Polycystic Ovary Syndrome (PCOS) is the most common endocrinopathy in reproductive aged women, with an estimated prevalence ranging from 4% to 20% and affecting more than 66 million worldwide in 2019 (1–5). PCOS is associated with increased incidence of cardiovascular disease, infertility, and of endometrial cancer (6–9). Its public health burden is immense, with nearly eight billion US dollars spent in 2020 to manage PCOS-related symptoms among women in the United States alone (10).

The diagnosis of PCOS is based on clinical criteria, with the Rotterdam criteria/International PCOS criteria (11, 12) being the most widely accepted. PCOS is characterized by the presence of a combination of hyperandrogenism, ovulatory dysregulation, and polycystic ovarian morphology (PCOM) (13–15). This already heterogenous clinical phenotype is complicated further by the elaborate interplay of genetic and environmental factors, such as diet related obesity or lifestyle factors, which affect clinical presentation (16). The criteria-based diagnosis of PCOS is complicated by variations in the clinical assessment of hyperandrogenism and determination of menstrual irregularity. Furthermore, the variation in normative standards for PCOM compounds these challenges (17). Estimates suggest that diagnosis is delayed by more than two years in one third of women with PCOS; yet this is likely an underestimation (18).

Artificial intelligence (AI) refers to simulation of human intelligence by computer based systems (19). On the other hand, machine learning (ML) is a subdivision of AI focused on learning from previous events and applying this knowledge to future decision making (20). ML techniques can be sub-classified as either supervised or unsupervised (21). The revolutionary advances in AI and ML over the last decade promise to rapidly advance our ability to diagnose and manage PCOS. This is in part due to the ability of AI to process massive amounts of disparate data, making it an ideal aid in the diagnosis of heterogeneous disorders like PCOS.

Several studies have investigated the ability of ML models to synthesize such disparate data as family genetic history, biomarkers, and demographic information into a unified algorithm for the diagnosis of PCOS, and make diagnostic predictions (22). Some pitfalls of these studies are their small size (22), lack of relevant comparators (23), use of varied diagnostic criteria (24, 25), and heterogeneity in reporting structures. Thus, the real gaps in knowledge and the full scope of AI/ML in the diagnosis of PCOS remain unclear. To better understand and summarize the body of evidence related to the application of AI/ML in PCOS, we conducted a systematic review of all relevant studies published up to January 1, 2022.

We performed a systematic review of AI/ML interventions in PCOS. All included studies were observational and retrospective. A small number used standard inclusion criteria such as the NIH, Rotterdam, or International PCOS criteria for diagnosis. Most studies achieved a high ability to diagnose PCOS or ‘classify’ patients as having PCOS using AI informed by clinical, radiological, electronic health records or biochemical data. Among the ten studies that used standardized criteria, the area under the receiver operator curve ranged from 80% to 100%, diagnostic accuracy from 89% to 100%, sensitivity from 87% to 100%, specificity from 90% to 100%, and positive predictive value from 68% to 81%. The most common AI/ML methods were SVM in 13 (42%) studies, K-nearest neighbor in eight (26%) studies, and regression models in seven (23%) studies. Importantly, a large number of the studies analyzed in the current review were able to achieve a high degree of diagnostic accuracy relative to standardized criteria. For instance, Deshpande et al. (2014) attained a 95% diagnostic accuracy against the Rotterdam criteria using an SVM algorithm using ultrasound imaging, clinical, and biochemical data (44). Similarly, Bharti et al. (2020) employed multiple ML algorithms to a dataset of 364 women with and without PCOS using clinical and imaging data and reported a > 90% diagnostic accuracy for the best SVM model (28).

AI/ML-based screening techniques for diabetic retinopathy and colorectal cancer have previously been found to be highly cost-effective (59, 60). In the case of colorectal cancer, cost savings of 400 million USD have been estimated when comparing next generation sequencing approaches to AI-based screening techniques (61). The potential use of AI/ML in the diagnosis and management of endocrine disorders has sparked intense research activity. A recent review reported that among the 611 ML-based endocrinology studies published between 2015 and 2020, 52% focused on diabetes, 14% on retinopathy, 14% on thyroid dysfunction, 8% on endocrine-related carcinoma, 7% on osteoporosis, and 5% on other disease states (62). Despite a growth in such studies, FDA-approved applications of AI for diagnostic or therapeutic purposes have lagged and approved devices employing AI/ML are concentrated in the management of diabetes and related conditions (63, 64).

In comparison, polycystic ovarian syndrome represents an ideal setting for future AI-based tools, given its high prevalence, significant healthcare burden, delayed detection, and complex diagnostic criteria spanning clinical, biochemical, and radiological domains. The diagnostic delay of greater than two years in a third of women reporting PCOS symptoms is a potent target for AI/ML-based approaches (18). Furthermore, geographical heterogeneity in clinical features of PCOS suggests an additional role of environmental influences, which may be overcome through adoption of AI/ML (65). Together, high costs and diagnostic delays in PCOS present a major unmet need which could be filled by the adoption of AI technology, as effectively demonstrated in other diseases. AI holds especially high potential for the diagnosis of PCOS because of its heterogeneous nature, with clinical, biochemical and radiological features each being incorporated into its diagnostic criteria (12). The use of AI on electronic health record (EHR) systems holds the potential to integrate these features while reducing diagnostic delays in PCOS.

The current body of research on AI in PCOS has revealed high rates of sensitivity and accuracy of PCOS detection. This implies that a well-designed AI/ML based program has the potential to significantly enhance our capability to diagnose PCOS early, with associated cost savings and a reduced burden of PCOS on patients and on the health system. However, several gaps remain in the domain of AI/ML based detection of PCOS. First, we noted that only a third of studies (32%) used standardized criteria such as the Rotterdam, NIH and International PCOS criteria as reference standards when evaluating AI in PCOS. This presents a high possibility misclassification of disease and biased detection estimates. Second, there was considerable heterogeneity in assessed AI-based studies, with some relying exclusively on a single parameter of PCOS diagnosis such as radiological, biochemical, or clinical features, despite Rotterdam criteria recommending diagnosis based on more than one of these elements. Third, a large number of assessed studies did not exhaustively report methodology/algorithms for AI based diagnosis, presenting concerns about the reproducibility of their findings. Most studies also relied on observational/retrospective data without use of prospective studies or validation datasets, limiting their applicability (66). A fourth major gap was the inadequate utilization of electronic health records, one of the most promising avenues for AI integration due to their potential for synthesizing clinical, biochemical, radiological, and genetic information and reducing lead time to the diagnosis in PCOS. This warrants further investigation in future studies. Finally, we noted that a vast number of AI/ML based studies were conducted in non-healthcare settings (71%) with non-healthcare investigators (97%). This raises the possibility of reduced applicability and relevance of studies in the clinical management of PCOS since such studies, while being technically robust, may not account for clinically important variables and outcomes. It is therefore important for physicians to become more aware of the advantages of AI/ML based methodologies and for physicians and computational scientists interested in AI/ML to work together to optimize the power of these new tools. Moreover, future AI/ML studies with applications for PCOS or other conditions, should make greater efforts to increase the methodological quality to increase the validity of the results. For this, we recommend the following five measures to improve the applicability of AI/ML for diagnosing PCOS and improving its care.

1. Increase collaboration between clinicians, researchers, and computational biologists.

2. Set up combined registries of data that include defined clinical, radiological (including images), and laboratory data (with reference values) of PCOS patients.

3. Use standardized criteria to train machine learning models as the standard reference and perform robust training and validation studies in PCOS patients.

3. Since some of the data used to develop the model may have some variation by time, it is important that future studies also test for performance (accuracy measures) consistency across time.

4. Enhance integration of population-based studies [e.g. All of Us, NHANES (67, 68)] with electronic health datasets to identify risk factors and risk enhancers for PCOS.

5. Include commonly used biochemical tests such as AMH, gonadal hormones, markers of insulin resistance and others in AI/ML to identify reliable biomarkers that can aid the diagnosis of PCOS.

To our knowledge, this is the first systematic review of AI/ML in the diagnosis of PCOS, spanning all published studies to date. We followed the methodological standards for systematic reviews proscribed by PRISMA guidelines. Despite the absence of a methodological assessment tool for evaluation of AI/ML based studies at the time of execution of this review, we performed a thorough evaluation of the quality by adapting the QUADAS-2 tool and adding relevant questions for the AI/ML interventions evaluated. Although not a weakness of our methods, confidence in our results is limited by the relatively small number of studies conducted on this subject, the heterogeneity of available data, and the risk of bias in primary studies. Broadly, poor dataset sourcing using non-standardized criteria, inconsistent use of best-practice machine learning methods, and limited clinical affiliations among authorship all undermined confidence in our selected studies.

In conclusion, our findings suggest that there is a high potential of AI/ML based programs in the diagnosis and care of PCOS, but that future studies should focus on enhancing methodological robustness and incorporating variables and outcomes of clinical importance.

The original contributions presented in the study are included in the article/ Supplementary Material . Further inquiries can be directed to the corresponding author.

FB and SS designed the study. FB and EB participated in all phases of the conduction of the study. FB, EB, SS, AR, JO, HP, and EL participated in screening, data extraction, and manuscript writing. JH, RR-G and SS reviewed the final version of the manuscript. All authors contributed to the article and approved the submitted version.