In this paper, we investigate video-based deception detection using multiple feature selection methods. Firstly, the eye gaze, head pose and facial action unit (AU) features are extracted based on the open source tool OpenFace, while the video-based heart rate (HR) features are extracted using remote photoplethysmography (rPPG) technique. Then, multiple Wrapper methods are employed to screen the feature sets for feature selection with K-nearest neighbor (KNN) and support vector machines (SVM) classifiers. A large multi-modal deception detection dataset is collected to conduct the feature selection study. Experimental results demonstrate that the symbiotic organisms search (SOS) feature selection method with SVM classifier achieves the best performance as 83.27% for AUC (area under the curve) and 83.33% for ACC (accuracy). In particular, video-based HR feature plays an important role in enhancing the accuracy for visual deception detection.