Depression is a significant mental health issue that profoundly impacts people’s lives. Diagnosing depression often involves interviews with mental health professionals and surveys, which can become cumbersome when administered continuously. Digital phenotyping offers an innovative approach for detecting and monitoring depression without requiring active user involvement. This study contributes to the detection of depression severity and depressive symptoms using mobile devices. Our proposed approach aims to distinguish between different patterns of depression and improve prediction accuracy. We conducted an experiment involving 381 participants over a period of at least three months, during which we collected comprehensive passive sensor data and Patient Health Questionnaire (PHQ-9) self-reports. To enhance the accuracy of predicting depression severity levels (classified as none/mild, moderate, or severe), we introduce a novel approach called symptom profiling. The symptom profile vector represents nine depressive symptoms and indicates both the probability of each symptom being present and its significance for an individual. We evaluated the effectiveness of the symptom-profiling method by comparing the $F_1$ score achieved using sensor data features as inputs to machine learning models with the $F_1$ score obtained using the symptom profile vectors as inputs. Our findings demonstrate that symptom profiling improves the $F_1$ score by up to 0.09, with an average improvement of 0.05, resulting in a depression severity prediction with an $F_1$ score as high as 0.86.