Deep convolutional neural networks can discriminate pain from other aversive states in EEG data.

1469 

Abstract Submission 

Alyson Champagne¹, Mathilda Buschmann², Mathieu Roy³, Michel-Pierre Coll¹

¹Laval University, Québec, Quebec, ²Universität Osnabrück, Osnabrück, Germany, ³McGill University, Montreal, Quebec

Alyson Champagne  
Laval University
Québec, Quebec

Mathilda Buschmann  
Universität Osnabrück
Osnabrück, Germany

Mathieu Roy, Dr.  
McGill University
Montreal, Quebec

Michel-Pierre Coll  
Laval University
Québec, Quebec

Pain is experienced differently by each person, depending on their previous experiences, emotions and mood. In research contexts, pain is assessed by scales or verbal reports, measures with several limitations (Smith et al., 2016). Thus, the scientific community has invested considerable effort in creating pain biomarkers based on brain activity (Davis et al., 2020; Wager et al., 2013) recorded by electroencephalography (EEG; Mari et al., 2022) that could provide a complementary way of measuring pain. However, these attempts are limited by the use of univariate analyses that do not utilize the rich multivariate nature of EEG signals and the lack of adequate control conditions (Mari et al., 2022). Here, we collected EEG data while individuals experienced various levels of pain and other aversive control stimulations. The aim of this study was to assess the sensitivity and specificity of various machine learning algorithms applied to EEG data to discriminate between pain and other aversive states within and across individuals.

EEG data (64 channels, Brain Vision Acticap) was collected from 65 healthy individuals (39 females, mean age: 24.2 +/- 6.14) under three conditions: resting state (five minutes), experiencing tonic thermal pain for eight minutes, or listening to an unpleasant auditory stimulus for eight minutes. Twenty-two of the participants were tested using a different EEG system of the same model in a different location. These participants served as the test sample. During all tasks except resting, participants first experienced the stimulation passively and then experienced it a second time while providing a continuous intensity or unpleasantness rating using a visual analog scale.

We preprocessed EEG data using a semi-automated approach, with steps including bandpass and powerline noise filtering, removal of bad channels and artifactual independent components and rejection of bad trials (Jas et al., 2017; Pion-Tonachini, Kreutz-Delgado & Makeig, 2019). The continuous recordings in each condition were split into 4-second epochs with no overlap, and classifiers were applied to sensor space data to attempt to classify these epochs according to their pain condition (pain vs no pain) in the training sample (n = 43). We compared the performance of a random forest classifier, a shallow convolutional neural network (CNN) and a deep CNN (Schirrmeister et al., 2017). Nested cross-validation was used to optimize hyperparameters and perform model comparisons. Within-participant accuracy was assessed using 10-fold cross-validation in each participant of the training sample, and between-participant accuracy was assessed in the independent test sample (n = 22).

In line with previous research using machine learning on EEG (Engemann et al., 2022), the shallow CNN showed the best classification performance compared to the deep CNN and random forest classifier. The accuracy of the shallow CNN was significantly above chance (0.5) for the classification performed within participants (mean cross-validation balanced accuracy = 0.68 +/- 0.09, range: 0.54-0.90; Figure 1A). Between participants, classification performed in the independent test sample led to a drastically lower accuracy but still significantly above chance (mean balanced accuracy = 0.56 +/- 0.07, range: 0.42-0.73; Figure 1B).

Our results confirm the ability of convolutional models to distinguish EEG signals associated with experimental pain from those associated with other aversive conditions across individuals. However, considerable work remains necessary to improve the performance of the models and establish the best approach to identify and measure pain using EEG signals. Our future work will aim to create a large open database of EEG pain recordings to provide the community with adequate data to test different approaches to reach this goal.

Classification and Predictive Modeling ¹

EEG/MEG Modeling and Analysis

EEG

Perception: Pain and Visceral ²

ADULTS

Electroencephaolography (EEG)

Machine Learning

Multivariate

NORMAL HUMAN

Pain

Perception

Other - Classification