Human gait has become of special interest to health professionals and researchers in recent years, not only due to its relation to a person's quality of life and personal autonomy, but also due to the involved cognitive process, since deviation from normal gait patterns can also be associated to neurological diseases. Vision-based abnormal gait detection can provide support to current human gait analysis procedures providing quantitative and objective metrics that can assist the evaluation of the geriatrician, while at the same time providing technical advantages, such as low intrusiveness and simplified setups. Furthermore, recent advances in RGB-D devices allow to provide low-cost solutions for 3D human body motion analysis. In this sense, this work presents a method for abnormal gait detection relying on skeletal pose representation based on depth data. A novel spatio-temporal feature is presented that provides a representation of a set of consecutive skeletons based on the 3D location of the skeletal joints and the motion's age. The corresponding feature sequences are learned using a machine learning method, namely BagOfKeyPoses. Experimentation with different datasets and evaluation methods shows that reliable detection of abnormal gait is obtained and, at the same time, an outstandingly high temporal performance is provided.