Full and original articles related to ML techniques on complications during pregnancy published in English from 2015 to 2020 were searched on PubMed, Web of Science, and Scopus databases. Search terms were chosen and searches performed in an iterative process, initially using word headings associated with ML, such as “machine learning,” “deep learning,” “artificial intelligence,” and related to perinatal complications, such as “complications in pregnancy” and “pregnancy complications,” and excluding articles related to postpartum and congenital complications. For PubMed, the MESH terms were used to include associated synonyms in the search, and for Scopus and Web of Science, the terms of interest mentioned before with Boolean operators were used (Table 1). The search and final collection of articles were 98 articles, of which 20 were excluded by duplication.

The included criteria for the articles searched were 1) English original articles, 2) access to full text, 3) studies based on humans, 4) studies using machine learning methods to predict complications in pregnancy, and 5) complications during pregnancy and at term in the mother and the newborn. The exclusion criteria applied were 1) systematic reviews, meta-analysis, and bibliographic reviews; 2) articles that included postpartum complications; 3) maternal congenital disease that increases the risk of perinatal complications; and 4) fetal congenital diseases. Articles were added manually according to the aforementioned criteria.

All articles found were uploaded to the Mendeley desktop platform, where they were saved in a dedicated folder for the present systematic review. After eliminating the duplicate articles, a total of 78 articles remained. Then 16 articles were excluded by title, 18 were excluded by criteria, and 19 were excluded after reading. Finally, 31 articles for the review were selected. The selected articles were classified by the ML model used, type of features used, outputs, and performance metrics, in order to estimate which methods are the most accurate in the context of predicting perinatal complications.

The 31 articles were subjected to the CASP checklist, which contains 11 questions to help evaluate a clinical prediction rule (CASP, 2017). Study quality was scored according to the CASP critical score: if the criterion was met entirely = 2 points; criterion partially met = 1 point; and criterion not applicable/not met/not mentioned = 0. Finally, study quality was ranked: a total score of 22 = high quality; 16–21 = moderate quality; and ≤15 = low quality.

To optimize the visualization of the results obtained in the systematic review, several tables were made according to the terms addressed in the search, showing complications that the models seek to predict, input characteristics for the training of the ML model, the type of ML used, and its validation and performance metrics.

To apply the PRISMA method, the articles have been classified according to the criteria mentioned before: title, abstract, and the full article. A total of 84 articles were found, of which 52 were excluded because they did not meet the search criteria of interest. Of these, 16 were eliminated by title, 18 after reading the abstract, and 19 after reading the entire article, leaving 31 articles to analyze (Figure 1 and Supplementary Table S2). The type of studies in the manuscripts analyzed were mainly cohort (87.2%) and retrospective (96.8%). The populations studied were primarily from Asia and Europe (both 32.3%), followed by North and South America (22.5 and 6.5%, respectively). An increased rate of studies was observed during 2019 (35.5%) (Table 2). The features mainly used to predict perinatal complications are electronic medical records (48%) and then medical images (29%), biological markers (19%), and 4% are based on another type of feature, in this case, sensors (Moreira et al., 2016a) and fetal heart rate (Zhao et al., 2019). Two studies contemplate two features: electronic medical records and medical images (Nair, 2018; Lipschuetz et al., 2020).

According to the CASP checklist, one article met the total score and was classified as a high-quality article (Gao et al., 2019). The rest of the items were classified as moderate quality and none as low quality according to the evaluation criteria (average total score = 18–19). It is essential to mention that the “non-compliance” items were not being mentioned or not applicable to the study. The item asking whether the sample was randomized in 15 articles does not apply since analyzed retrospective electronic health records or images. Regarding using a comparison group, 12 reports do not apply due to retrospective data and data management for the prediction model (Supplementary Figure S1).

The choice of informative, discriminatory, and independent characteristics is crucial to achieving effective algorithms for recognizing, classifying, and regression patterns. Thus, the four types of features analyzed in the articles were electronic medical records (EMRs) (Table 3), medical images (recordings, ecotomographs, ultrasound, resonance, etc.) (Table 4), biological markers (Table 5), and others (sensors and fetal heart rate) (Table 6).

These have been divided into 16 main prediction outputs: prematurity, pre-eclampsia, adverse delivery, size for gestational age, gestational diabetes mellitus, neonatal mortality, fetal acidemia, fetal hypoxia, placental accreta, pulmonary diseases, cesarean section, placental invasion, congenital anomaly, severe maternal morbidity, spontaneous abortion, and trial of labor after cesarean (TOLAC) failure (Figure 2). The main perinatal complications considered in the application of ML are prematurity (7 studies) and pre-eclampsia (6 studies).

Validation methods are strategies that allow the estimation of the predictive capacity of ML models. Fifty-five percent use training tests and the cross-validation method as a validation method with greater reliability in results, while 41.8% use a single validation method and 3.2% do not use any validation method (neither training tests nor cross-validation).

In the present review, 67.7% of the articles used AUC and 61.3% used the accuracy metric. Sensitivity was only evaluated in 61.3% of the studies. While all studies assess results with at least one performance metric, reports of predictive accuracy were often incomplete, with a total of 38.7% of studies reviewing at most two performance methods. According to the studies, none had a clinical application, they only functioned to establish precise prediction systems in the diagnosis of the different perinatal complications presented.

Twenty-one different ML methods were used to predict these 16 perinatal complications. Placental invasion is referred to as placental adhesive disorders observed in women with placenta previa or prior cesarean section that lead to complications such as perinatal hemorrhage and visceral injuries, where an early diagnosis is necessary for appropriate treatment (Sun et al., 2019). Excellent performance of placental invasion can be observed with an AUC and an accuracy of 0.980 and 95.2%, respectively, using the Tree-based Pipeline Optimization Tool (TPOT) (Sun et al., 2019). To predict fetal acidemia, using convolutional neural networks, an AUC and accuracy of 0.978 and 98.4% are achieved, respectively (Zhao et al., 2019). Only one study of the six attempting to diagnose pre-eclampsia had a performance considered as good, using the AdaBoost model, with an AUC of 0.964 and an accuracy of 89% (Munchel et al., 2020). The prediction of prematurity has excellent results in two studies; the one that uses SVM achieves an AUC of 0.952 and an accuracy of 95.7% (Sadi–Ahmed et al., 2017), and the study that uses stacked sparse autoencoder achieves an AUC of 0.900 and an accuracy of 90% (Chen et al., 2019). For the prediction of neonatal mortality, through sociodemographic records using XGBoost, an AUC of 0.842 and an accuracy of 99.7% were obtained (Hamilton et al., 2020). Regarding the performance of the predictions included in the greatest number of studies, prematurity outperformed pre-eclampsia according to the AUC (Table 7).

It was decided to corroborate the performance of the methods based on deep learning. Only four studies used deep learning methods. They all had an excellent performance. For the prediction of fetal acidemia, a deep convolutional network was used with an AUC of 0.978 and an accuracy of 98.4% (Zhao et al., 2019). For the prediction of spontaneous abortion, multilayer perceptron and radial-based networks were used, with an accuracy of 90.9% (Paydar et al., 2017). And as mentioned above, for the prediction of pre-eclampsia, using biological markers and multilayer perceptron, an AUC of 0.908 was obtained (Nair, 2018). For the prediction of neonatal mortality, through sociodemographic records using XGBoost, an AUC of 0.842 and an accuracy of 99.7% were obtained (Hamilton et al., 2020) (Table 8).

The interpretability of ML models refers to the degree to which a human being can consistently predict the outcome of the model (Kim et al., 2016), which has been well accepted by the clinical team. In this systematic review, we found that 24% of the studies use AI-interpretable ML models. The ML methods that were the most used in the prediction of perinatal complications were the random forest, logistic regression, neural networks, and support vector machine (SVM).

Forty-eight percent of the studies explain the main characteristics of pregnant women that could be relevant to predict some conditions. Characteristics and antecedents such as gestational diabetes, cardiovascular disease, underlying diseases, and the age of the mother, as well as the presence of chronic arterial hypertension, are considered high-ranking features for the prediction of premature births; and the father’s nationality is very important to differentiate the provider-initiated spontaneous preterm births (Khatibi et al., 2019).

On the other hand, important predictors to determine the likelihood of a newborn to be small for gestational age (SGA) were smoking, a particular amount of gestational weight gain, and low–birth weight newborn. The body mass index (BMI) before pregnancy, gestational weight gain, and a macrosomic newborn in a previous delivery were the strongest predictors to determine large for gestational age (LGA) newborns (Kuhle et al., 2018). To predict fetal macrosomia, the determining variables were age ≥30, multiparity, 12 kg of total weight gain during pregnancy, abdominal circumference >95 cm (at the last perinatal checkup), and a gestation period over 39 weeks (Shigemi et al., 2019).

In order to predict pre-eclampsia, the most influential variables were systolic blood pressure, serum levels of ureic nitrogen and creatinine, platelet count, serum potassium level, leukocyte count, blood glucose level, serum calcium, and proteinuria levels in the early second trimester (Jhee et al., 2019). Interestingly, high pre-pregnancy BMI and previous preterm births (Pan et al., 2017) were able to predict whether pregnant women will have an adverse pregnancy outcome (preterm, low birth weight, neonatal/infant death, stay in the neonatal intensive care unit) and indicate the main risk characteristics.

Furthermore, in order to predict TOLAC, the determining factors in the prediction model were parity, age, vaginal birth with cesarean section in the past, gestational weeks, minimum gestation week in previous deliveries, the weight of the newborn from the previous delivery, dilation, and head position (Lipschuetz et al., 2020). To predict pregnancy complications associated with placental alterations (pre-eclampsia, GDM, fetal growth restriction, macrosomia), maternal age, BMI, newborn weight, and the results of adverse events in previous pregnancies were the most influential characteristics in the study (Guo et al., 2020).

To predict gestational age at delivery (if the newborn will be preterm) variables such as the date of the mother’s last menstruation, birth weight, delivery of twins, maternal height, hypertension during labor and HIV serological status were decisive in the ML model (Rittenhouse et al., 2019). To determine preterm birth, the presence of premature rupture of membranes and/or vaginal bleeding, ultrasound cervical length, gestation week, fetal fibronectin, and serum C-reactive protein were the determining variables (Mailath-Pokorny et al., 2015). In another study, prediction of preterm birth considered the most relevant variables to be maternal age, whether the mother was black, Hispanic, Asian, born in the United States, delivered by herself or assisted by a physician, presence of diabetes mellitus, chronic arterial hypertension, thyroid dysfunction, asthma, previous stillbirth, fetal weight loss, in vitro fertilization, nulliparity, being a smoker during the first trimester, and BMI (Weber et al., 2018).

Stillbirth can potentially be identified prenatally considering the combination of current pregnancy complications, congenital anomalies, maternal characteristics, and medical history (Malacova et al., 2020). Determining factors for the prediction of fetal acidemia were maternal age, gestational age, pH, extracellular fluid deficit, pCO2, base excess, APGAR 1 and 5 min, parity, gestational diabetes, birth weight, child sex, and the type of delivery (Zhao et al., 2019).

In the case of the prediction of severe maternal morbidity, the following characteristics were determining factors: ventilator dependence, intubation, critical care, acute respiratory failure, ventilation, trauma and postoperative pulmonary failure, fluid and electrolyte disorder, systemic inflammatory response syndrome, acidosis, and septicemia (Gao et al., 2019).

According to the studies, none had clinical application; they only served to establish precise prediction systems to diagnose the perinatal complications presented.

Machine learning plays a vital role and offers solutions with many applications, for example, image detection, data mining, natural language processing, and disease diagnosis (Maity and Das, 2017). This systematic review provides a study of different ML techniques for the diagnosis of different perinatal complications and frames a contribution to women’s health. A total of sixteen perinatal complications predicted by various ML models were detected, among which the most studied were prematurity and pre-eclampsia.

ML can significantly improve health care; however, it is necessary to consider the disadvantages of AI in health. Ethical dilemmas need to be addressed and the potential for human biases when creating computer algorithms (Ho et al., 2019). Health-care predictions can vary based on race, genetics, gender, and other characteristics, which could lead to the overestimation or underestimation of patient risk factors if not considered. When it comes to AI analysis in health care, it will be the physician’s responsibility to ensure that AI algorithms are developed and applied appropriately (Jordan and Mitchell, 2015).

In the present systematic review, the main data collection method was the use of electronic medical records. ML techniques can establish patterns from a data set based on electronic medical records (EMRs). Pattern recognition from these records supports in predicting and making decisions for diagnosis and treatment planning (Johnson et al., 2016). The application of EMR-based ML methods can be combined with other sources of large medical data, such as genomics, and medical imaging, which through predictive algorithms could improve clinical diagnosis and treatment systems, when used as complementary information (Barak-Corren et al., 2017). EMR data usually include demographics data, diagnoses, biochemical markers, vital signs, clinical notes, prescriptions, and procedures, which are generally easy to obtain and reduce transfer errors when handling large amounts of information. Previously, several studies have described medical diagnosis prediction tools mediated EMRs (McCoy et al., 2015; Osborn et al., 2015; Nguyen et al., 2017; Rajkomar et al., 2018); furthermore, in the present systematic review, 48% of the features for the diagnosis prediction model to perinatal complications came from EMRs, of which the most used features were sociodemographic maternal characteristics. Thus, this tool can predict perinatal complications common in a given population, contributing to the overall improvement of perinatal public health.

Output variables were usually binary outputs (with complication or without complication). However, some studies quantified the risk, for example, the risk of TOLAC was classified as high, medium, or low (Lipschuetz et al., 2020), and in studies of gestational diabetes, one article quantified it as high risk or low risk (Cömert et al., 2018). The most frequently predicted perinatal complications in ML models were prematurity and pre-eclampsia. According to the literature, the high rate of preterm birth is a public health problem, since these newborns suffer substantial morbidity and mortality in the neonatal period, which translates to high medical costs (McCormick et al., 2011). Pre-eclampsia is a pregnancy disorder characterized by the new onset of hypertension after 20 weeks gestation and organ damage with underlying causes being endothelial dysfunction (ACOG (American College of Obstetricians and Gynecologists), 2020; Carrasco-Wong et al., 2021; Roberts, 1998). It is the leading cause of maternal and neonatal mortality and morbidity (Salsoso et al., 2017; Fondjo et al., 2019). Thus, prediction of the risk for developing pre-eclampsia can be performed in the first half of pregnancy.

Diagnostic accuracy is the ability of a test to discriminate between the target condition and health. This discriminative potential can be quantified by several performance tools, such as sensitivity and specificity, AUC, accuracy metric, and other measurements (Šimundić, 2009). While all studies assess results with at least one performance metric and just 38.7% assess at least two performance methods, reports of predictive accuracy were often incomplete. With this observation, it is imperative to show the same performance tools on the different prediction models to evaluate accuracy compared between them.

In this systematic review, several ML methods were used. One of the better performances was obtained by the Tree-based Pipeline Optimization Tool (TPOT) to predict placental invasion (Sun et al., 2019), which was previously used in the investigation of novel characteristics in data science, providing optimization of the studied parameters (Le et al., 2020). Another excellent performance observed was the convolutional neural network (CNN) to predict fetal acidemia (Zhao et al., 2019). The CNN has gained much attention from attempts made at harnessing its power to automatically learn intrinsic patterns from data, which can avoid time-consuming manual functions engineering, and capture hidden intrinsic patterns more effectively (Oquab et al., 2014). Moreover, in the health-care field, CNN has been shown to capture more hidden data patterns and learn high-level abstraction in problem-solving (Zhang et al., 2017).

It is essential to mention that it is difficult to reach a consensus on the best method for predicting perinatal complications, since not all of them had the same input variables, type of records, and a number of samples. However, the best performance metrics observed were the prediction model of prematurity from medical images using the SVM technique with an accuracy of 95.7% and the prediction of neonatal mortality using the XGBoost technique with an accuracy of 99.7%. SVM has shown simplicity and flexibility to address several classification problems and also offers balanced predictive performance even in studies where sample sizes may be limited (Alkhaleefah and Wu, 2018). The XGBoost technique is a very effective and widely used ML method that data scientists use to achieve state-of-the-art results in many ML challenges (Wang et al., 2020).

Despite the recognition of the value of ML in medical care, impediments persist for its greater acceptance within medical teams (Holzinger et al., 2019). A fundamental impediment relates to the nature of the black box, or “opacity,” of many ML algorithms. The term refers to a system in which only the inputs and outputs are observable, while the question of what is transforming the inputs into the outputs cannot be fully understood (Molnar, 2019). Therefore, new techniques have been developed to facilitate the understanding of the internal functioning of the model, granting interpretability, which seeks to provide transparency to the black box (Freitas, 2014; Doshi-Velez et al., 2017; Lipton, 2018), so that the end-user can understand the model and may even improve the ML system (Freitas, 2014). The improvement in the precision of the prediction will depend on the interpretability of the model to be used. This means that with ML interpretability, clinical staff could know which variables are involved in the prediction of a diagnosis.

Regarding the predictive variables, while most of them agreed with current knowledge, it was also shown that ML models contributed new variables of relevance, which would be interesting to observe in controlled clinical studies (Table 9). For example, pre-eclampsia was found to be predictable based on systemic blood pressure, platelet count, and urinary protein levels as influential variables, with lesser influence found from glucose levels, leukocytes count, serum calcium, and potassium levels (Jhee et al., 2019). Other innovative variables of interest found using ML in the prediction of perinatal complications were newborn sex for the prediction of fetal acidemia (Liu et al., 2019), and father’s nationality and mother’s age for the prediction of provider-initiated spontaneous preterm delivery (Malacova et al., 2020). Nevertheless, some prediction models lack variable measurements, making them impossible to apply in a clinical setting. For example, “weight gain” is mentioned as a predictor for SGA and LGA, but the article does not specify whether it was inadequate or excessive (Kuhle et al., 2018). It is also stated that the underlying disease of the mother influences the delivery initiated by the provider; however, it is not detailed which underlying disease is considered in this association (Khatibi et al., 2019). Also, some studies describe obvious associations, such as low birth weight is associated with SGA, or fetal macrosomia is associated with LGA (Kuhle et al., 2018). pH was also a predictor of fetal acidemia, which is logical since this condition is associated with pH changes (Zhao et al., 2019). Since the engineering team behind these investigations emphasizes these characteristics in the results, without taking this obviousness into account, it is imperative to include clinical experts on women’s health into AI and data science teams.

Only 6.4% of the studies were case–control studies, while the vast majority were cohort studies. This may limit the use of these results in clinical practice (Salazar et al., 2019). Only one study was multicenter for predicting neonatal morbidity (Khatibi et al., 2019), representing higher quality evidence. Among the best performing studies, it is noteworthy that most had less than 1,000 patients, and only one based on XGBoost to predict neonatal mortality had over 10,000 patients. This may be risky since the sample size may not be representative for a given geographic group, representing one of the limitations of ML in health (Vayena et al., 2018). Also, another significant limitation of the present systematic review is that all studies included have different baselines, variable inputs, and separate complications (endpoints) assessed in their prediction, making it difficult to compare them.

It is essential to mention that all the studies reviewed have not been applied in a clinical phase; however, the majority mention that to optimize the results obtained, and the models should be used in hospitals or health services that care for pregnant women. Future prospective studies and additional population studies are needed to assess the clinical utility of the model for the real world (Liu et al., 2019; Malacova et al., 2020).

Few systematic reviews have addressed the use of AI in pregnancy. The first one describes how AI has been applied to evaluate maternal health during the entire pregnancy process and helped to understand the effects of pharmacological treatments during this stage (Davidson & Boland, 2020). The second systematic review concluded that using ML algorithms is better than using multivariable logistic regression for prognostic prediction studies in pregnancy care, focusing mainly on decision-making for the medical team (Sufriyana et al., 2020). Furthermore, the third one performed exclusively on neonatal mortality reported that ML models can accurately predict neonatal death (Mangold et al., 2021). Last, the use of modern bioinformatics methods analyzing ML models as non-invasive measures of heart rate variability to monitor newborns and infants was reported (Chiera et al., 2020). Although this body of evidence does not focus on predicting pregnancy complications, it encourages the clinical use of IA to support women’s health during pregnancy.