In this thesis, we study deep learning-based diagnostics with non-intrusive continuous sensing for mobile health- care devices. Our objective is to enable comfortable remote medical care at home with signals easily measured at home to continuously monitor and detect diseases. We address the problem in the following medical applications: (i) Atrial fibrillation (AF) detection, (ii) Sleep stage classification. First, we tackle the problem of distinguishing AF PPG signals from PPG signals of normal but high-risk factor groups. Premature atrial contraction (PAC), which is common in high-risk factor populations, makes the rhythm irregular, making it difficult to distinguish from AF. We proposed an end-to-end deep learning framework that not only considers irregularities of the rhythms but learns temporal patterns of PAC to classify AF. Second, we aim for trustable use of deep learning-based AF detection to be integrated into a ring-type wearable device. We focused on the explainable analysis of a deep learning-based AF detector and revealed the patterns and features it focus on to classify AF. Lastly, we propose a non-contact sound-based sleep stage classifier to enable low-cost at-home continuous sleep tracking. We proved that our model is robust to environmental noise so that it applies to smartphones or smart speakers.