Prediction meets time series with gaps: User clusters with specific usage behavior patterns

Area of Science:

• Digital Health
• Machine Learning
• Time Series Analysis

Background:

• Mobile health (mHealth) apps collect real-world data for treatments but suffer from fluctuating engagement and high user dropout rates.
• These data challenges hinder machine learning (ML) analysis and understanding of user adherence.
• Identifying user disengagement is crucial for effective mHealth interventions.

Purpose of the Study:

• To develop a method for identifying and predicting varying dropout rates in mHealth app datasets.
• To predict periods of user inactivity based on their current engagement state.
• To analyze the evolution of adherence within different user clusters.

Main Methods:

• Utilized change point detection to identify distinct phases of user engagement and dropout.
• Employed time series classification to predict user phases based on their activity.
• Addressed challenges of uneven, misaligned time series with missing values.
• Evaluated the method on an mHealth app dataset for tinnitus management.

Main Results:

• Successfully identified phases with different dropout rates and predicted future user behavior.
• Demonstrated the ability to predict expected inactivity periods for individual users.
• Showcased the evolution of adherence across different user clusters.
• Validated the approach's effectiveness on real-world mHealth data.

Conclusions:

• The proposed method effectively handles uneven, unaligned time series data common in mHealth apps.
• This approach is suitable for studying user adherence in datasets with missing values and variable lengths.
• The findings can improve the reliability of data analysis and intervention strategies in mHealth.
• Accurate prediction of user dropout and adherence is vital for optimizing digital health tools.