Automatic detection of viewer interest while watching video contents can enable multimedia applications, such as online video streaming, to recommend contents in real time. However, there is yet a generic model for detecting viewer interest that is independent of subject and content while using noninvasive sensors in near-natural settings. This paper is the first attempt at solving this issue by investigating the feasibility of a generic model for detecting viewer interest based on facial expression and heart rate features. The proposed model adopts deep learning features, which are trained and tested using multisubjects' data across different video stimuli domains. The experimental results show that the generic model can reach a similar accuracy to a domain-specific model.