Deviant Tone Detection in lnfants by P3a ERP Classification

The aim of this study was to implement a machine learning classifier for infant EEG data. Similar classifiers for adults are commonly utilized by brain-computer-interfaces. Three experiments were conducted to test our implementation, with the goal of classifying standard versus deviant/target stimuli. Experiment 1 was conducted with 72 infants (mean age: 8.4 month) performing a passive auditory two-stimulus oddball paradigm. Experiment 2 tested our classifier implementation on publicly available adult three-stimulus oddball data. Experiment 3 tested the oddball paradigm previously used with the infants on a healthy adult in both passive and active conditions. Eight different classifier configurations were compared.
None of the configurations was able to significantly classify the infant data above chance level on average. Still, there was a significant positive correlation to an overall attentiveness score. Results for the adult on the same passive paradigm were similar. For the active oddball parts of experiments 2 and 3, the classifier worked well, consistently reaching BACC scores for standard versus target detection above .8 and .9, respectively. Even for standard versus distractor classification in experiment 2, BACC scores of above .65 were consistently larger than in our passive oddball paradigm. Results suggest that attention to the stimuli is necessary for classification to work. Future research could manipulate stimuli to be intrinsically interesting to target groups like infants to increase classification performance in passive paradigms.