Classifying brain states and pupillary responses associated with the processing of old and new information

Memory retrieval of consolidated memories has been extensively studied using "old-new tasks", meaning tasks in which participants are instructed to discriminate between stimuli they have experienced before and new ones. Significant differences in the neural processing of old and new elements have been demonstrated using different techniques, such as electroencephalography and pupillometry. In this work, using the data from a previously published study (Campos-Arteaga, Forcato et al. 2020), we investigated whether machine learning methods can classify, based on single trials, the brain activity and pupil responses associated with the processing of old and new information. Specifically, we used the EEG and pupillary information of 39 participants who completed an associative recall old-new task in which they had to discriminate between previously seen or new pictures and, for the old ones, to recall an associated word. Our analyses corroborated the differences in neural processing of old and new items reported in previous studies. Based on these results, we hypothesized that the application of machine learning methods would allow an optimal classification of old and new conditions. Using a Windowed Means approach (WM) and two different machine learning algorithms - Logistic Regression (WM-LR) and Linear Discriminant Analysis (WM-LDA) - mean classification performances of 0.75 and 0.74 (AUC) were achieved when EEG and pupillary signals were combined to train the models, respectively. In both cases, when the EEG and pupillary data were merged, the performance was significantly better than when they were used separately. In addition, our results showed similar classification performances when fused classification models (i.e., models created with the concatenated information of 38 participants) were applied to individuals whose EEG and pupillary information was not considered for the model training. Similar results were found when alternative preprocessing methods were used. Taken together, these findings show that it is possible to classify the neurophysiological activity associated with the processing of experienced and new stimuli using machine learning techniques. Future research is needed to determine how this knowledge might have potential implications for memory research and clinical practice.