Passive Sensor Data Based Future Mood, Health, and Stress Prediction: User Adaptation Using Deep Learning.

Abstract

Predicting one's mood, health, and stress in the future may provide useful feedback before wellbeing related problems become severe. Previously, researchers developed participant-dependent wellbeing prediction models using mobile and wearable sensors, where the models were trained and tested with the same group of people. However, in real-world applications, it is essential to consider the adaptability of the developed models to new users for predicting new users' wellbeing immediately and accurately. In this paper, we built wellbeing prediction models using passively sensed data from wearable sensors, mobile phones, and weather API, and deep learning methods, and evaluated the models with the data from new users. We compared deep long short-term memory (LSTM) network and the combination of convolutional neural network (CNN) and the LSTM model. We found that our deep LSTM model provided performances, in mean absolute error (MAE), as 15.7, 15.6, and 16.8 out of 100 in predicting self-reported mood, health, and stress respectively for new users. Furthermore, we applied a fine-tuning transfer learning method based on our deep LSTM model, which provided new participants with more accurate predictions, especially when the volume of new participants' data was limited. The transfer learning model improved the MAE performances to 13.5, 13.2, and 14.4 out of 100 for mood, health, and stress, respectively.