Measuring the Severity of Depression from Text using Graph Representation Learning

The common practice of psychology in measuring the severity of a patient's depressive symptoms is based on an interactive conversation between a clinician and the patient. In this dissertation, we focus on predicting a score representing the severity of depression from such a text. We first present a generic graph neural network (GNN) to automatically rate severity using patient transcripts. We also test a few sequence-based deep models in the same task. We then propose a novel form for node attributes within a GNN-based model that captures node-specific embedding for every word in the vocabulary. This provides a global representation of each node, coupled with node-level updates according to associations between words in a transcript. Furthermore, we evaluate the performance of our GNN-based model on a Twitter sentiment dataset to classify three different sentiments and on Alzheimer's data to differentiate Alzheimer’s disease from healthy individuals respectively. In addition to applying the GNN model to learn a prediction model from the text, we provide post-hoc explanations of the model's decisions for all three tasks using the model's gradients.