AI, machine learning may aid in PCOS detection, reducing undiagnosed burden

Key takeaways :

• AI/machine learning demonstrated high performance in detecting polycystic ovary syndrome.
• Use of standard diagnostic criteria could make AI a useful clinical tool for PCOS.

SEATTLE — Artificial intelligence and machine learning showed high performance in diagnosing and classifying polycystic ovary syndrome, according to study data.

Use of standardized diagnostic criteria could make AI and machine learning more useful as tools for detecting PCOS and reducing diagnostic delays, according to a presenter at the American Association of Clinical Endocrinology annual conference.

“The main takeaway from our study is that across a range of diagnostic and classification modalities, there was an extremely high performance of AI and machine learning in detecting PCOS,” Skand Shekhar, MD, assistant research physician and endocrinologist at the National Institute of Environmental Health Sciences at the NIH, told Healio. “Specifically, for those studies that employed standardized criteria to evaluate the performance of AI models in PCOS, such as the international PCOS criteria and Rotterdam criteria, we found that AI/machine learning could distinguish those with PCOS from those without PCOS extremely well.”

In this systematic review and meta-analysis, Shekhar and colleagues searched Embase, Cochrane Register, Web of Science and the Institute of Electrical and Electronics Engineers (IEEE) Xplore Digital Library and identified 31 observational studies, from inception to January 2022, that evaluated the performance of AI/machine learning in the detection of PCOS. Studies that used clinical PCOS diagnostic criterion, such as NIH, Rotterdam or Revised International PCOS classification, were considered to diagnose PCOS, and studies without these criterion were considered to have classified PCOS.

Studies included from 9 to 2,000 participants. Overall, 23% of studies were multicenter studies and most were conducted in India (29%) or China (16%) with a median age of 29 years among PCOS participants. A total of 32% of studies diagnosed PCOS using established diagnostic criteria, and the remaining 68% of studies classified PCOS.

Clinical data with or without imaging were used in 55% of studies. The most common AI/machine learning techniques employed were support vector machine (42%), K-nearest neighbor (26%) and regression models (23%). Researchers observed an area under the receiver operating characteristic curve ranging between 73% and 100% in seven studies, a diagnostic accuracy of 89% to 100% in four studies, sensitivity ranging from 41% to 100% in 10 studies, specificity between 75% to 100% in 10 studies, positive predictive value between 68% and 95% in four studies and negative predictive value between 94% and 99% in two studies.

“We were pleasantly surprised by the very high efficacy of AI and machine learning in detecting PCOS across a range of technologies and data. These findings promise to open up an exciting new avenue to advance the health and well-being of millions of women worldwide who suffer from PCOS but remain undiagnosed or experience delays in getting a diagnosis,” Shekhar said.

“Future studies should employ AI in electronic health record settings and involve more clinician input so that these technologies can be effectively employed in the clinical care of those with PCOS,” he said.

For more information:

Skand Shekhar, MD, can be reached at skand.shekhar@nih.gov.