Use of Machine Learning to Mine User-Generated Content From Mobile Health Apps for Weight Loss to Assess Factors Correlated With User Satisfaction

IMPORTANCE The effectiveness of mobile health (mHealth) apps for reducing obesity is not ideal in daily life. Therefore, it would be useful to explore factors associated with user satisfaction with weight loss apps. Currently, research on these factors from the perspective of user-generated content is lacking. OBJECTIVE To mine the themes and topics frequently discussed in user-generated content in mHealth apps for weight loss, explore correlations of the topics with user satisfaction and dissatisfaction, and assess whether these correlations were asymmetric. DESIGN, SETTING, AND PARTICIPANTS In this population-based cross-sectional study, unsupervised machine learning was used to identify themes and topics in online discussions generated between January 1, 2019, and May 20, 2021, by Chinese users of mHealth apps for weight loss. MAIN OUTCOMES AND MEASURES Based on the 2-factor theory, a tobit regression model was used to explore the correlation of various app discussion topics with user satisfaction and dissatisfaction. Differences of the coefficients in models of positive rating deviation (PD) and negative rating deviation (ND), defined as the difference between the users' rating of the app and the app's comprehensive rating in the app stores, were analyzed by the Wald test. RESULTS In total, 191619 reviews and ratings from unique usernames were collected for 2139 weight loss apps; 86 423 reviews (45.1%) from 339 apps (15.8%) were included in the study. Most users (65 249 [75.5%]) were satisfied with the mHealth app. Eighteen topics were identified and summarized into 9 themes. Nine topics had significant positive correlations with the PD of user satisfaction, and 6 had significant negative correlations. The factor with the strongest positive correlation with the PD was celebrity effect (beta= 0.307; 95% CI, 0.290-0.323), and the factor with the weakest correlation was economic cost (beta = -0.426; 95% CI, -0447 to -0.406). Nine topics had significant positive correlations with the ND of user satisfaction, whereas 7 topics had significant negative correlations. The factor with the strongest positive correlation with the ND was fitness effect (beta = 1.369; 95% CI, 1.283-1.455). and the factor with the strongest negative correlation was economic cost (beta = -2.813; 95% CI, -2.875 to -2.751). There were significant differences in the PD and ND of user satisfaction. Nine motivation factors (ie, value-added attributes) and 7 hygiene factors (ie, user-expected attributes) for mHealth apps were identified. CONCLUSIONS AND RELEVANCE In this cross-sectional study, 16 factors had asymmetric correlations with user satisfaction and dissatisfaction with weight loss apps; 7 were related to basic expected attributes of the apps and 9 to value-added attributes. By distinguishing between expected and value-added factors, the use of weight loss apps may be improved.