Mental health issues such as stress, anxiety, and depression are growing concerns in today’s world. Traditional methods for depression monitoring are based on structured clinical assessments, which are inherently sparse, rely on self-reporting, and provide limited insight into individual symptoms. This study presents a deep representation learning framework that employs multimodal wearable data to monitor depression, allowing for more objective, continuous, and symptom-focused analysis. We intend to reduce our reliance on traditional clinical assessments by implementing unsupervised and semi-supervised learning approaches. The framework was tested on symptom sub-typing and identification tasks with promising results, demonstrating its potential to significantly improve mental health monitoring.