JMIR Publications

ABSTRACT

Background:

Generally, adherence to diet and physical activity in behavioral weight management programs predicts weight loss. Meeting attendance and platform and/or app use predict weight loss. A lot of weight loss prediction studies primarily use regression. However, little is known about the levels of adherence to specific online weight management program tools, and behavioral and psychosocial variables that predict significant weight loss using machine-learning approaches.

Objective:

We aimed to examine variables that can predict clinically significant weight loss (≥5%) in a multi-modal commercial digital weight management program, including levels of adherence to intervention tools and changes in behavioral and psychosocial states using machine-learning approaches.

Methods:

We performed secondary analyses using data from a one-arm trial online WW (formerly Weight Watchers) weight-loss intervention program lasting 6-months that recruited US adults with BMI between 25-45 kg/m2. We used a WW Bluetooth scale and digital intervention tools such as mobile app for point tracking, weekly virtual workshops, weekly wellness check-ins and a Facebook group, and changes in psychosocial and behavioral variables (i.e., food craving, pittsburgh sleep quality index, diet, hunger, and physical activity). Using the receiver operating characteristics (ROC) curve, we identified the predictors of significant weight loss, as well as the associated cut-points (CP) and area-under-curve (AUC) values for each variable. We further used a classification tree to confirm the importance of these predictors and assessed the out-of-sample prediction accuracies using 5-fold cross-validation.

Results:

Participants (N=153) were 70% female, 66% White, with a mean age of 41.09 (SD=13.78), and had a mean BMI of 31.8 (SD=5.0). About 51% of participants lost ≥5% weight. Using ROC curve, food tracking (CP ≥ 39.4%; AUC = 0.744; p<.001), increase in self-weighing (CP ≥ 0.5; AUC = 0.733; p<.001), wellness check-in attendance (CP ≥ 31.3%; AUC = 0.723; p<.001), and workshop attendance (CP ≥ 35.4%; AUC = 0.699; p<.001) were identified as significant predictors of achieving ≥5% weight loss. We confirmed the importance of these variables using classification tree, and together they achieved an out-of-sample prediction accuracy of 78%.

Conclusions:

ROC curve and classification tree provided consistent results of predictors of clinically significant weight loss (i.e., food tracking, increase in self-weighing, workshop attendance, and wellness check-in attendance). Our study extends the weight management literature by utilizing machine-learning approaches to identify significant weight loss predictors and specific levels of these predictors needed to achieve clinically significant weight loss.