| Abstract [eng] |
The aim of this work is to develop a system which detects and classifies pain into 4 categories, based on finger photoplethysmographic signal. To accomplish the task first medical literature was analyzed to find how to experimentally stimulate pain and how the body reacts to pain, what are expected physiological changes and how currently pain is measured, it was found that with pain there is an increased activation of sympathetic nervous system, and with it changes in cardiovascular system. Currently the golden standard for pain estimation is patient self-report, in cases where this is not possible there are scales based on patient’s movement, body position, sound, etc. Then technical literature review was performed to find state of the art in automatic pain recognition and classification., it was found that almost all research is focused on using facial video recordings and very few methods used physiological signals, and only one used photoplethysmogram signals, this is due to limited available databases and open databases having only ECG, EDA, and EMG signals. Other methods relied heavily on EDA signal due to its large initial reaction to pain. Methods were constructed to extract and process heartbeat pulses from PPG signals, features were then extracted to describe the pulse morphology. The extracted pulses were then passed through a quality control algorithm which determined is the PPG pulse artefact free. The extracted pulses were then separated into datasets by person. The task of 4 classes was split into 2 neural networks, an initial binary classifier for pain/no pain detection, and a secondary trinary classifier for pain class detection. Three different neural network architectures were designed with 10 different types of input. Results showed that all neural networks performed well in binary classification with feature based networks performing better than signal based networks. With highest achieved accuracy of 0.92 in testing dataset and 1.00 in training dataset. In trinary classifier all networks performed poorly, with signal based networks performing better than feature based. With highest achieved accuracy of 0.61 in testing dataset and 1.00 in training dataset. |
