Digital technologies and artificial intelligence in the diagnosis of cardiovascular complications in pregnancy: a review

Full Text

INTRODUCTION

Pregnancy in women with cardiovascular diseases (CVD) is one of the most challenging areas of high-risk obstetrics. According to global statistics, CVD complicates up to 4% of pregnancies [1] and remains the leading cause of indirect maternal mortality. In the Russian Federation in 2019, diseases of the circulatory system were the leading cause of maternal mortality due to non-obstetric disease, accounting for 57.6% [2]. The growing prevalence of this condition is attributable both to the increasing number of women with congenital heart disease surviving to reproductive age and to the trend toward delayed childbearing, which is associated with higher rates of hypertension, obesity, diabetes mellitus, and other comorbidities [3]. Contemporary epidemiological studies highlight the substantial contribution of acquired heart disease to maternal mortality rates [2].

The spectrum of CVD in pregnant women is broad. In some patients, CVD first presents during pregnancy or in the postpartum period. In Russia, a substantial proportion of CVD-related maternal deaths occur in women whose disease had not been diagnosed before pregnancy and first presented during labor or in the postpartum period [4]. Diagnostic challenges arise from the overlap between disease symptoms and the physiological changes of pregnancy, as well as from the poor correlation between clinical presentation and disease severity. In a study of peripartum cardiomyopathy, one quarter of patients with a left ventricular ejection fraction < 25% reported only mild symptoms [5]. At the same time, timely diagnosis is crucial: a retrospective analysis of maternal mortality cases in France showed that almost half (47%) of such outcomes could have beenpreventable with earlier recognition [6]. At present, no universal screening protocols are available, and decisions regarding further diagnostic evaluation are most often guided by clinical suspicion and patient-reported symptoms. The key diagnostic tools remain electrocardiography, echocardiography, and, when indicated, magnetic resonance imaging [1].

Effective management of pregnant women with heart disease requires a multidisciplinary approach involving cardiologists, obstetricians, anesthesiologists, and other relevant specialists [1]. We believe that a central role is played by anesthesiologist, who is responsible for peripartum risk assessment, selection of the optimal analgesic strategy, and maintenance of stable hemodynamics during labor or surgical procedures.

Particular attention should be paid to the implementation of artificial intelligence (AI) technologies, which have seen increasing use in clinical medicine in recent years. Systems based on machine learning methods and advanced analytical methods enable the integration and interpretation of large volumes of clinical and imaging data, offering new opportunities for early diagnosis, risk stratification, and outcome prediction [7, 8]. It should be emphasized that AI does not replace clinical thinking but rather expands its capabilities by providing decision-making support. Given the importance of cardiovascular imaging and diagnostics in cardiology, the prospects for using AI in high-risk obstetrics among women with CVD appear particularly promising [4, 6, 9].

This review examines current approaches to pregnancy management in women with CVD and promising opportunities for integrating AI into this field.

ARTIFICIAL INTELLIGENCE IN RISK STRATIFICATION

Cardiovascular risk assessment in pregnant women is a key component of prenatal counseling and an important tool for planning monitoring in the antenatal, intrapartum, and postpartum periods. Several risk prediction models have been developed for patients with heart disease to stratify the risk of complications. For example, the CARPREG (Cardiac Disease in Pregnancy Study) I and II models [10, 11], based on clinical characteristics, diagnostic findings, and imaging data, are used to estimate the risk of cardiac complications during pregnancy. The ZAHARA (Zwangerschap bij Aangeboren Hartafwijking) risk score [12] focuses primarily on women with congenital heart disease. In recent years, the modified World Health Organization (WHO) classificationhas been widely adopted, classifying women into five risk categories according to the type of underlying heart disease, each associated with a corresponding risk of complications [13]. Use of this system improves the accuracy of counseling, helps determine whether pregnancy is advisable, and supports development of an optimal management strategy.

Traditional prognostic models, despite their clinical value, have limited ability to identify nonlinear relationships among risk factors [11, 12]. In contrast, AI algorithms can analyze large volumes of multimodal data, capture temporal changes, and identify complex patterns that cannot be detected using conventional multivariate analysis methods [2, 14, 15]. For this reason, AI is regarded as a promising tool for dynamic risk stratification and early detection of complications in pregnant women with CVD [2, 6, 9].

N. Shara et al. [7] presented a machine learning model developed using an analysis of more than 6000 electronic medical records, including 604 cases of cardiovascular complications among pregnant women. The most common complications were preeclampsia (90.6%), thromboembolic events (2.7%), and acute kidney injury or renal failure (2.2%). The model incorporated both static variables (medical history and demographic characteristics) and dynamic variables (clinical symptoms and ongoing clinical findings). Validation showed that the model was able to predict preeclampsia a mean of 62 days before diagnosis was documented in the medical record, myocardial infarction 66 days before diagnosis, and cardiomyopathy and heart failure 13 days before diagnosis. Despite its promising potential, the clinical applicability of these reported lead times remains uncertain. The model is likely to be most useful as a dynamic early warning tool used throughout pregnancy.

S. Zahid et al. [8] analyzed data from more than 2 million pregnant women, 7% of whom developed acute cardiovascular and related complications, including preeclampsia or eclampsia, peripartum cardiomyopathy, acute heart failure, acute coronary syndrome, arrhythmias, acute kidney injury, pulmonary edema, and venous thromboembolism. Based on 14 variables (demographic characteristics, medical history, and laboratory parameters), the authors developed a predictive model with a 100-point risk scoring system. The model achieved an ROC-AUC (Area Under the Receiver Operating Characteristic Curve, ROC-AUC) of 0.68 on validation, indicating moderate predictive performance and the need for further refinement before clinical implementation. Despite the limited predictive accuracy, this study demonstrated the potential of AI in processing large datasets and rapid risk stratification in the postpartum setting.

Thus, the use of AI technologies in risk stratification for pregnant women with CVD is a promising area. However, their widespread implementation requires greater predictive accuracy, model standardization, and confirmation of their clinical efficacy in prospective studies.

CARDIAC ARRHYTHMIAS DURING PREGNANCY AND THE ROLE OF ARTIFICIAL INTELLIGENCE IN ARRHYTHMIA ASSESSMENT

Physiological changes during pregnancy place substantial stress on the cardiovascular system and can provoke both the onset of arrhythmias and the exacerbation of pre-existing arrhythmias. Women with congenital heart disease, cardiomyopathies, and pre-existing arrhythmias are at the highest risk. Supraventricular tachycardias occur most frequently during pregnancy, while ventricular arrhythmias are less frequent. According to the ROPAC (Registry of Pregnancy and Cardiac Disease) registry, their frequency during pregnancy in women with CVD is 1.7% and 1.6%, respectively [9].

Management of arrhythmias in pregnant women has specific features and differs from standard protocols used in non-pregnant women. Particular attention is given to the safety of both the mother and the fetus, which limits the use of certain antiarrhythmic drugs because of their potential teratogenic effects. In cases of refractory arrhythmias, cardioversion and catheter ablation remain effective and relatively safe treatment options [1, 16].

Diagnosis of arrhythmias in pregnant women is traditionally based on recording a standard electrocardiogram (ECG) or Holter monitoring. These methods are simple, widely available, and non-invasive; however, their interpretation may vary depending on the physician’s experience and expertise [17]. At this stage, AI holds considerable promise for standardizing and improving the accuracy of ECG interpretation.

Despite the scarcity of studies on the use of AI technologies in the diagnosis of arrhythmias in pregnant women, numerous works in non-pregnant population demonstrate promising results [14, 18–21]. Neural networks have shown the ability to detect arrhythmias from both 12-lead and single-lead ECG recordings with accuracy comparable to, or even exceeding, that of experienced cardiologists [14]. Moreover, models have been developed that are capable of predicting the development of atrial fibrillation from a standard resting ECG recorded during sinus rhythm [14, 18, 19]. These capabilities have been confirmed both in the analysis of routine 12-lead ECGs and during short-term ambulatory monitoring or the use of home devices for recording single-lead ECGs [20, 21]. Thus, AI can identify patients at high risk of developing clinically significant arrhythmias even in the absence of overt manifestations.

The potential of this approach is increasing with the growing adoption of wearable devices that monitor heart rhythm using photoplethysmography. Results from large-scale projects such as the Huawei Heart Study [22] and the Apple Heart Study [23] have shown that wearable devices can effectively detect atrial fibrillation, thereby enabling new approaches to remote screening and monitoring. Applying AI models to data derived from wearable devices may further improve diagnostic sensitivity and specificity while enabling more reliable long-term monitoring.

Although studies on the use of AI to detect arrhythmias in pregnant women are still lacking, the findings from the general population indicate that this is a promising field (Table 1). In the future, the integration of AI with wearable devices and standard diagnostic methods could represent an important step toward personalized monitoring of pregnant women at high risk of arrhythmias and improving outcomes for both the mother and the fetus.

 

Table 1. Cardiac Arrhythmias During Pregnancy and Potential Applications of Artificial Intelligence

Condition/Technology	Frequency/characteristic	Diagnostic Methods	Potential role of artificial intelligence	Reference
Supraventricular tachycardia	~ 1.7% in pregnant women with CVD	Standard electrocardiography, Holter monitoring	Can automate ECG interpretation and detect occult arrhythmias	[9]
Ventricular arrhythmias	~ 1.6% in pregnant women with CVD	Electrocardiogram, inpatient monitoring	Machine learning models can predict life-threatening arrhythmias	[9]
12-lead ECG	A basic, widely available, noninvasive method	Interpretation varies between specialists	Neural networks can detect arrhythmias with accuracy comparable to or exceeding that of experienced cardiologists	[14, 17]
Prediction of atrial fibrillation	Risk can be identified even from a normal ECG recorded during sinus rhythm	Deep learning algorithms identify latent patterns	Early risk stratification for atrial fibrillation	[14, 18 19]
Outpatient and home ECG Monitoring (Lead I)	Used for long-term monitoring	ECG recorders, portable devices	Identifies patients at high risk of subclinical atrial fibrillation	[20, 21]
Wearable devices with photoplethysmography	Widespread use; a convenient self-monitoring tool	Heart rate and rhythm monitoring via smartwatches and wristbands	Improves sensitivity and specificity for atrial fibrillation detection	[22, 23]

 

HEART FAILURE DURING PREGNANCY AND THE POTENTIAL ROLE OF ARTIFICIAL INTELLIGENCE

Heart failure during pregnancy is one of the most serious complications in obstetric practice. Its development is caused by a combination of congenital and acquired heart disease, and unique physiological changes of pregnancy. In some women, heart failure first becomes clinically apparent during pregnancy, often in association with previously undiagnosed cardiomyopathies. In others, the hemodynamic burden of pregnancy, including increased circulating blood volume and cardiac output, together with hormonal and immunologic changes, may precipitate decompensation of pre-existing cardiovascular disease [24].

Two forms of heart failure encountered predominantly in obstetric practice deserve special attention.

• Peripartum cardiomyopathy is a form of systolic heart failure associated with left ventricular dysfunction and reduced left ventricular ejection fraction (< 45%) that develops during the last month of pregnancy or within 5 months after delivery. The disease is often severe and is associated with a high risk of life-threatening arrhythmias, thromboembolic events, and progression of heart failure [25].
• Heart failure with preserved ejection fraction is characterized by the symptoms of heart failure despite a normal or near-normal left ventricular ejection fraction. It is most commonly associated with hypertension, obesity, and preeclampsia [24, 26].

Clinical management of patients with heart failure during pregnancy is complicated by the limited therapeutic options because of the potential teratogenic effects of medications. Treatment is based on diuretics to reduce volume overload, β-blockers to control heart rate and reduce myocardial oxygen demand, digoxin for systolic dysfunction, and hydralazine with nitrates to reduce afterload in severe cases [27]. Despite the availability of these therapeutic approaches, timely diagnosis remains critical, as it allows early detection of peripartum cardiomyopathy and prompt initiation of treatment before irreversible complications develop [1, 2, 5].

Traditional diagnosis of heart failure is based on clinical assessment, echocardiography, and measurement of natriuretic peptide levels [1, 28]. However, symptoms such as shortness of breath, fatigue, and edema often mimic normal manifestations of pregnancy, while laboratory markers and echocardiography are not always available, especially in resource-limited regions [1, 4, 28]. These limitations have fueled growing interest in AI-based technologies, which can uncover latent patterns in electrocardiographic data and predict impaired myocardial contractile function [2, 5, 6].

Several recent studies have shown the high efficacy of AI in the diagnosis of peripartum cardiomyopathy [29, 30]. For example, D. A. Adedinsewo et al. [29] analyzed the ECG data from 1807 women and compared them with echocardiographic findings. The deep learning model enabled identification of left ventricular ejection fraction ≤ 35% (ROC-AUC = 0.92), demonstrating higher accuracy than both a model based on natriuretic peptide levels and a multivariable model incorporating demographic and clinical parameters (ROC-AUC = 0.85). In addition, the model predicted left ventricular ejection fraction < 45 and < 50%, with ROC-AUC values of 0.89 and 0.87, respectively. In the study by Y. Lee et al. [30], the authors used more than 58000 pairs of ECGs and echocardiographic findings to train a model that achieved an ROC-AUC of 0.87 on external validation among 157 pregnant and postpartum women.

Studies evaluating the applicability of simplified recording methods are also of interest. For example, Y. M. Jung et al. [31] evaluated the performance of AI-based software for analysis of both 12-lead and single-lead (lead I) ECGs for screening of peripartum cardiomyopathy. Among 204 women, the model demonstrated high diagnostic accuracy: ROC-AUC was 0.979 and 0.944, respectively, highlighting the potential of these technologies in wearable devices. I. Karabayir et al. [32] developed a model using ECG-derived data from non-pregnant women and validated it in two cohorts of pregnant women. Despite the differences in the samples, the model showed comparable accuracy (ROC-AUC from 0.73 to 0.94), supporting the generalizability of the approach.

Particular attention should be paid to the innovative method proposed by D. A. Adedinsewo et al. [33]. The authors used a digital stethoscope to simultaneously record a single-lead ECG and a phonocardiogram in 100 women during pregnancy and the postpartum period. The AI model achieved an ROC-AUC of 1.0 when analyzing 12-lead ECGs and 0.987 when analyzing single-lead ECGs combined with phonocardiography. This approach enables bedside diagnosis without the need for complex equipment.

An important milestone was the first randomized clinical trial conducted in Nigeria [34]. It included 1232 women randomized into two groups: the first received standard obstetric care, whereas the second received standard care supplemented by AI-based technology, including ECG analysis, digital stethoscope evaluation, and echocardiography when indicated. Use of AI-based technologies nearly doubled the detection rate of peripartum cardiomyopathy compared with standard practice; however, differences between AI-assisted ECG interpretation and conventional clinical assessment did not reach statistical importance.

Overall, the available evidence suggests that AI models demonstrate high diagnostic accuracy (ROC-AUC > 0.9) for detecting peripartum cardiomyopathy and left ventricular dysfunction [31, 34]. Their use is particularly promising in resource-limited settings and with simplified technologies such as single-lead ECG, digital stethoscopes, and wearable devices (Table 2). However, the integration of AI into clinical algorithms requires large multicenter studies, standardization of methodologies, and assessment of their impact on maternal and infant outcomes.

 

Table 2. Heart Failure During Pregnancy and the Role of Artificial Intelligence

Authors	Year	Study Design	Sample Size	Method	Main Findings	ROC-AUC
D. A. Adedinsewo et al. [29]	2021	Retrospective study	1807 women; ECG and echocardiography	Deep learning analysis of 12-lead ECGs	Identification of LVEF ≤ 35%	0.92
Y. Lee et al. [30]	2022	Retrospective study	58350 paired ECG and echocardiographic datasets; external validation in 157 women	Neural network for LVEF prediction	Diagnosis of PPCM	0.87
Y. M. Jung et al. [31]	2023	Retrospective study	204 women	AI-based analysis of 12-lead and single-lead ECGs	Screening for PPCM	0.979/0.944
I. Karabayir et al. [32]	2024	Single-center retrospective study	Two cohorts of pregnant women; model trained on non-pregnant women	12-lead ECG	Diagnosis of PPCM	0.73–0.94
D. A. Adedinsewo et al. [33]	2024	Prospective study	100 women	12-lead and single-lead ECG; digital stethoscope	Identification of LVEF < 50%	1.0/0.987
D. A. Adedinsewo et al. [34]	2024	Randomized clinical trial	1232 women	12-lead ECG; digital stethoscope	Higher detection rate of PPCM than with standard practice	>0.9

Note. PPCM, peripartum cardiomyopathy; LVEF, left ventricular ejection fraction; ECG, electrocardiogram; AI, artificial intelligence; ROC-AUC, area under the receiver operating characteristic curve, a measure of the diagnostic or prognostic accuracy of a model.

 

VALVULAR HEART DISEASE AND THE POTENTIAL ROLE OF ARTIFICIAL INTELLIGENCE

Valvular heart disease account for a substantial proportion of cardiovascular disorders in pregnancy, representing approximately 27% of all cases of maternal heart disease [35]. The etiology varies by region: in high-income countries, congenital valvular abnormalities predominate [3, 10], whereas in low- and middle-income countries, rheumatic heart disease continues to play a leading role [1, 3, 10].

Severe valvular diseases are associated with a high risk of both maternal and perinatal complications. In stenotic lesions, such as mitral or aortic stenosis, fixed obstruction to blood flow limits the capacity to increase cardiac output, leading to elevated atrial pressures, pulmonary hypertension, pulmonary edema, and arrhythmias as early as the second trimester of pregnancy. In regurgitant lesions, particularly mitral or aortic regurgitation, the clinical course is generally more favorable because of the reduction in afterload during pregnancy; however, severe regurgitation may still lead to cardiac decompensation during late pregnancy [28].

The management of pregnant women with valvular heart disease largely depends on the severity of the lesion. In most cases, patients can be managed medically with diuretics, β-blockers, and antiarrhythmic agents; however, life-threatening conditions may require intervention, including balloon valvuloplasty or transcatheter aortic valve implantation during pregnancy, or surgical treatment in the postpartum period [1, 28].

In recent years, the role of AI in the diagnosis and treatment of valvular heart disease has attracted increasing attention [36, 37]. In the general population, AI has already demonstrated substantial advantages: AI-assisted interpretation of phonocardiograms obtained using digital stethoscopes was more than twice as sensitive as primary care physicians in detecting valvular heart disease [36]. Furthermore, AI has shown high efficacy in the test of echocardiographic findings. G. Holste et al. [37] developed a model capable of diagnosing severe aortic stenosis based on two-dimensional echocardiographic images without the need for Doppler imaging.

Although such developments have not yet been systematically tested in pregnant women, they offer substantial potential for obstetric practice (Table 3). We believe that applying AI technologies to the analysis of auscultatory findings and echocardiographic data in pregnant women may standardize interpretation, improve diagnostic accuracy, and facilitate early detection of critical valvular heart disease. This, in turn, may improve risk stratification and optimize management decisions, including timely referral to a specialized center.

 

Table 3. Valvular Heart Disease in Pregnant Women and the Potential Role of Artificial Intelligence

Type of Lesion	Hemodynamic changes during pregnancy	Risk of complications	Clinical management	Potential role for artificial intelligence
Mitral stenosis	Fixed obstruction to blood flow, increased left atrial pressure, and risk of pulmonary hypertension	Pulmonary edema, arrhythmias, and heart failure as early as the second trimester	diuretics, β-blockers; in severe cases, balloon valvuloplasty	AI-assisted interpretation of phonocardiograms for early detection of stenosis; analysis of echocardiographic findings to assess severity
Aortic stenosis	Limited increase in cardiac output and increased left ventricular load	Heart failure, syncope, and risk of sudden death	medical therapy; in critical cases, transcatheter aortic valve implantation	Analysis of echocardiographic findings for automated diagnosis of severe aortic stenosis
Mitral regurgitation	Left ventricular volume overload, partially compensated by reduced afterload	Decompensation later in pregnancy; arrhythmias	conservative treatment; surgical correction usually in the postpartum period	Echocardiographic analysis for longitudinal monitoring of disease progression
Aortic regurgitation	Left ventricular volume overload, usually better tolerated because of reduced afterload	In severe cases, heart failure and adverse perinatal outcomes	medical therapy; surgical intervention after delivery	Standardization of echocardiographic interpretation and detection of occult left ventricular dysfunction

 

Thus, AI technologies may eventually become an important adjunct to conventional diagnostic methods for valvular heart disease in pregnant women, helping improve quality of care and reduce the risk of adverse maternal and fetal outcomes.

HYPERTENSIVE DISORDERS DURING PREGNANCY AND THE POTENTIAL ROLE OF ARTIFICIAL INTELLIGENCE

Hypertensive disorders of pregnancy are among the leading causes of maternal and perinatal morbidity and mortality, complicating 6%–8% of all pregnancies [26]. They include chronic hypertension, gestational hypertension, preeclampsia, eclampsia, and HELLP syndrome. Although these conditions differ in their underlying pathophysiology, they are all associated with a high risk of adverse maternal and fetal outcomes [1, 4]. In mothers, hypertensive disorders during pregnancy are associated with cardiovascular decompensation, stroke, and multiorgan failure, whereas in the fetus they are associated with intrauterine growth restriction, preterm delivery, and antenatal death [6, 26].

Traditional diagnostics include blood pressure measurement using a sphygmomanometer, laboratory testing, and ultrasound examination. However, these methods have limited predictive value. AI has opened new perspectives in the monitoring and prediction of hypertensive disorders during pregnancy. Blood pressure prediction using surrogate signals, including photoplethysmography and single-lead ECG, has been demonstrated, with these modalities offering the potential for integration into wearable devices for continuous monitoring [38, 39]. Furthermore, models have been developed to predict individual response to antihypertensive therapy, paving the way for personalized treatment [40, 41].

Studies on ECG-based diagnosis of preeclampsia are of particular interest. For example, L. Butler et al. [42] trained a neural network using data extracted from short-duration 12-lead ECG recordings (904 recordings) and validated it in an independent cohort (817 ECG recordings). The model demonstrated high accuracy, with ROC-AUC values of 0.85 and 0.81, respectively, including analysis of recordings obtained 30–90 days before diagnosis and in cases of early-onset preeclampsia (< 34 weeks). This finding is likely explained by alterations in cardiac electrical activity during preeclampsia, including prolongation of the QT/QTc intervals, increased interval dispersion, and changes in P-wave morphology, which reflect repolarization abnormalities and electrical remodeling associated with endothelial dysfunction [43, 44].

Another approach is the integration of demographic characteristics and biomarkers into prognostic models. For example, M. M. Gil et al. [45] used data from more than 30000 pregnant women, including uterine artery pulsatility index, mean arterial pressure, placental growth factor, and pregnancy-associated plasma protein-A. At a false positive rate of 10%, the model detected early preeclampsia in 75.3% of cases (ROC-AU = 0.909). Validation in the Spanish PREVAL study, which included 10110 women, showed comparable efficacy, with detection rates of 84.4% for early-onset preeclampsia (< 34 weeks; ROC-AUC = 0.920), 77.8% for preterm preeclampsia (< 37 weeks; ROC-AUC = 0.913), and 55.7% for all cases of preeclampsia (ROC-AUC = 0.846). These findings were comparable to those obtained with the Fetal Medicine Foundation competing-risks model; however, AI showed superior accuracy in predicting early-onset preeclampsia.

Thus, AI possesses substantial potential in the prediction and diagnosis of hypertensive disorders during pregnancy:

• identification of subclinical ECG changes long before clinical onset;
• integration of demographic and biochemical data for risk stratification;
• use of wearable devices for continuous monitoring;
• prediction of individual response to therapy.

Although most models remain at the validation stage, current findings support the potential integration of AI into early screening algorithms and surveillance strategies for pregnant women at high risk [9, 43, 44].

ACUTE CORONARY SYNDROME DURING PREGNANCY AND THE ROLE OF ARTIFICIAL INTELLIGENCE

Acute coronary syndrome during pregnancy is relatively rare, accounting for less than 2% of all cardiovascular complications in this group [46]. Unlike the general population, in which atherosclerosis is the primary cause of acute coronary syndrome, spontaneous coronary artery dissection (SCAD) is the leading cause in pregnant and postpartum women, accounting for more than 40% of cases [47]. Increased vascular wall vulnerability during this period is associated with hormone-induced changes in connective tissue under the influence of estrogen and progesterone, as well as increased hemodynamic stress [48]. Known risk factors for SCAD include connective tissue disorders and a history of prior dissection; however, in a substantial proportion of patients, SCAD occurs in the absence of traditional predictors [49].

Its early and accurate diagnosis is critical for selecting the appropriate treatment strategy. Although coronary angiography remains the "gold standard", its two-dimensional images do not always allow for the detection of intramural hematoma or atypical forms of dissection, which can lead to diagnostic errors. Intravascular ultrasound and optical coherence tomography have higher sensitivity; however, their use is limited by invasiveness, technical difficulties, and increased risk during pregnancy [50]. These limitations underscore the need for improved interpretation of coronary angiography.

In recent years, AI-based quantitative coronary angiography technologies have been increasingly developed to automate image analysis. The DeepDiscern model, presented by J. Chae et al. [51], demonstrated an accuracy of 87.6% in detecting coronary abnormalities, including dissections. Y. Kim et al. [52] showed that AI-based quantitative coronary angiography was non-inferior to optical coherence tomography guidance for stent optimization during percutaneous coronary intervention, suggesting that it may serve as a potential alternative in certain clinical scenarios. Although these approaches have not yet been specifically tested for the diagnosis of SCAD in pregnant women, they may represent a promising avenue for improving diagnostic accuracy and facilitating timely intervention.

AI is also useful for predicting the risk of adverse outcomes. In a retrospective study by C. Krittanawong et al. [53] including 375 patients with SCAD (mean age, 52.2 years; 64.3% women), the mortality rate was 11.5%. The deep learning model demonstrated substantially higher predictive accuracy (ROC-AUC = 0.98) than traditional methods, including logistic regression. Key predictors of increased mortality included elevated C-reactive protein levels, atrial fibrillation, hypertension, and steroid use. These data open opportunities for developing specialized risk models for pregnant women with SCAD, where risk factors may differ.

Thus, acute coronary syndrome during pregnancy, despite its rarity, is of considerable clinical importance and is characterized by distinct pathophysiologic features. Integration of AI technologies into analysis of coronary angiographic images and outcome prediction represents a promising area that may improve diagnostic accuracy, reduce the risk of missed SCAD, and improve outcomes for both the mother and the fetus.

PULMONARY HYPERTENSION DURING PREGNANCY AND THE POTENTIAL ROLE OF ARTIFICIAL INTELLIGENCE

Pulmonary hypertension is a severe, life-threatening condition characterized by persistent elevation of pulmonary arterial pressure, resulting in right ventricular overload and progressive right heart failure. Pregnancy in women with pulmonary hypertension is classified as risk class IV according to the modified WHO classification and is considered contraindicated because of the high risk of maternal and perinatal mortality. According to a systematic review by R. Ma et al. [54], the mean maternal mortality rate among women with moderate-to-severe pulmonary hypertension is approximately 9%.

Nevertheless, if pregnancy occurs and is continued, therapeutic options are limited because of the teratogenic effects of certain drugs Endothelin receptor antagonists and soluble guanylate cyclase stimulators are contraindicated during pregnancy. In contrast, phosphodiesterase-5 inhibitors may be used to reduce pulmonary arterial pressure, whereas prostacyclin analogs such as epoprostenol¹ and treprostinil², owing to their potent vasodilatory effects, are considered relatively safe an may be used for hemodynamic control in pregnant women [13].

Outside pregnancy, AI-based approaches to the diagnosis and screening of pulmonary hypertension are being actively investigated. M. Elgendi et al. [55] demonstrated highly accurate detection of pulmonary hypertension through analysis of phonocardiograms recorded with a digital stethoscope; the deep learning model identified a specific acoustic pattern (signature) associated with elevated pulmonary arterial pressure. J. Kwon et al. [56] reported strong performance of a neural network model in the analysis of both single-lead and 12-lead ECGs, with AI detecting features of pulmonary hypertension that may be missed on conventional visual interpretation. Furthermore, machine learning models applied to echocardiographic data were able to automate estimation of the probability of pulmonary hypertension and were more accurate than subjective physician assessment [57]. In addition, AI-based analysis of chest radiographs also demonstrated potential for early detection of structural features suggestive of pulmonary hypertension [58].

Collectively, these findings highlight the broad range of AI-based approaches that could be integrated into clinical practice for early detection and monitoring of pulmonary hypertension. In obstetric practice, these approaches are particularly promising because they enable the use of noninvasive and readily available modalities, including electrocardiography, phonocardiography, and radiography, for screening women at high risk [55–57].

Thus, despite limited therapeutic options during pregnancy, integration of AI technologies into pulmonary hypertension diagnosis may improve the accuracy of early detection, facilitate timely management decisions, and potentially improve maternal and fetal outcomes.

LIMITATIONS OF ARTIFICIAL INTELLIGENCE FOR CARDIOVASCULAR DISEASE IN PREGNANCY

Data Quality and Bias

Despite its substantial potential, integration of AI into medicine is associated with several important challenges. One of the major barriers remains the quality of the underlying data. Development of reliable models requires large, representative datasets derived primarily from electronic health records. However, these data are often fragmented, incomplete, and heterogeneous across healthcare systems, thereby undermining model accuracy and reproducibility. Moreover, electronic health records reflect existing systemic inequalities, including disparities in access to care and quality of care across social, ethnic, and demographic groups. Models trained on such data risk not only reproducing these biases but also amplifying them, potentially leading to inequitable clinical recommendations and adversely affecting vulnerable patient populations [59].

Interpretability, Trust, and Regulation

Another major challenge is model interpretability. Most AI models function as “black boxes”, in which the pathway from input data to prediction remains opaque [59]. This undermines physician trust, as clinicians may be reluctant to rely on system-generated recommendations without a clear understanding of their underlying rationale. The lack of a robust regulatory framework also hinders implementation: standardized validation procedures, post-market surveillance, and legal mechanisms for assigning responsibility in the event of adverse outcomes remain insufficiently developed [1, 6].

Patients also express concerns regarding both the use of AI in clinical decision-making and the use of their personal data for model training. Protecting confidentiality and ensuring transparent consent mechanisms remain major ethical priorities [1, 59].

FUTURE DIRECTIONS

Screening

To date, no standardized screening protocols for CVD in asymptomatic pregnant women have been established. The California Maternal CVD Screening Algorithm³ currently used in the United States provides a structured approach but does not integrate the capabilities of modern digital technologies. At the same time, studies have shown that AI can detect cardiac abnormalities from ECG data. We believe that, if its performance is validated in pregnant women, AI-based analysis of a short ECG recording could become an accessible and scalable method for early screening. This is particularly important for women traditionally considered to be at low risk who nonetheless develop severe complications, such as peripartum cardiomyopathy or SCAD.

Imaging

Radiography and echocardiography are among the most promising imaging modalities for AI applications, given the large volume of standardized data available. AI models have already demonstrated strong performance in automated segmentation, structural recognition, and pathology detection using magnetic resonance imaging and echocardiographic data [60, 61]. These technologies can be readily adapted for pregnant women, as imaging protocols in this population do not differ from those used in the general population. We believe that integration of AI may accelerate diagnostic reporting and improve interpretive accuracy, which is particularly important for timely detection of complications during pregnancy.

Remote Monitoring

AI-enabled wearable devices create new opportunities for outpatient monitoring. These systems have been shown to predict heart failure decompensation [15] and detect atrial fibrillation [62]. This capability is particularly valuable during pregnancy, as arrhythmias are often asymptomatic or intermittent [16]. We believe that continuous AI-assisted ECG monitoring may facilitate early detection of arrhythmias, reduce the need for hospital-based monitoring, and alleviate the burden on the healthcare system.

Table 4 summarizes the challenges and future directions for AI in CVD during pregnancy.

 

Table 4. Challenges and Future Directions for Artificial Intelligence in Cardiovascular Disease During Pregnancy

Clinical relevance	Barriers	Potential solutions
Data Quality and Bias
Diagnostic and prognostic errors; risk of inequitable clinical recommendations	Fragmented electronic health records; social and ethnic disparities	Development of unified registries, data standardization, and quality control
Interpretability of Algorithms
Physician distrust; limited implementation in clinical practice	AI functions as a “black box”	Development of interpretable AI methods; physician training
Regulatory Framework
Legal liability and patient safety risks	Lack of standardized validation procedures and post-market surveillance	Development of a legal framework and international standards for AI evaluation
Ethical and Legal Issues
Reduced patient trust; risk of data breaches	Privacy and informed consent concerns	Data protection laws; transparent consent mechanisms
Screening
Identification of women at latent risk	Lack of universal screening protocols	Use of AI for analysis of ECGs and biomarkers
Imaging
Timely diagnosis of complications	High specialist workload; delays in reporting	Automated segmentation and AI-based image analysis
Remote Monitoring
Early detection of arrhythmias and heart failure in pregnant women	Limited access to inpatient monitoring	AI-enabled wearable devices; outpatient monitoring

 

CONCLUSION

AI offers new opportunities for the diagnosis, risk stratification, and monitoring of CVD during pregnancy. Emerging evidence highlights its potential to improve the accuracy of early detection and optimize referral pathways for affected patients. At the same time, safe and equitable integration of these technologies into clinical practice will require further research, rigorous clinical validation, and robust regulatory oversight.

Integration of AI technologies into obstetric and cardiovascular practice may reduce morbidity and mortality, improve the quality of care, and lower healthcare costs. Realizing this potential requires interdisciplinary collaboration among specialists in cardiology, obstetrics and gynecology, medical informatics, bioethics, and law.

Thus, AI may become an important tool in protecting maternal and fetal health; however, its widespread use must be accompanied by sound scientific justification, ethical review, and the development of an effective regulatory framework.

ADDITIONAL INFORMATION

Author contributions: Yu.A. Trusov: conceptualization, data curation, writing — original draft, supervision; Kh.T. Shamsueva: conceptualization, data curation, visualization; M.Z. Kolkhidova: data curation, writing — original draft; A.T. Inderbieva: data curation, writing — original draft; E.D. Baryshnikova, S.A. Sanakoev: writing — original draft, writing — review & editing; K.A. Khusnutdinova: data curation (English), writing — original draft; A.E. Rasponomareva, A.R. Shabazgerieva: writing — original draft; Kh.K. Radzhabov: formal analysis, writing — review & editing; V.Kh. Kudzieva, A.V. Ponomareva: visualization, writing — review & editing; N.K. Kozyreva: writing — review & editing. All the authors approved the version of the manuscript to be published and agreed to be accountable for all aspects of the work, ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Ethics approval: Not applicable.

Funding sources: No funding.

Disclosure of interests: The authors have no relationships, activities, or interests for the last three years related to for-profit or not-for-profit third parties whose interests may be affected by the content of the article.

Statement of originality: No previously published material (text, images, or data) was used in this study or article.

Data availability statement: The editorial policy regarding data sharing does not apply to this work.

Generative AI: Generative artificial intelligence technologies (ChatGPT 4.0, OpenAI) were used during manuscript preparation. The use of artificial intelligence was limited to language editing and the identification of stylistic and spelling errors. Data analysis and interpretation were performed by the authors; all scientific conclusions presented in the manuscript are the sole responsibility of the authors.

Provenance and peer-review: This article was submitted unsolicited and reviewed following the fast-track procedure. The peer-review process involved two external reviewers and member of the Editorial Board.

¹ This medicinal product is not registered in the Russian Federation.

² This medicinal product is not registered in the Russian Federation.

³ CMQCC and California Department of Public Health. Cardiovascular Disease Assessment in Pregnant and Postpartum Women [Internet]. Palo Alto: CMQCC and California Department of Public Health; 2017. Available from: https://www.cmqcc.org/resource/cardiovascular-disease-assessment-pregnant-and-postpartum-women. Accessed November 21, 2025.

About the authors

Yurii A. Trusov

Samara State Medical University

Email: yu.a.trusov@samsmu.ru
ORCID iD: 0000-0001-6407-3880
SPIN-code: 3203-5314
Russian Federation, Samara

Khadizhat T. Shamsueva

North-Ossetian State Medical Academy

Author for correspondence.
Email: shamsuevakh1@gmail.com
ORCID iD: 0009-0003-7173-9072
Russian Federation, Vladikavkaz

Marianna Z. Kolkhidova

North-Ossetian State Medical Academy

Email: mari_kolxi@mail.ru
ORCID iD: 0009-0003-1301-8795
Russian Federation, Vladikavkaz

Albina T. Inderbieva

North-Ossetian State Medical Academy

Email: dr.inderbieva@mail.ru
ORCID iD: 0009-0007-8620-4790
Russian Federation, Vladikavkaz

Elizaveta D. Baryshnikova

The Russian National Research Medical University named after N.I. Pirogov (Pirogov University)

Email: mazhirinal2013@yandex.ru
ORCID iD: 0009-0007-6066-9354
Russian Federation, Moscow

Kamilya A. Khusnutdinova

The Russian National Research Medical University named after N.I. Pirogov (Pirogov University)

Email: Khusnutdinova.k@mail.ru
ORCID iD: 0009-0004-7935-9477
Russian Federation, Moscow

Aleksandra E. Rasponomareva

Professor V.F. Voino-Yasenetsky Krasnoyarsk State Medical University

Email: aleksandrarasp@mail.ru
ORCID iD: 0009-0008-3382-1493
Russian Federation, Krasnoyarsk

Alina R. Shabazgerieva

North-Ossetian State Medical Academy

Email: alyashabazgerieva@gmail.com
ORCID iD: 0009-0003-5687-7098
Russian Federation, Vladikavkaz

Khadzhimurad K. Radzhabov

Penza State University

Email: khadzhimurad.radzhabov@mail.ru
ORCID iD: 0009-0004-3169-951X
Russian Federation, Penza

Stanislav A. Sanakoev

North-Ossetian State Medical Academy

Email: stas.sanakoev-2018@mail.ru
ORCID iD: 0009-0003-1058-9226
Russian Federation, Vladikavkaz

Valeria Kh. Kudzieva

Rostov State Medical University

Email: kudzievavaleria@mail.ru
ORCID iD: 0009-0006-6371-6112
Russian Federation, Rostov-on-Don

Alina V. Ponomareva

Kuban State Medical University

Email: alipon4@yandex.ru
ORCID iD: 0009-0004-3936-7982
Russian Federation, Krasnodar

Natalia K. Kozyreva

Kuban State Medical University

Email: Nata05042004@yandex.ru
ORCID iD: 0009-0003-1517-0539
Russian Federation, Krasnodar

References

Supplementary files