The gait kinematics of an individual is affected by various factors, including age, anthropometry, gender, and disease. Detecting anomalous gait features aids in the diagnosis and treatment of gait-related diseases. The objective of this study was to develop a machine learning method for automatically classifying five anomalous gait features, i.e., toe-out, genu varum, pes planus, hindfoot valgus, and forward head posture features, from three-dimensional data on gait kinematics. Gait data and gait feature labels of 488 subjects were acquired. The orientations of the human body segments during a gait cycle were mapped to a low-dimensional latent gait vector using a variational autoencoder. A two-layer neural network was trained to classify five gait features using logistic regression and calculate an anomalous gait feature vector (AGFV). The proposed network showed balanced accuracies of 82.8% for a toe-out, 85.9% for hindfoot valgus, 80.2% for pes planus, 73.2% for genu varum, and 92.9% for forward head posture when the AGFV was rounded to the nearest zero or 1. Multiple anomalous gait features were detectable using the proposed method, which has a practical advantage over current gait indices, including the gait deviation index with a single value. The overall results confirmed the feasibility of using the proposed method for screening subjects with anomalous gait features using three-dimensional motion capture data.