AI Implementation in Obstetrics and Gynecology

Welcome to Health Connect, the podcast for health professionals where we will share the latest news and information on science and technology in the medical field. In this episode, we’ll give you an overview of the current developments in artificial intelligence for obstetrics and gynecology.

Did you know that novel deep learning and machine learning approaches have been successfully applied to predict patient survival using multi-omic data?¹

The integration of artificial intelligence algorithms and multi-omics, referring to the combination of two or more omics datasets, could create predictive or prognostic models for specific diseases. These models would be more accurate than those obtained through traditional clinical procedures.¹

This technology could enable the discovery of new biomarkers with great potential for disease prediction, patient stratification and precision medicine, which could be particularly useful in he face of the growing global trend towards reproductive complications, which is in part due to increasing maternal age and the prevalence of obesity, which requires closer monitoring of female reproductive health.¹

Despite the promise of predictive machine learning models, a limited number of efforts in this area have been made. For instance, the 2021 retrospective cohort study of Escobar and colleagues successfully predicted about 52 percent of obstetric complications. The researchers used machine learning on electronic health record data from approximately 304,000 deliveries.¹

This technology could be particularly important in low- and middle-income countries, having the highest rates of fetal and neonatal deaths. Using machine learning on cohort data from more than 500,000 pregnancies, the models of Shukla and colleagues achieved an acceptable predictive accuracy for both intrapartum fetal death and neonatal mortality. The area under the receiver operating characteristic curve was 0.71, with birth weight as the most significant predictor of neonatal mortality.¹ 

Another influential factor that needs to be considered when it comes to neonatal and perinatal deaths is preterm birth, the causes of which are still largely unknown. Thanks to the work of Jehan and colleagues in 2020, machine learning based on transcriptomic and proteomic profiling of plasma and metabolomic analysis of urine has made it possible to predict accurately preterm birth, with an area under the curve of 0.83.¹ 

Pre-eclampsia is yet another major pregnancy complication, which is currently not detectable before the onset of symptoms. However, changes between the first and early second trimester in maternal plasma FLT1, soluble endoglin and placental growth factor, in combination with clinical features have shown promise in predicting early-onset pre-eclampsia.¹ 

Recently, demonstrating the exciting potential of artificial intelligence in this area, a machine learning-based multi-omics model for pre-eclampsia risk achieved a high accuracy with an area under the curve of 0.94. This was done by analyzing six -omics datasets from a longitudinal cohort of pregnant women in which the risk of pre-eclampsia was examined.¹

All these ongoing investigations show us various clinical challenges, where applying solutions based on artificial intelligence could be decisive in predicting pregnancy disorders and complications. Currently, according to the systematic review led by Stepan Feduniw in 2022, various investigations on the application of these technologies are focused on defined obstetric areas: predicting general risk characteristics of pregnancy, prenatal diagnostics, hypertensive disorders of pregnancy, fetal growth, stillbirth, gestational diabetes, preterm birth, and delivery route.²

According to these findings, the best application of artificial intelligence to assess medical conditions are artificial neural networks, the most popular form of artificial intelligence applied in medicine, with an average accuracy of 80 to 90 percent. These types of models are inspired by the animal nervous system and can learn and analyze imprecise information, allowing the analysis of large amounts of medical data and assisting in diagnostic and treatment decisions.²

Focusing on predicting the overall risks during pregnancy, it is interesting to consider the different predictors used in constructing the five studies evaluated by Feduniw’s team. The most relevant were:²

● Clinical parameters and existing health conditions of the mother, including maternal and gestational age, pregnancy and parity, body mass index, and weight gain.
● Vital parameters such as blood pressure, lifestyle, and diet.
● Pregnancy-related complications, including threatened preterm delivery, hemorrhage during pregnancy, pregestational and gestational diabetes, various hypertensive and hepatic disorders, cholestasis, thrombophilia, autoimmune diseases, an age greater than 35 years, multiple pregnancies, and a parity interval of 10 years or more.
● Laboratory parameters.
● Fetal monitoring parameters such as basal fetal heart rate and variability.
● Ultrasound and Doppler parameters. 

However, to increase the representativeness and accuracy of the model, many other variables can be used to measure the overall risk of pregnancy. Collecting these potential risk factors and using trained and validated software could lead to an early diagnosis of pregnancy complications or even prevent them.²

While today’s decisions are based on the experience and expertise of physicians, the human brain is unfit for processing large amounts of data from prospective studies to calculate the exact risk factor in each medical case, and it is in this scenario that digital software could help improve the prediction of medical conditions.²

Now, let's look at a different approach to predicting preterm delivery, proposed by the team led by Sunwha Park, Jeongsup Moon, and Nayeon Kang in 2022. The authors aimed to select markers by analyzing vaginal microbiome data of pregnant women and develop a machine learning-based model with a high prediction rate by combining clinical information.³

In this multicenter case-control study, 150 singleton pregnant women with a gestational age between 17 and 32 weeks were recruited between 2018 and 2020; 54 were assigned to the preterm delivery group and 96 to the full-term delivery group. Subjects diagnosed with gestational diabetes mellitus, pre-eclampsia, or inadequate medical history were excluded.³ 

Clinical characteristics and demographic profiles were then recorded, including variables such as age, body mass index, educational level, and history of preterm delivery. White blood cell count and cervical length were also recorded. To analyze the microbiome profiles, cervicovaginal fluid was collected at mid-pregnancy.³

To build the machine learning model, participants were randomly divided into training and test groups in a two-to-one ratio. Univariate analysis was performed to select markers to predict preterm delivery using seven statistical tests. The researchers also employed an evidence-based medicine method through the existing literature review to select markers related to preterm birth that could adequately complement the machine learning method. The final markers from the microbiome included: Atopobium vaginae, Gardnerella vaginalis, Lactobacillus crispatus, Lactobacillus fornicalis, Lactobacillus gasseri, Lactobacillus iners, Lactobacillus jensenii, Peptoniphilus grossensis, Prevotella timonensis, and Ureaplasma parvum.³

The test area under the curve of the logistic regression model was 0.72 when the model was trained with 17 of the selected markers. However, the random forest model had the highest area under the curve of 0.84 when combining white blood cell count and cervical length with the seven markers of the vaginal microbiome. To further develop the predictive models, the authors proposed an improved version of a single-tree model called "GUIDE", which can still be used for interpretability with improved performance of a test area under the curve of 0.77.³

Cervical length measurement was performed to predict the risk of preterm birth in the second trimester. As we know, if the length of the cervix is less than 25 millimeters, it is considered high risk, and if it is less than 15 millimeters, hospitalization is recommended. Remarkably, using ten selected markers, cases with a cervical length shorter than 17.5 millimeters highly correlated with preterm birth. On the other hand, in cases where Ureaplasma and Prevotella increased, and the participant had a cervical length greater than 17.5 millimeters, there was a trend toward preterm delivery.³

Another relevant finding was that Prevotella timonensis, Ureaplasma parvum and Peptoniphilus lacrimalis played the most important role in preterm birth. According to previous studies, if bacterial diversity is high, the distribution of other pathogens increases, which translates into a higher risk of preterm birth.³

As this study aimes to develop a non-invasive method, the possibility of developing a predictive model for preterm birth by using the results of the microbiome analysis, the results of blood tests and the measurement of cervical length are very promising. Moreover, it showed better predictive power than the existing method for predicting preterm delivery.³

When it comes to gynecological imaging, artificial intelligence is a relatively understudied area compared to breast imaging.⁴

Although gynecologic cancers have a lower incidence than breast cancer, they have higher rates of morbidity and mortality.⁴

In a systematic review published in 2022 comparing the two fields, the use of artificial intelligence in breast cancer imaging yielded 767 studies dating back to the 1990s, whereas gynecologic cancer imaging yielded only 194 studies, mostly dated in the last 2 years.⁴

Thus, although an area of growing interest, gynecologic imaging has been relatively neglected in the field of artificial intelligence applied to imaging in women's health.⁴

First, let's talk about the importance of artificial intelligence in the field of diagnostic imaging and how it enhances, rather than replaces, the work of clinicians. Conventional diagnostic imaging relies on visual pattern recognition by a skilled radiology technician to make inferences from multiple data sources. Artificial intelligence technology helps standardize and streamline this process, introducing more advanced, faster and more reliable methods.⁴

Although fully automated cancer diagnosis generally remains beyond our reach, AI tools for characterizing malignancies with radiomics, which we can define as quantitative methods, standardize measurements and reporting, and improve sensitivity. These improvements can be extremely useful to the practicing radiologist.⁴

For example, Shrestha and colleagues reviewed 61 articles related to gynecologic cancers, including endometrial, cervical, ovarian, uterine, and vulvar cancers. Most studies used radiomic methods based on magnetic resonance imaging, and most used classical machine learning models.⁴

Let's start with high-risk endometrial cancer, where early and timely diagnosis is of the essence. Preoperative imaging risk stratification can help select patients at high risk of metastasis who may benefit from oncologic referral. However, magnetic resonance imaging currently has limitations for staging compared to pathology, which may be related to subtle textural patterns that go unnoticed by the human eye.⁴

Not surprisingly, most of the studies reviewed by the team aimed to classify uterine masses as malignant or benign, or into high- and low-grade groups. Created artificial intelligence models claimed similar or even higher diagnostic accuracy and net clinical benefit compared to experienced radiologists.⁴ 

For example, in the work led by Yan in 2021, the team built a deep learning model for the preoperative diagnosis of pelvic lymph node metastases, and further demonstrated that radiomics correlates with immunohistochemical indices such as estrogen receptor, progesterone receptor, P53, and Ki-67.⁴

Meanwhile, also in 2021, Jacob and colleagues developed an MRI-based radiomic model for predicting survival in patients with endometrial cancer. The researchers examined the model's association with gene expression profiles and clinical parameters for better prediction of disease outcomes. They found that 46 genes were associated with worse prediction of disease survival as well as a significant association with advanced stage disease, deep myometrial invasion, lymph node metastases, and worse 5-year survival.⁴ 

Relatedly, in 2020, Bereby-Kahane and colleagues analyzed MRI-based texture features in correlation with tumor parameters and found that the best predictor of higher-risk cancers was a short axis greater than 20 millimeters.⁴

In addition, other studies have incorporated known clinical factors such as obesity, hormone therapy, race, menstrual and reproductive history, and endometrial hyperplasia into their radiomic models. Most studies demonstrated better clinical prediction than radiomics alone.⁴ 

One last example is the study by Yan and colleagues, a large multicenter study of patients with histopathologically confirmed endometrial cancer. The study combined selected radiomic features with clinical factors to create a preoperative endometrial cancer prediction nomogram that improved surgical planning in 11 to 15% of patients compared with actual surgery. ⁴

Cápsula
The reliance on the human annotation of image data for primary cancer and/or lymph nodes or distant metastases is one of the common challenges in AI applied to medical imaging. This time-consuming task often limits the availability of training and test data, which further limits the performance of the models and the conclusions drawn.⁴

The role of convolutional neural networks is crucial. Their ability to learn directly from images could theoretically allow the annotation step to be skipped if a sufficiently large dataset is available; on the other hand they have shown excellent performance compared to radiomics and traditional machine learning techniques. However, only a few studies using this technology are available so far.⁴
Fin de cápsula

Welcome back. 
Cervical cancer, the fourth most common gynecologic cancer worldwide in terms of incidence and mortality, has also seen advances in the development of artificial intelligence models. Most studies have focused on predicting and detecting lymph node metastasis, while a few have explored areas such as predicting cancer versus non-cancer, lymphatic involvement, risk prediction, and treatment outcomes.⁴

Importantly, in all studies radiomic methods have been a valuable adjunct to diagnosis, with performance equal to or better than subjective diagnosis by radiologists. While manual segmentation is most common, some studies have used semi-automated or automated techniques for segmentation, and most have achieved better performance and accuracy in complex multi-parameter models.⁴

Now, moving on to ovarian cancer, the fifth leading cause of cancer death in women, artificial intelligence methods that can help differentiate benign from malignant tumors would be very useful in the early stages of the disease due to the typically advanced stage at diagnosis and a lack of effective screening options. Let us highlight the study by Wang and team, which found that deep learning models improved the performance of junior radiologists and matched the performance of senior radiologists.⁴ 

Another example are the studies by Li in 2020 and Park in 2021, that combined radiological features with age to build models to distinguish malignant from benign or borderline lesions. Both studies found age to be an important complementary feature.⁴

Another notable study by Yi and colleagues in 2021 selected 12 single nucleotide polymorphisms of the human sulfatase 1 for genotyping. The study combined the polymorphism analysis with radiomic and clinical features. Yi’s team built a model that showed promising results for predicting platinum resistance in ovarian cancer patients, with an interesting clinical application.⁴

Although the studies we have featured today show the vast progress of artificial intelligence integrated into obstetrics and gynecology in recent years, a vast opportunity remains for new research to develop more efficient approaches and set best practices.⁴

One of the most significant barriers that the clinical application of artificial intelligence faces is reproducibility. The lack of reproducibility prevents doctors from being able to test the researched methods in their institutions. Even if they decide to try a reimplementation, it might fail due to differences in image acquisition technique or patient populations.⁴

Thanks for joining us on this episode of Health Connect. Don't miss our on our next episode.
Discover more medical news on Viatris Connect.

References:

1. 1. Kharb S, Joshi A. Multi-omics and machine learning for the prevention and management of female reproductive health. Front Endocrinol (Lausanne). 2023; 14:1081667. Available at: https://www.frontiersin.org/articles/10.3389/fmed.2023.1098205/full
2. 2. Feduniw S, Golik D, Kajdy A, Pruc M, Modzelewski J, Sys D, et al. Application of Artificial Intelligence in Screening for Adverse Perinatal Outcomes-A Systematic Review. Healthcare (Basel). 2022; 10(11):2164. Available at: https://www.mdpi.com/2227-9032/10/11/2164
3. 3. Park S, Moon J, Kang N, Kim YH, You YA, Kwon E, et al. Predicting preterm birth through vaginal microbiota, cervical length, and WBC using a machine learning model. Front Microbiol. 2022;13:912853. Available at: https://www.frontiersin.org/articles/10.3389/fmicb.2022.912853/full
4. 4. Shrestha P, Poudyal B, Yadollahi S, E Wright D, V Gregory A, D Warner J, et al. A systematic review on the use of artificial intelligence in gynecologic imaging - Background, state of the art, and future directions. Gynecol Oncol. 2022;166(3):596-605. Available at: https://www.gynecologiconcology-online.net/article/S0090-8258(22)00496-6/fulltext

Os links para todos os sites de terceiros são oferecidos como um serviço aos nossos visitantes e não implicam endosso, indicação ou recomendação do Health Connect. Os artigos vinculados são fornecidos apenas para fins informativos e não visam implicar uma atribuição pelo autor e/ou editor. O Health Connect se isenta de qualquer responsabilidade pelo conteúdo ou pelos serviços de outros sites. Recomendamos que você analise as políticas e condições de todos os sites que escolher acessar.

NON-2023-9605
NON-2023-2512