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

Although the growing prevalence of obesity has led to an increased reliance on health-tracking apps for weight management, their effectiveness remains limited owing to missing data and irregular sampling of user-reported records. These issues highlight the need for more sophisticated predictive models to address real-world data limitations and offer personalized interventions. This study developed a gated recurrent unit-ordinary differential equation (GRU-ODE)-Bayes-based deep learning framework to predict successful weight loss using real-world lifelog data.

We analyzed a retrospective cohort of Noom Coach users, who logged data at least twice a month for six months between 2012 and 2014. We included demographic and self-monitoring variables, with weight loss ≥5% in three months as the primary outcome. We evaluated the model performance using the area under the receiver operating characteristic curve (ROC AUC) and precision-recall curve (PRC AUC).

This study utilized a large-scale dataset (N = 34,322) that was subdivided into training (N = 24,292), validation (N = 6074), and test sets (N = 3375). Participants who frequently logged their weight, exercise, meals, and snacks were more likely to achieve weight loss. The model achieved an ROC AUC of 0.830 [95% confidence interval 0.819-0.840] and PRC AUC of 0.727 [0.707-0.746] in the validation set, and an ROC AUC of 0.821 [0.806-0.835] and PRC AUC of 0.717 [0.689-0.744] in the test set, using only the first four weeks of lifelog data. Early weight change and initial weight were the most important features as determined by the integrated gradients.

The proposed model predicted early outcomes in weight management and might contribute to developing effective intervention methods for participants at risk of failure and reducing the burden of frequent self-reporting.