF-ET. The study only considered women with viable singleton and 6-12 weeks of pregnancy (79). The authors were able to predict embryonic development after transfer using six different ML-classifiers, with AUCs ranging from 0.91 to 0.97 when fetal heart rate (FHR) was included among the predictors. The most accurate prediction was obtained by RF at the 10th week after embryo transfer, with an AUC of 0.99. Other example is the article of Huang et al., who used deep learning to predict pregnancy outcomes in patients with recurrent reproductive failure (RRF), including recurrent pregnancy loss (RPL) and recurrent implantation failure (RIF). The study defined RPL as two or more SA before 20 weeks of pregnancy, and RIF as couples unable to conceive after multiple IVF-ET cycles. The authors analyzed EHR data with sparse coding, and predicted four pregnancy outcomes: biochemical pregnancy, clinical pregnancy, ongoing pregnancy and live birth. They got testing accuracies that ranged between 54.2% and 89.7% for the different pregnancy outcomes. Notably, the best model for the prediction of biochemical pregnancy was obtained with a panel of 10 endometrial immunological markers, while the best models for the other three outcomes, were obtained with a panel of 15 autoantibodies. The authors discussed that this knowledge could help clinicians to plan a more personalized diagnosis and treatment for patients with RRF (80).

The WHO defines PTB as the delivery of alive babies before 37 weeks of pregnancy are completed (56). Based on gestational age, it can be sub-categorized as: extremely preterm, before 28 weeks; very preterm, between 28 and 32 weeks; and moderate to late preterm, between 32 and 37 weeks. Most of preterm deliveries are spontaneous, although some are provider-initiated (56).

PTB is the main cause of death in children under 5 years of age worldwide. Furthermore, it has short and long-term consequences on newborns’ health, which imply a significant psychological and economic burden to families and health systems (84). The development of PTB predictive tests could be useful to identify high risk pregnancies, which could guide the healthcare personnel to offer prophylactic interventions and make antenatal management decisions (85).

ML has already been applied to develop predictive models for PTB ( Table 5 ). For instance, Khatibi et al. aimed to predict spontaneous and provider-initiated PTB with data from the Iranian Maternal and Neonatal registry, which includes information of more than 1,400,0000 deliveries and 112 features. The authors used different big data ML algorithms to classify pregnant women in two steps. In the first step, all subjects were classified into term or PTB; and in the second step, the subjects classified as PTB in the first step, were then sub-classified as spontaneous or provider-initiated. Their best model, an ensemble of DT, SVM and RF, achieved a weighted average accuracy of 81%, and an AUC of 68% (86). Similarly, Belaghi et al. used first and second trimester information from the Ontario’s Better Outcomes Registry and Network database, and different ML methods to predict overall and spontaneous PTB. The investigation considered 112,963 pregnancies. For overall cases, the best models were obtained by ANN, and reached AUCs of 60.3% and 79.8% in the validation cohort at the first and second trimester, respectively. For spontaneous cases, the best results were obtained by LR, and got validation AUCs of 59.4% and 64.5% at the first and second trimester, respectively (87). A different approach was followed by Gao et al., who used EHR text data and deep learning ML methods to predict extreme PTB. Their dataset involved 10 years of EHR information from 25,689 deliveries at the Vanderbilt University Medical Center. The long short-term memory (LSTM) recurrent neural networks (RNN) ensemble model allowed to predict extreme PTB with an AUC of 0.744 in the validation cohort, greater than the obtained by LR, SVM and GB (88). This is an interesting result, although this work didn’t differentiate spontaneous from provider-initiated cases. Likewise, Zhang et al. aimed to predict PTB with continuous EHR data and LSTM. Their dataset included first and second trimester medical parameters from more than 25,000 pregnant women who received antenatal care and had vaginal delivery at the Hangzhou Women’s Hospital. Notably, the time-series deep learning technique LSTM achieved a better predictive performance than the traditional cross-sectional ML technique XGBoost, with cross-validation AUCs of 0.651 and 0.516-0.601, respectively (89).

All the aforementioned studies based their predictive models on clinical and biochemical maternal information available in databases. However, other articles have assessed alternative types of data to predict PTB. Such studies are very useful to find novel biomarkers for PTB, and to propose informed hypotheses about its causes and underlying mechanisms, which are not fully understood (84, 85). For instance, Aung et al. measured an extensive set of 65 urine and plasma biomarkers, and combined them with ML to predict PTB at 26 weeks of gestation. They tested three ML methods: LR, adaptive EN and RF. The best validation results were obtained with the latter. The combination of all the biomarkers with RF yielded AUCs of 0.85 and 0.79 for overall and spontaneous PTB, respectively. Then, the authors divided the biomarkers into five groups, i.e. DNA damage markers, angiogenic factors, protein damage markers, inflammatory markers and lipid damage markers. The best predictive performances were obtained with lipid damage markers and RF, with AUCs of 0.84 and 0.79 for overall and spontaneous cases, respectively. Furthermore, the study identified the enzymatic pathway that contributed the most to that prediction: the eicosanoid lipoxygenase pathway. The combination of 15 lipoxygenase metabolites with RF got AUCs of 0.83 and 0.82 for overall and spontaneous PTB, respectively (90). Another example is the study of Chen et al., who applied untargeted LC-MS plasma metabolomics to identify metabolites that could be related to PTB, at 24-28 gestational weeks. The authors identified 17 and 16 biomarkers for overall and spontaneous cases, respectively, and tested their predictive performance with seven ML classifiers. The best results were obtained by RF, with AUCs of 0.92 and 0.89 in the testing dataset. Interestingly, most of the identified biomarkers were fatty acids, which suggests their involvement in the pathogenesis of PTB (91). Similarly, Jehan et al. performed an early pregnancy multiomics characterization of PTB. The authors applied untargeted transcriptomics and targeted proteomics on plasma samples, and untargeted metabolomics on urine specimens. They used a 2-step ML algorithm, in which a model was first trained for each omics dataset, and then combined into a final model. The integrated model achieved a cross-validation AUC of 0.83, higher than the obtained for the different omics datasets alone. The work also identified the features that were more associated with PTB: a proteomics inflammatory module, including IL-6, IL-1RA, G-CSF, RARRES2 and CCL3; and an urine metabolomic module, enriched for glutamine and glutamate metabolism, and valine, leucine and isoleucine biosynthesis pathways (92).

Some less common approaches have also been applied in the context of PTB prediction. For example, Despotovic et al. tested EHG recordings to predict PTB. They built ML models using k-nearest neighbors (KNN), SVM, RF, RF with synthetic minority oversampling technique (SMOTE), and RF with adaptative synthetic (ADASYN) sampling. Their RF-ADASYN model allowed to predict PTB at 22-25 weeks of pregnancy, with an accuracy of 99.23% and an AUC of 0.999 in cross-validation (93). A different case is the work of Rawashdeh et al., who combined 19 clinical maternal parameters with ML methods to predict PTB in a high risk cohort. They developed two different strategies to analyze their data. The first one aimed to predict whether the pregnancy would continue beyond 26 gestational weeks (the lower limit for PTB in this study) and the potential value of performing cervical cerclage to prolong the pregnancy. For this first aim, the authors tested four different classification ML methods, DT, RF, KNN and NN; solo and with SMOTE. The highest testing AUC was obtained by the KNN-SMOTE model, with a value of 1.000. The second strategy of the authors aimed to predict the timing of spontaneous delivery after cervical cerclage, an approach that wasn’t assessed in any of the previously discussed articles. For this second aim, they tested five different regression ML methods, linear regression, Gaussian process, RF, K-star and locally weighted learning. The best correlation with the actual gestational age at delivery was obtained by the RF model, with a value of 0.752 in the testing dataset. Such a regression ML model could help physicians to define prophylactic interventions timely, and reduce PTB-related perinatal morbidity and mortality (94).

Cesarean section is an effective mean to solve medical and surgical complications during dystocia and severe pregnancy disorders, and has an irreplaceable role (95). Delivery through cesarean section reduces the risk of maternal-fetal morbidity and mortality, when is medically indicated (96). Emergency cesarean section (EMCS) can be a procedure that saves lives if pregnant women experience abnormal conditions during vaginal delivery, such as fetal suffering, eclampsia or severe preeclampsia (97). Deciding to perform an EMCS is a complicated process, occurring only in specific obstetric conditions, and requires awareness and rapid assessment of the risk of the situation (98). Failure to perform EMCS on time can lead to postpartum mental disorders and other severe adverse maternal and fetal outcomes (99, 100). Recognizing an acute situation during pregnancy, labor or delivery, that can be life threatening and that could require an EMCS, is considered one of the most challenging tasks in obstetrics (101).

Visual inspection of CTG traces by obstetricians and midwives is the gold standard for monitoring the wellbeing of the fetus during antenatal care (102). One of the areas in which mathematical and computational tools for data analysis, such as ML methods, excel is in the analysis of instrumental continuous signals ( Table 6 ). Several output data from instruments used in clinical diagnosis or monitoring are composed of this type of signals, in which between any two points there can be a large amount of data points, as large as allowed by the signal resolution, or even an infinite amount in the case of analog instruments. CTG traces are a great example of this type of data in obstetrics. The problem with this type of data is that its visual interpretation is highly dependent on the observer’s experience and can be strongly subjective. Most importantly, clinical decisions such as pregnancy intervention through cesarean section are made using visual inspection of CTG traces. It has been reported that the positive predictive value produced by obstetricians to anticipate negative outcomes that require cesarean section deliveries is only 30% (110). However, although human eye may fail to provide a reliable and objective interpretation, mathematical tools for pattern recognition are not subjected to the observer’s bias.

Two different articles published by the same group in Liverpool have addressed the observer variability of CTG traces using a ML approach (102, 103). The authors applied signal processing techniques to extract relevant features from CTG traces and modeled the data using different ML methods, such as DNN, LDA, RF, SVM and ensemble classifiers. They were able to classify cesarean section and vaginal deliveries from CTG traces with cross-validation AUC values of 96-99%. Other study performed by an Italian research group used a similar methodology and obtained consistent results, that is, a cross-validation AUC value of 96.7% by RF (104). Likewise, a Chinese study that proposed a comparable strategy to classify normal and abnormal CTG traces reported an AUC of 0.95 by CNN in cross-validation (105). Their results demonstrate that ML methods significantly improve the prediction efficiency of necessary cesarean sections, and that their use provide a valuable decision support tool to minimize subjective interpretations of CTG traces from medical practitioners.

Besides CTG traces analysis, ML methods have been applied on EHR information to predict cesarean section and identify important variables, as well as to understand the interaction between those variables. The model developed by Clark et al. using a classification and regression tree had an AUC value of 0.7, which was considered acceptable (106). The three features that contributed the most to that model were hospital type, maternal BMI and intrapartum oxytocin dose.

Other uses of ML have been tested in the context of cesarean sections. For example, a decision-support ML-based model for assessing intrathecal hyperbaric bupivacaine dose using physical variables during cesarean section was developed, providing the anesthesiologists a new tool that gives new insights into the potential impact of controversial parameters (107). The least absolute shrinkage and selection operator regression model got a mean squared error of 0.0087. ML has also been applied to predict surgical site infection in cesarean section wounds, which is a leading cause of mortality and an important health concern in low-resource countries (108). The best model was obtained with mobile device images and LR, and achieved an AUC of 1.0. Prediction of likelihood of a successful vaginal birth after former cesarean deliveries has also been addressed using ML, which may help as a decision-making tool that could contribute to a reduction in cesarean deliveries rates (109). The EHR-based RF model reached an AUC of 0.69, better than the obtained by DT and LR.