Integrating Supervised and Self-Supervised Models for an Enhanced PCOS Detection: A Data-Driven Approach with Machine Learning Perspective

Polycystic Ovary Syndrome (PCOS) is a common ovarian dysfunction that leads to various difficulties, including missed or irregular menstrual periods, acne, depression, mood swings, and excessive facial hair. Due to the wide range of symptoms, diagnosing PCOS can be challenging. In addition, the process can be costly, and there is a high rate of false positives in PCOS diagnoses. Machine Learning (ML) can aid both patients and clinicians in diagnosing PCOS by using labeled data to analyze patient history effectively. However, the Self-Supervised model uses labeled and unlabeled data to understand new patient cases better. Incorporating unlabeled information during model training has become increasingly reliable for identifying a new patient case study. In this study, we employed the publicly available KagglePCOS tabular dataset to evaluate various modeling approaches. Our findings indicated that the Support Vector Classifier (SVC) model achieved a commendable accuracy of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 3. 8 \%}$</tex>. In comparison, the Gaussian Naive Bayes (GNB) model exhibited notably lower performance, with an accuracy of 54.1%. Among the Self-Supervised models, the Autoencoder showcased the most effective balance between accuracy and recall, achieving an accuracy of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{8 2. 7 \%}$</tex>. Although other models, such as Simple Contrastive Learning of Representations (SimCLR) and Bootstrap Your Own Latent (BYOL), demonstrated lower F1 scores and accuracy, they highlighted the potential of Self-Supervised methods in identifying new cases within unlabeled datasets. While traditional Supervised models remain highly effective, SelfSupervised learning offers a promising direction, especially for diagnosing cases where labeled data is unavailable.