A deep learning-based early prediction framework for weight management using real-world lifelog data: GRU-ODE-Bayes...

Area of Science:

• Computational biology
• Health informatics
• Machine learning for healthcare

Background:

• Obesity prevalence necessitates effective weight management tools.
• Health-tracking apps show promise but are limited by data gaps.
• Sophisticated predictive models are needed for personalized interventions.

Purpose of the Study:

• To develop a deep learning framework for predicting successful weight loss.
• To address limitations of real-world lifelog data in health apps.
• To utilize a gated recurrent unit-ordinary differential equation (GRU-ODE)-Bayes model for prediction.

Main Methods:

• Analysis of a retrospective cohort of Noom Coach users (N=34,322).
• Inclusion of demographic and self-monitoring variables.
• Evaluation using ROC AUC and PRC AUC metrics.

Main Results:

• The GRU-ODE-Bayes model achieved strong predictive performance (ROC AUC ~0.82, PRC AUC ~0.72) on test data.
• Frequent logging of weight, exercise, meals, and snacks correlated with successful weight loss.
• Early weight changes and initial weight were key predictors.

Conclusions:

• The model accurately predicts early weight management outcomes.
• This approach can inform personalized interventions for at-risk individuals.
• It may reduce the need for extensive self-reporting.