A Generative Modeling Approach to Calibrated Predictions: A Use Case on Menstrual Cycle Length Prediction.

Area of Science:

• Machine Learning
• Healthcare Analytics
• Biostatistics

Background:

• Uncertainty quantification and calibration are vital in medical predictive modeling.
• Healthcare data, from electronic health records to mobile health (mHealth) self-tracking, presents inherent variability and reporting uncertainties.
• Accurate estimation of uncertainty is essential for reliable clinical decision-making.

Purpose of the Study:

• To explore methods for quantifying uncertainty in healthcare predictive models.
• To develop a flexible generative model for accurate and well-calibrated prediction of menstrual cycle length using self-tracked data.
• To demonstrate the utility of flexible generative models in accommodating mHealth data idiosyncrasies and personalizing uncertainty estimates.

Main Methods:

• Utilized a flexible generative model with two degrees of freedom to fit the mean and variance of observed cycle lengths.
• Employed machine learning techniques to achieve state-of-the-art predictive accuracy and enable well-calibrated predictions.
• Evaluated the model on real-world menstrual cycle length data from a popular mHealth application.

Main Results:

• The proposed flexible generative model provides accurate and well-calibrated predictions for menstrual cycle length.
• The model successfully accommodates under-dispersed distributions common in physiological data.
• Demonstrated the ability to adjust predictive uncertainty to individual user's cycle length patterns.

Conclusions:

• Flexible generative models offer a powerful approach for both accurate prediction and robust uncertainty quantification in healthcare.
• This methodology is well-suited for handling the complexities of mHealth data, improving personalized health insights.
• Enhanced, less uncertain cycle length predictions can benefit menstrual health research, mHealth users, and developers, leading to improved mHealth solutions.