Artificial Intelligence in Obstetrics and Gynaecology: Is This the Way Forward?

Artificial Intelligence: Is This the Way Forward?

Artificial intelligence (AI) is defined as the use of complex algorithms in order for machines to reason and perform cognitive functions, including problem solving and decision making (7). AI is defined as comprising four essential core elements: machine learning (ML), natural language processing, artificial neural networks (ANNs) and computer vision (7). ML relies on subtle patterns and inference to make future predictions using complex datasets. ML has been shown to outperform logistic regression for prediction of surgical site infection by analysing multiple data consisting of laboratory values, diagnosis and treatment (8). Natural language processing is designed to understand human language, including incoherent speech (7). ANNs are inspired by biological nervous systems and comprise many computational units (i.e. ‘neurones’). A feed-forward type of ANN developed by a team in Edinburgh has proven to predict mortality from acute pancreatitis more accurately than the use of a scoring tool (APACHE II and Glasgow Severity) (9). Computer vision is the understanding of images and videos, such as facial recognition technology. An example which incorporates computer vision is real-time analysis of laparoscopic videos for identifying steps of a sleeve gastroscopy, with overall reported accuracy up to 92.8% (10).

Aim of this Review

We explore how AI applications are addressing common challenges in obstetrics and gynaecology. The main objective of this narrative review is to describe some important applications of AI in obstetrics and gynaecology, discuss whether AI can lead to deeper understanding of several pathophysiological aspects of these specialities, and recommend areas in obstetrics and gynaecology where AI may be used to improve healthcare provision.

AI in Obstetrics

Cardiotocography (CTG) monitors the fetal heart and uterine contractions and functions as an indirect tool for screening for the development of hypoxia. It is currently used as the major tool in intrapartum decision making. CTG sensitivity is estimated to be around 60%, and it continues to be debated whether this has improved perinatal outcomes since its introduction over 45 years ago (11). CTG interpretation is prone to human error and misrepresentation because of unavoidably high inter- and intra-observer variability. CTG interpretation, in other words, has a major element of subjectivity. There have been many studies looking at CTG data in order to predict fetal outcomes and help decision making. These aimed to demonstrate how ML can be an adjunct in fetal monitoring, and objectively help determine whether a caesarean section is necessary during intrapartum care.

In 2010, researchers converted fetal heart rate variability (HRV) into energy bands of movement and associated these bands with other activity of the foetus. Using these parameters as inputs, they created data segments (epochs) for normal and pathological cases. This form of supervised learning (ML) enabled detection of half of the pathological cases, with an acceptable false-positive rate of 7.5%. The combination of HRV plus other parameters, compared to HRV and the individual parameters alone, led to superior predictions (3).

In 2014, a large CTG dataset was devised into a random forest classifier, a form of decision tree, using latent class analysis to identify normal and abnormal CTG patterns. In this study, the sensitivity and specificity of correct CTG interpretation were 72% and 78% respectively; those values were calculated based on a complex ‘Aggregated Confusion Matrix model’ (12). Another study in 2017, where CTG data, age of mother, pH of umbilical artery, Apgar scores, base excess and deficit were used to classify normal vaginal and caesarean section deliveries. This study processed these parameters using subsets of ML algorithms including random forest classification, Fisher's linear discriminant analysis and deep learning. The deep-learning pattern approach seemed to produce superior sensitivity and specificity scores, 94% and 91% respectively, compared to the former studies (11).

AI in Gynaecological Oncology

Prognostication in gynaecological cancer is currently based on FIGO classification. A modern shift towards novel radiological or molecular biomarkers is emerging as the next step in treatment stratification. In a recent review, it was shown that several studies have explored the prognostic impact of p53 or Kirstein rat sarcoma viral oncogene homolog (KRAS) gene mutations in endometrial cancer (13). A classic example of novel radiological biomarkers includes the stratification of pelvic tumours based on extramural vascular invasion. Despite those efforts, the complexity of the carcinogenesis pathway, as well as lack of understanding of several multifactorial mechanisms involved in cancer progression artist to creating a huge challenge that needs to be addressed in order to achieve the vision of personalised medicine. Moreover, predicting response to neoadjuvant or adjuvant treatment for individualised care provision in gynaecological oncology requires a similar or sometimes deeper understanding of the relevant molecular mechanisms involved. Therefore, AI algorithms are being developed to face this challenge.

Researchers at Imperial College London used radiomics-determined mathematical descriptors of primary epithelial ovarian cancer risk phenotype to provide a reliable and promising predictive value of prognosis (14). Software, TexLab 2.0, was employed to summarise 657 features relating to the size, shape, texture, wavelet, intensity and decomposition of 364 preoperative contrast-enhanced computed tomographic scans. These radiomic profiles were investigated alongside molecular profiling and gene expression, in particular copy-number alterations. High-grade serous ovarian cancer had distinct radiomic features that were frequently associated with copy-number alterations and had worse outcomes. A radiomic prognostic vector was therefore devised, which possessed significantly better prognostic power compared to cancer antigen-125 (CA-125) and was also predictive of an association with chemotherapy resistance and poor surgical outcomes.

Another team investigated various models to predict 5-year survival in patients who had undergone radical hysterectomy for treatment of cervical cancer (15). Twenty-three items of demographic and perioperative data, and tumour-related parameters of 102 patients were collected. The team found the best model for prediction of 5-year overall survival was using a probabilistic neural network.

Other groups have developed a clinical decision scoring system using ANN for women with cervical intraepithelial neoplasia (CIN). They utilised data of human papillomavirus (HPV) biomarkers (E6 and E7 mRNA, and p16^INK4A) and colposcopic findings of 2,267 women. In comparison to cytology with or without HPV, the developed ANN had sensitivity and specificity of 93% and 99.2%, respectively, for prediction of CIN2 (16), suggesting enormous potential for improving patient care.

The Vision of Personalised Medicine

Hanahan and Weinberg discussed the importance of the six fundamental hallmarks in cancer biology. This has been the basis of the current molecular oncology principles, and an essential step to pursue the vision of being able to offer individualised approach for each cancer patient (17). Personalised medicine can be defined as the use of combined knowledge about a person (e.g. genetics, medical history) to predict disease susceptibility, prognosis or treatment response (18). Prognostic and predictive biomarkers can guide patient management and forecast outcomes. Mutations of breast cancer gene 1 and 2 (BRCA1 and BRCA2) are a successful example of parameters for provision of personalized treatment based on individual predictors or prognosis and response to treatment (19). KRAS mutations in endometrial cancer and WNT signalling in ovarian, endometrial and cervical cancer are also potential future targets based on which prognosis of cancer progression might be predicted on an individual basis and initial treatment options therefore be stratified accordingly (13, 20).

Modern medicine has shifted from developing treatments after the fact, to preventing, personalising and delivering precision care. This requires vast amounts of data to increase available knowledge on disease processes. Examples of processable data include proteomic, genomic and transcriptomic biomarkers, and baseline characteristics of patients. AI has been declared as the primary tool to synthesise data on complex cancer oncology and achieve the vision of personalised medicine.

AI in IVF

The selection of a viable embryo remains a principle challenge of IVF. This is essential in order to predict outcomes which could lead to shorter time to pregnancy and result in a healthy, live-birth. In 1997, Kaufmann et al. developed an ANN with a predictive power of 59% using only four inputs (age, number of eggs recovered, number of embryos transferred and whether there was embryo freezing) in order to predict the likelihood of successful IVF (21). Obstacles to greater sensitivity in this area include the many unknown factors which lead to a successful outcome for IVF which are needed to ‘teach’ for the purpose of ML (22). It is for this reason that large datasets, such as computer vision, are being used to construct an ANN in an attempt to increase predictive power.

Time-lapse imaging of embryos and image data have been used to obtain large datasets. Several studies were presented at the 2018 Annual Congress of the American Society for Reproductive Medicine and the European Society for Human Reproduction and Embryology. More specifically, one of those studies achieved 83% overall accuracy in predicting live birth by looking at 386 time-lapse images of single blastocyst transfers (23). Another study reviewed 50,392 images from 10,148 embryos and managed to obtain 97.53% accuracy in discriminating between a poor and good blastocyst (24). Others used pre-treatment characteristics of known cycles to predict first cycle success, which had an accuracy of 81% (25).

Limitations

We acknowledge a series of limitations in this narrative review. Firstly, despite the fact that a formal key word strategy was followed, this article has discussed only those studies which were considered to follow a robust research methodology and deemed essential for the understanding of the field. Secondly, the vast majority of the included studies related to data in an experimental phase and are yet to be applied clinically; some studies had small sample numbers and others oversampled data. In particular, with IVF studies there was a lack of explanation of how the training of machinery took place and the specific AI architecture used. There are also the limitations of AI itself, such as systematic biases and improper labelling of data, leading to skewed results. Finally, AI success really depends on large datasets, thus the speed of discovery is dependent on the speed of data availability.