Machine-Aided Self-diagnostic Prediction Models for Polycystic Ovary Syndrome: Observational Study

Area of Science:

• Digital health and artificial intelligence applications in women's healthcare.
• Development of predictive models for endocrine disorders.

Background:

• The COVID-19 pandemic accelerated advancements in AI and digital health for medical diagnosis and treatment.
• Polycystic Ovary Syndrome (PCOS) diagnosis and management remain a significant challenge in women's health.

Purpose of the Study:

• To develop machine learning models for the self-diagnosis of PCOS using noninvasive measures.
• To create prediction models for PCOS diagnosis utilizing both noninvasive and invasive clinical data for healthcare providers.
• To compare the performance of patient-centric and provider-centric PCOS prediction models.

Main Methods:

• Retrospective analysis of a publicly available dataset of 541 women's health records.
• Application of the CatBoost classification method and K-fold cross-validation for model performance evaluation.
• Utilized SHAP values for variable importance analysis and k-means clustering with Principal Component Analysis for BMI subgroup analysis.

Main Results:

• Achieved 81%-82.5% accuracy for PCOS prediction using noninvasive measures (patient model).
• Attained 87.5%-90.1% accuracy for PCOS prediction using all variables (provider model).
• Identified key noninvasive predictors (acanthosis nigricans, acne, hirsutism, menstrual irregularities) and invasive markers (ovarian follicle count, anti-Müllerian hormone).

Conclusions:

• Proposed AI models can serve as a digital platform for PCOS pre-diagnosis and risk assessment.
• The tools facilitate convenient, at-home PCOS risk evaluation for women before seeking medical care.
• Clinical providers can leverage these models to aid in the diagnosis of PCOS.