Mental health is a complex process, consisting of a causal relationship between (1) context, e.g, external circumstances and personality traits, which determines the mental health, and (2) symptoms, i.e., behavioral, psychosocial, and physiological responses, by which mental health is observed. SOTA mental health research is focused on finding relationships between affective states and physiological responses. However, transitioning this knowledge to real-life settings poses several problems; (1) physiological responses do not solely depend on mental health, (2) people can be subject to intra-user variance, and (3) machine learning models are mostly black-box and therefore give little to no insights. I will tackle these problems by constructing a multimodal and dynamic hierarchical sensing framework for personalized (mental) health monitoring in real-life. Multimodal sensing will be used to detect non-physiological symptoms and thus incorporate context. The multimodality of incoming data streams will be handled by this hierarchical framework, where dynamic questioning enables capturing intra-user variance. By fusing hierarchical anomaly detection with behavior modelling, using an active learning approach, I will determine the optimal moment to gather user feedback. Finally, I will focus on providing sensible insights to both physicians and patients using expert knowledge. This research will be conducted and validated on two mood disorders; stress and depression.