Real-time pain detection in facial expressions for health robotics

Automatic pain detection is an important challenge in health computing. In this paper we report on our efforts to develop a real-time, real-world pain detection system from human facial expressions. Although many studies addressed this challenge, most of them use the same dataset for training and testing. There is no cross-check with other datasets or implementation in real-time to check performance on new data. This is problematic, as evidenced in this paper, because the classifiers overtrain on dataset-specific features. This limits realtime, real-world usage. In this paper, we investigate different methods of real-time pain detection. The training data uses a combination of pain and emotion datasets, unlike other papers. The best model shows an accuracy of 88.4% on a dataset including pain and 7 non-pain emotional expressions. Results suggest that convolutional neural networks (CNN) are not the best methods in some cases as they easily overtrain if the dataset is biased. Finally we implemented our pain detection method on a humanoid robot for physiotherapy. Our work highlights the importance of cross-corpus evaluation & real-time testing, as well as the need for a well balanced and ecologically valid pain dataset.