From Detection to Discovery: A Joint Learning Framework for Medical Knowledge Discovery and Depression Detection Using User-generated Content

Abstract

Researchers have long recognized that integrating domain knowledge into machine learning can enhance the efficiency of disease detection using user-generated content. However, a critical yet often overlooked aspect of research on combining machine learning with user-generated content is knowledge discovery: extracting new knowledge directly from user-generated content using machine learning techniques, thereby contributing back to medical knowledge. In this study, we use depression as a research case and develop a joint learning and knowledge graph-based framework, namely, Joint Depression Detection and Knowledge Completion (JDeC), to facilitate the iterative loops of predicting depression and discovering new medical knowledge from user-generated content. Specifically, we create a closed-loop joint training framework that combines ontology-based knowledge graph construction and learning, knowledge learning from user generated content on social media, and knowledge completion by integrating recognized entities from user generated content into the domain ontology.