Undertreatment or overtreatment of pain will cause severe consequences physiologically and psychologically. Thus, researchers have made great efforts to develop automatic pain assessment approaches based on physiological signals using machine learning techniques. However, state-of-art research mainly focuses on verifying the hypothesis that physiological signals can be used to assess pain. The critical assumption of these studies is that training data and testing data have the same distribution. However, this assumption may not hold in reallife scenarios, for instance, the adoption of machine learning model by a new patient. Such real-life scenarios in which user's data is unlabeled is largely neglected in literature. This study compensates for the rift by proposing an adaptive transfer learning based pain assessment system (ATLAS), a novel adaptive learning system based on the transfer learning algorithm Transfer Components Analysis (TCA) to minimize the distance between training data and unlabeled testing data. Experiments were conducted on BioVid database, and the results showed our approach outperforms three existing traditional machine learning-based approaches and achieves an accuracy just 2.0% below the accuracy with labeled data.