Benchmarking machine learning architectures for menstrual recovery prediction using physiologically informed synthetic wearable data

Secondary amenorrhea is a heterogeneous condition with implications for reproductive, cardiovascular, and bone health. Existing machine learning approaches in menstrual health focus on cycle prediction rather than recovery modeling in pathological conditions. We present a proof-of-concept framework to model menstrual recovery within three months from non-invasive wearable-derived physiological features and self-reported inputs, including heart rate variability, resting heart rate, sleep, physical activity, skin temperature, perceived stress, age, and duration of amenorrhea. Using a synthetically generated dataset of 5000 individuals encoding physiologically informed feature-outcome relationships, twelve models were evaluated across baseline and longitudinal configurations. The best-performing model (XGBoost) achieved an AUC of 0.914, with ablation analysis confirming baseline features capturing the majority of learnable signal (\(\Delta\)AUC = 0.020). Permutation-based null models confirmed non-trivial predictive structure (AUC = 0.503), and XGBoost outperformed rule-based baselines (\(\Delta\)AUC = 0.044). SHAP analysis identified perceived stress and heart rate variability as dominant predictors, consistent with the data-generating structure. As the wearable-derived features are routinely captured by consumer devices and the self-reported inputs require brief periodic assessment, this framework establishes a foundation for wearable-based modeling of menstrual recovery, with future work required for real-world clinical validation and integration into health monitoring systems.