A less-invasive method for the diagnosis of the major depressive disorder can be useful for both the psychiatrists and the patients. We propose a machine learning framework for automatically discriminating patients suffering from the major depressive disorder (n = 14) and healthy subjects (n = 17). To this end, spontaneous physical activity data were recorded via a watch-type computer device equipped by the participants in their daily lives. Two machine learning models are investigated and compared, i. e., support vector machines, and deep recurrent neural networks. Experimental results show that, both of the two methods, i. e., the static model fed with human hand-crafted features, and the sequential model fed with raw data can reach a promising performance with an unweighted average recall at 76.0 % and 56.3 %, respectively.