Can Data Augmentation Improve Daily Mood Prediction from Wearable Data? An Empirical Study

Mobile sensing data, approximating human behavior and physiology, can be processed by machine learning models to predict mental health symptoms. While these models are accurate in smaller samples, their generalization accuracy decreases in larger samples, potentially because it is difficult to collect enough mobile sensing and mental health outcomes data at scale to enable generalization. In this study, we hypothesized that augmenting training data with synthetic data samples could improve the generalizability of these machine learning models. We created a data augmentation system that generated synthetic mobile sensing and mental health outcomes data, and evaluated the utility of this system via the downstream machine learning task of predicting daily mood from wearable sensing data. We experimented with both simple (e.g. noise addition) and novel generative data augmentation methods, based upon conditional generative adversarial networks and multi-task learning. Our initial findings suggest that the data augmentation system generated realistic synthetic data, but did not improve mood prediction. We propose future work to validate our findings and test other methods to improve the generalizability of mental health symptom prediction models.