Personalized multimodal physiological modeling of menstrual cycle dynamics using endocrine, metabolic and wearable signals

Menstrual cycle tracking remains challenged by the limited physiological scope and inconsistent generalizability of existing wearable‐based approaches, which often rely on individual sensing modalities that incompletely capture the complexity of menstrual physiology. This study presents a personalized multimodal machine learning framework for menstrual phase classification by integrating endocrine, autonomic, metabolic, and respiratory physiology through hormonal, glucose, heart rate variability, and respiration‐derived signals. Using the open‐source mcPHASES dataset, 126 menstrual cycles from 42 participants were analyzed over fixed 90‐day monitoring periods. Daily physiological measurements were transformed into statistical, spectral, and nonlinear descriptors and modeled using an Extra Trees ensemble classifier with leakage‐controlled five‐fold cross‐validation and fold‐specific Synthetic Minority Oversampling Technique balancing. Hormone‐derived features consistently provided the strongest individual predictive performance, reflecting the central role of endocrine regulation in menstrual physiology. However, multimodal integration produced the highest overall classification performance, yielding a consistent median improvement of +0.0156 macro F1 over the strongest unimodal models (Wilcoxon signed‐rank test, p = 8.0 × 10 −6 ; effect size r = 0.87). Although these improvements were modest in absolute magnitude, they were remarkably consistent across participants, indicating that multimodal integration primarily enhanced prediction robustness through complementary physiological information rather than large performance gains. The substantial variability in participant‐specific optimal sensing configurations (normalized entropy = 0.834) further demonstrated pronounced inter‐individual physiological heterogeneity. Collectively, these findings suggest that menstrual phase prediction is more appropriately viewed as a personalized multimodal inference problem than a conventional sensor optimisation task. Endocrine biomarkers provide the principal representation of ovarian activity, while autonomic, metabolic, and respiratory signals contribute complementary, context‐dependent information that refines participant‐specific prediction. This adaptive multimodal framework provides a foundation for scalable precision menstrual health analytics, longitudinal reproductive monitoring, and next‐generation personalized wearable health technologies.