Cycle-phase supervision improves transferable physiological representations from wearable data for women's health

Abstract The menstrual cycle can be tracked through physiological signals captured by consumer wearables overnight --- skin temperature, breathing rate, heart rate and heart rate variability --- which act as proxies for the underlying hormonal variations. Other reproductive states, such as pregnancy or menopause, are often studied in isolation rather than as perturbations of the normal menstrual hormone rhythm, even though they are all governed by the same endocrine system. Here, we present CHARM (Cyclic Hormone-Aware Representation of the Menstrual cycle), a framework that learns menstrual cycle representations from wearable data by combining two training objectives: reconstruction of masked sensor channels, and a menstrual cycle phase (luteal or follicular) supervised loss. Phase supervision improves recovery of the underlying hormone dynamics from the learned representation. Without further training, frozen CHARM embeddings also generalise to multiple reproductive states and symptoms not seen during model training. The phase objective substantially improves this generalisation and is highly label-efficient, matching a strong baseline while using around 5-fold fewer labels for perimenopause and outperforming the baseline on menopause across all data regimes.