Development and validation of an interpretable Conditional RNN for weight change prediction an obesity management mobile app (Preprint)
BACKGROUND As an alternative to on-site obesity management, a mobile-based intervention has been given more attention. Despite the rise of mobile interventions for obesity, there are lost opportunities to achieve better outcomes due to the lack of a predictive model using currently existing health data collected longitudinally and cross-sectionally. OBJECTIVE This study aimed to develop a predictive model for weight to be used in mobile-based interventions using interpretable AI, and to explore the contributing factors to weight loss. METHODS Using lifelong of mobile application users (Noom) who used a weight loss program for 16 weeks in the U.S., an interpretable recurrent neural network for the prediction of weight after intervention considering both time-variant variables and time-invariant variables was developed. This interpretable model was trained and validated with fivefold cross-validation testing (training set: 70%; testing: 30%) using lifelog data of app users for weight loss. Mean average percent error (MAPE) between actual weight loss and predicted weight, and contribution coefficients for model interpretation. To better understand the behavior factors to weight loss or gain, the contributing factors were calculated by the contribution coefficients in test sets to interpret the effects of contributing factors to weight loss. RESULTS A total of 17,867 eligible users were included in the analysis. The overall mean average percentage error of the model was 3.50% and the errors of the model declined from 3.78% to 3.45% by observing the data at the end of the program. The time level contribution was shown to be equally distributed at 0.0625 in each week, but this gradually decreased as it approached 16 weeks. Factors such as usage pattern, weight input frequency and meal input adherence, exercise, and sharp decreases in weight trajectories had negative contribution coefficients of -0.021, -0.032, -0.015, and -0.066, respectively. As for time-invariant variables, males had a -0.091 contribution coefficient. CONCLUSIONS An interpretable artificial intelligence to utilize both data and time fixed data can forecast weight loss precisely after obesity management application while preserving model transparency. This week to week prediction model is expected to improve weight loss and provide a global explanation of contributing factors, leading to better outcomes.