A predictive model for next cycle start date that accounts for adherence in menstrual self-tracking

Area of Science:

• Digital Health
• Machine Learning in Healthcare
• Reproductive Health Technology

Background:

• Mobile health (mHealth) apps for menstrual tracking collect valuable self-tracked data.
• User adherence to tracking can impact data reliability, necessitating methods to disentangle physiological patterns from user behavior.
• Accurate prediction of menstrual cycle start dates is crucial for understanding reproductive health.

Purpose of the Study:

• To develop a predictive model for next menstrual cycle start date using self-tracked mobile health data.
• To explicitly account for self-tracking adherence in predictive modeling.
• To enable interpretable, evolving predictions and model individual cycle length history with population data.

Main Methods:

• Utilized a large dataset from a popular menstrual tracker (186,000 users, >2 million cycles).
• Developed a machine learning model that incorporates self-tracking adherence and updates predictions dynamically.
• Incorporated individual cycle history and population-level information into the model.

Main Results:

• The developed model outperformed five baseline models (mean, median, CNN, RNN, LSTM) in prediction accuracy.
• The model demonstrated consistent performance improvement as the menstrual cycle evolved.
• The model successfully provided predictions for skipped tracking probabilities.

Conclusions:

• Disentangling physiological menstrual patterns from adherence is essential for accurate mHealth app predictions.
• The proposed model enhances prediction of menstrual cycle start dates and user awareness of tracking behavior.
• This approach provides deeper insights into the underlying data structure of self-tracked reproductive health information.