The goal of this project is to develop techniques to remove identifying information from wearable and phone data to protect patients’ privacy, yet preserve information to predict depressive symptoms. We propose a novel method following an adversarial training framework combined with representation learning using an autoencoder.
We depict the identifiability of users from the wearable and phone sensors. We show that common dimensionality reduction techniques still preserve identifying information for most users. We show that our method is successful in masking identifying information as measured by accuracy and F1 score of IC(z) while being able to retain information relevant to HAMD regression task compared to baseline.
This research is supported by National Institute of Health, MIT-MGH Grand Challenge, and MIT J-Clinic.