Author(s)

Yizhen Fu

Corresponding Author

Yizhen Fu

Abstract

The menstrual cycle is a key indicator of women’s reproductive and endocrine health, influenced by hormones, lifestyle, and exercise. Accurate prediction is crucial for clinical decision-making and personal health management. Existing studies, however, often rely on small samples focus on single populations, or depend on noisy self-reported app data, lacking integration of multimodal features and medical interpretability. To address these gaps, we propose a multilevel AI framework that combines statistical analysis of exercise-hormone relationships with an LSTM-based time-series prediction model, incorporating physical, lifestyle, and hormona indicators. Furthermore, we fine-tune LLaMA3 using LoRA to build a domain-specific language model for menstrual health question answering and prediction interpretation. Experiments demonstrate a significant negative correlation between daily steps and progesterone, while estradiol shows no strong association. Our LSTM model outperforms ARIMA, GRU, and other baselines on MSE, MAE, RMSE, and R^2, while the fine-tuned LLaMA3 achieves superior PPL BLEU,ROUGE-L, and BERTScore compared to the original model. Overall, this study reveals exercise-hormone associations, proposes a robust cycle prediction method, and develops an intelligent health-oriented QA system, offering new tools for women’s health management.

Keywords

Menstrual cycle prediction; Women’s health; Hormone fluctuations; LSTM; LoRA fine-tuning