Activity level, sleep behavior and weigh-in routine are known to be associated with weight goal achievement. The aim of this study is to investigate the use of pure end-to-end deep learning on raw data coming from consumer-grade digital health devices to predict weight goal achievement.


Anonymized data from 1.5 million users of Withings connected scales and sleep and activity trackers was analysed. We considered a binary discriminative formulation of the learning problem: given a user-defined weight goal and time series of activity, sleep, and weight measurements, we want to know if the user will succeed in losing (or gaining) the weight needed to achieve his goal.


The results of our systematic exploration into a broad range of machine learning methods for this weight goal prediction task show that a novel application of deep long short-term memory (LSTM) models is capable of reliable and accurate prediction. With this model, areas-under-ROC of 88% and accuracies of 80% after 10-fold cross-validation were reached – performance metrics that significantly outperform three “shallow” baseline approaches to sequence classification (GHMM, SVM and RF with 75%, 77%, 80% of areas-under-ROC respectively), as well as a common variation of deep neural networks (79%).


We can accurately predict if the user will achieve his weight goal based on activity level, sleep behavior and weigh-in routine. This is one of the first time that such quantities of longitudinal health data have been investigated, allowed by the advent of connected devices. While pure end-to-end deep learning approaches are routinely used within time-series domains like speech recognition, they are rarely applied to the emerging domain of digital health. We believe that the research presented here will be an informative step forward towards broader application of such techniques to other health domains.