Leveraging Unlabeled Data in Supervised Learning to Objectively Assess Depression

Abstract

Depression is the leading cause of disability, affecting over 250 million people worldwide [34]. Major Depressive Disorder, or MDD, is difficult to assess and diagnose due to lack of resources and the personal, private nature of the disease, which has symptoms that can vary greatly on a patient-by-patient basis [2]. Even so, the current standard methods of assessing depression are subjective and outdated, consisting of surveys and questionnaires first developed 60 years ago [21]. As part of a recent clinical study, data from wearable sensors, smartphones, and surveys were collected from a number of participants, and used to train classical machine learning models aimed at assessing depression. In this thesis, those methods are expanded upon with the intent of improving them, with varied success. Investigations conducted include training a small neural network on the same data, training a Multimodal Autoencoder on additional unlabeled data, and concatenating time features to utilize temporal information.