Adapting Generative Adversarial Networks to Predict Individual-Level Mental Health Symptoms (via Zoom)

Abstract: Continuous mobile sensing data streams, collected from smartphones and wearables, can capture information about human behavior, physiology, and well-being. Machine learning algorithms can then be trained on this data for predicting symptoms of mental health. Off-the-shelf machine learning algorithms are often not suited for this task. These data streams are composed of heterogeneous multivariate data distributions, and the relationships between the input data and output symptoms are idiosyncratic across both features and individuals. In his talk, Dan will speak about a specific project where he adapted an approach from the generative modeling literature for predicting changes in wearable data that occurred when individuals experienced prolonged stress. Specifically, he will speak about how he combined multitask learning with generative adversarial networks to predict individual-level multivariate data streams. He will also focus on how he created metrics that were clinically interpretable to assess generative model performance, and some of the ethical implications of continuous monitoring.

Bio: Dan Adler is a PhD student in the Information Science department at Cornell Tech. He is interested in how we can leverage passively collected data streams to design algorithms for mental health symptom prediction. In addition, he studies the implications of these methods, specifically the interpretability, security, and privacy. His recent work has focused on using generative models to predict well-being changes that occurred when individuals’ experienced prolonged stress, and designing an anomaly detection system using encoder-decoder networks to predict early warning signs of relapse in schizophrenia. Dan is advised by Dr. Tanzeem Choudhury.