Predicting human movement from neural multivariate time series with ResNet

When we perform an action, such as moving our finger, our brain starts prepar- ing the action a few hundreds of milliseconds before. It is possible to detect such motor preparation by recording brain signals with good temporal resolu- tion such as the ones recorded with magnetoencephalography (MEG). The aim of this thesis was to detect motor preparation from brain signal and predict as early and as accurately as possible whether a participant is about to perform an action or not. To tackle that task, we trained and adapted ResNet, a neu- ral network that has been shown to be particularly efficient with multivariate times series recording, on data acquired in a group of 16 participants. The aim of the model is to correctly classify short time windows of the signal as containing or not motor preparation. We employed two different methods to achieve our goal. Firstly training a specific classifier for each participant (individual models), secondly training a classifier on a group of participants (group models). We found that by training and testing one classifier per short time window, we can detect using group models significantly above chance earlier (500ms prior the action) compared with using within participant models (250ms before the action). The performance reached by both scenarios, at the time of the action, reaches an AUC value around 65%. Furthermore, using a single model to detect in real time a motor preparation, we saw that only within participant classifiers were able to detect the action up to 500ms before its enactment significantly above chance level. A similar performance around 65% has been reached at the time of the action by both individual and group models. These results are of interest to the field of neurology, such as the control of neuroprosthesis. They are also of interest in cognitive neuroscience, in relation to the mechanisms underlying the sense of agency (SoA), which is the feeling of being masters of our actions and their consequences.