Abstract:

Mobile health (mHealth) apps provide a rich source of self-reported observations: they provide day-to-day health indicators and behaviors, which can help shed light onto an individual's wellness and health over time. Menstrual trackers are an example of the promise and pervasiveness of mHealth apps, with millions of women around the world routinely tracking their menstrual cycles and a variety of contextual signs and symptoms.

 

Although the menstrual cycle is a powerful indicator of overall health in women, its characterization remains an open research question. In part, due to a lack of direct and reliable measurements of menstruation over-time and across individuals, a gap that menstrual trackers can help bridge. However, data from mHealth apps have questionable reliability, as they hinge on users' adherence to the app.

 

In this talk, I will describe how probabilistic machine learning can provide robust predictive models of self-tracking data in the context of menstrual trackers. Namely, I will present statistical generative models for personalized and accurate predictions of next cycle start date, accommodating mHealth variables that are subject to per-user missingness patterns. On the way, I will showcase the importance of disentangling physiological patterns from self-tracking adherence to provide informative predictions, as well as how statistical models enable well-calibrated predictions that accommodate the idiosyncrasies and uncertainties of mHealth self-tracked data.