A Novel Approach to Menstrual Cycle Prediction Using GAN-Generated Data and Transformer Model for Tabular Data

Accurate menstrual cycle prediction is crucial for women's health and fertility management. While prior studies have utilized machine learning models to predict cycle length and classify regularity, they often struggled with the low Observationto-Variable Ratio (OVR) inherent in menstrual cycle data. This study tackles this issue by employing Generative Adversarial Networks (GANs) to generate synthetic data and Transformerbased models for both classification a nd r egression t asks. We evaluated multiple machine learning models for menstrual cycle phase classification a nd l ength p rediction. F or t he classification of regular versus irregular cycles, TabNet significantly outperformed baseline models, including Logistic Regression, Random Forest, and LSTM, achieving 99.67% accuracy, an AUC of 0.96, an F1 score of 0.99, 92.37% specificity, a nd 99.66% sensitivity, with CTGAN-generated synthetic data enhancing its performance. In the regression task of predicting cycle length, TabNet demonstrated strong predictive power, with an MSE of 1.29, RMSE of 1.13, and an R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> score of 0.87. Although Random Forest performed slightly better with an MSE of 0.958 and an R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> score of 0.90, TabNet remained competitive. These results highlight the potential of GAN-generated data and advanced models like TabNet in overcoming low OVR challenges and improving menstrual cycle prediction. Future research may incorporate personalized health data to further enhance model accuracy and generalizability.