Hybrid RNN-ANN Based Deep Physiological Network for Pain Recognition.

Quantitative assessment of pain is vital progress in treatment choosing and distress relief for patients. However, previous approaches based on self-report fail to provide objective and accurate assessments. For impartial pain classification based on physiological signals, a number of methods have been introduced using elaborately designed handcrafted features. In this study, we enriched the methods of physiological-signal-based pain classification by introducing deep Recurrent Neural Network (RNN) based hybrid classifiers which combines auto-extracted features with human-experience enabled handcrafted features. A bidirectional Long Short-Term Memory network (biLSTM) was applied on time series of pre-processed signals to automatically learn temporal dynamic characteristics from them. The handcrafted features were extracted to fuse with RNN-generated features. Finely selected features from biLSTM layer output and handcrafted features trained an Artificial Neural Network (ANN) to classify the pain intensity. The handcrafted features enhance the RNN classification performance by complementing RNN-generated features. With our accuracy reaching 83.3%, comparison results on an open dataset with other methods show that the proposed algorithm outperforms all of the previous researches with higher classification accuracy. Therefore, this research is a good demonstration of introducing hybrid features for pain assessment.