Automatic Subject Independent Pain Intensity Estimation using a Deep Learning Approach

The notion of pain covers a multitude of distinct sensorial experiences that serve as a warning system to the human subject. Pain management is crucial because an untreated or intense pain can subdue the nervous system and the entire emotional universe of the suffering patient. In this paper we propose an automatic pain assistance solution, a visual recognition system for pain intensity estimation that targets patients unable to communicate that they are in pain. We trained from scratch a ResNet model on the BioVid Heat Pain data set, with different selections of hyper-parameters and data augmentation and we evaluated the results. We also analyzed the impact of constraining the test and train data sets with regards to using images of the same subject for both the train and the test sets comparing with the scenario where the images of a subject are used solely for testing or for training sets. The selection of parameters is of central importance in building a robust classifier that generalizes and led us to state of the art results in the context of the complex pain intensity estimation task on the BioVid Heat Pain data set.