Benchmarking machine learning architectures for menstrual recovery prediction using physiologically informed...

Area of Science:

• Reproductive Endocrinology
• Biomedical Data Science
• Machine Learning in Healthcare

Background:

• Secondary amenorrhea impacts reproductive, cardiovascular, and bone health.
• Current machine learning in menstrual health primarily focuses on cycle prediction, not pathological recovery.
• There is a need for advanced modeling of menstrual recovery in clinical settings.

Purpose of the Study:

• To develop and validate a proof-of-concept framework for modeling menstrual recovery within three months.
• To utilize non-invasive wearable-derived physiological features and self-reported data for prediction.
• To establish a foundation for wearable-based menstrual health monitoring.

Main Methods:

• A synthetically generated dataset of 5000 individuals was used.
• Twelve machine learning models were evaluated, including baseline and longitudinal configurations.
• XGBoost was employed, with SHAP analysis for feature importance and ablation studies.

Main Results:

• The best-performing XGBoost model achieved an AUC of 0.914.
• Baseline features significantly contributed to the predictive signal (ΔAUC = 0.020).
• Perceived stress and heart rate variability were identified as dominant predictors.

Conclusions:

• The developed framework demonstrates a viable approach for wearable-based menstrual recovery modeling.
• The findings highlight the potential of integrating routinely captured wearable data for health insights.
• Further clinical validation and integration into health monitoring systems are necessary.