Abstract #1413640: Artificial Intelligence and Machine Learning in the Detection of Polycystic Ovarian Syndrome: A Systematic Review and Meta-Analysis

Abstract

Polycystic Ovarian Syndrome (PCOS) affects 5-20% of reproductive aged women but remains widely underdiagnosed. Moreover, artificial intelligence (AI) and machine learning (ML) have emerged as great supplemental tools in healthcare settings. To identify the utility of AI/ML in the diagnosis or classification of PCOS, we conducted a systematic review of the literature. We searched the following databases by applying relevant keywords - Embase, Cochrane Register, Web of Science and the IEEE Xplore Digital Library. Relevant studies were extracted and synthesized from inception up to January 1, 2022. Studies that used a clinical comparator criterion (NIH, Rotterdam, or Revised International PCOS classification) were considered to ‘diagnose’ and those without such a comparator were considered to ‘classify’ PCOS. Performance characteristics for each study were extracted and analyzed. A total of 31 studies were included in the final analysis out of 135 studies that were initially screened and extracted. Sources of data for AI/ML interventions included clinical information, electronic medical records, biochemical and genetic data. A total of ten studies (32%) diagnosed PCOS used established diagnostic criteria while the remaining 21 (68%) classified women with PCOS. Additionally, 17 (55%) studies utilized clinical data with or without imaging. Support vector machine (42% of studies), K-nearest neighbor (26%), and regression models (23%) were the commonest AI/ML techniques. Area under the ROC (plotted against clinical diagnosis) ranged between 73% and 100% (n=7 studies), diagnostic accuracy was between 89% and 100% (n=4 studies). Moreover, sensitivity ranged between from 41% and 100% (n=10 studies), specificity between 75% and 100% (n=10), positive predictive value (PPV) between 68% and 95% (n=4), and negative predictive value (NPV) between 94% and 99% (n=2 studies). Artificial intelligence and machine learning had high-performance scores in detecting PCOS and thus hold promise in reducing the burden of underdiagnosed PCOS. However, future AI and ML based studies should employ established PCOS criteria as comparators and adhere to methodological guidelines to improve their application and deployment in clinical settings.