Evaluating Symptom and Wearable Features for Machine-Learning Based Menstrual Bleeding Day Classification for Variable Cycles
Advita Goyal
28/09/2026
Period-tracking applications are becoming increasingly common among younger generations. In period-tracking applications and models, accurately identifying menstrual bleeding days for highly variable cycles remains a weak spot. This can potentially disrupt lives for menstruators. Many studies have evaluated the use of different types of data to improve menstrual date classification, but there is a lack of research on types of data that specifically affect variable-cycle individuals. This study investigates the potential of symptom data (headaches, cramps, and more) and wearable device data (breathing rate, temperature, and more) to improve same-day menstrual bleeding classification compared to cycle history alone for variable cycles. In this study, three machine-learning models: logistic regression, gradient boosting, and random forest, were used to evaluate how same-day menstrual bleeding classification differed across four feature sets: cycle history only, cycle history + symptoms, cycle history + wearables, and all features. Among individuals with higher cycle variability, the highest AUC was achieved by the gradient boosting model using the cycle history + symptoms feature set (AUC: 0.831), but the cycle history feature set performed similarly (AUC: 0.827). In all three models, adding symptom and wearable features did not consistently improve performance compared to cycle history alone. This study served as a proof-of-concept exploratory study to potentially guide future research; future research should use larger and more diverse datasets to develop software to improve the classification of menstrual bleeding days among individuals with variable cycles to develop more accurate and clinically useful same-day menstrual bleeding classification tools.