Classification of noxious and non-noxious event-related potentials from S1 in pigs using a convolutional neural network

The purpose of this study was to develop a novel objective measurement of nociception using brain signals. Using deep learning, pain markers can possibly be extracted from event-related potentials (ERP). Therefore, this study aimed to develop a convolutional neural network (CNN) to classify between noxious and non-noxious stimuli based on ERPs from micro-electrocorticography (micro-ECoG) recordings in pigs. Micro-ECoG recordings were acquired from 13 experiments on 5 pigs. Subjects received electrical stimulation to the ulnar nerve, while micro-ECoG recordings were acquired using a 32-channel microelectrode array placed on the dura above the primary so-matosensory cortex. Each pig received three sets of both noxious and non-noxious stimulations. The micro-ECoG recordings were transformed into short-time Fourier transforms, which were used as input to the CNN. ERPs were classified with an accuracy of 73.5% and AUC of the receiver operating characteristic curve at 0.72. Additionally, the model was better at predicting non-noxious responses (85%) compared to noxious stimuli (62%). The developed CNN could classify between noxious and non-noxious ERP's recorded from pigs with a ROC-AUC at 0.72. In a further development process, the performance of the CNN model needs to be optimised and further research has to be conducted regarding the translation of the results from animal to human pain research.