Predicting Physical Exercise Adherence in Fitness Apps Using a Deep Learning Approach.

Area of Science:

• Digital Health
• Behavioral Science
• Machine Learning

Background:

• Mobile fitness apps surged in popularity, especially during the COVID-19 pandemic, due to gym closures and reduced mobility.
• While promoting active lifestyles, these apps face significant challenges with user attrition and adherence to training plans.
• Predicting user adherence is crucial for improving engagement and retention in the digital health sector.

Purpose of the Study:

• To develop an automatic classification system for predicting user adherence to mobile fitness apps.
• To forecast user behavior (adherence/non-adherence) in the fourth month based on the first three months of app usage.
• To assist fitness app developers in implementing strategies to reduce user attrition.

Main Methods:

• Utilized data from 246 users of the Mammoth Hunters Fitness app.
• Applied pre-processing and clustering techniques, including affinity propagation (AP), to group users based on their initial 90-day training behavior.
• Employed an ensemble regression model, specifically Long Short-Term Memory (LSTM) networks trained within clusters, to predict fourth-month adherence.

Main Results:

• The combined AP clustering and LSTM model achieved high predictive performance.
• The system demonstrated 87% accuracy and an 85% F1-score in classifying users' adherence status.
• The model successfully predicted future adherence or non-adherence based on early usage patterns.

Conclusions:

• The developed system can effectively anticipate future user adherence to mobile fitness applications.
• This predictive capability offers fitness app creators opportunities to proactively address attrition.
• Implementing such predictive models can lead to enhanced user retention and more effective digital health interventions.