Abstract
Abstract Background Is it possible to decode lying behavior from neural signatures of the electroencephalography (EEG)? Existing studies on this topic have methodological limitations: tasks lack the incentive to lie, the act of lying is confounded with memory recollection, there is no sufficient distinction between instructed versus spontaneous decisions, and participants' risk-taking tendency is not controlled for. To address these limitations, we introduce a novel interactive, two-player game, where successful lying is incentivized in the reward scheme and that has both instructed and spontaneous conditions for participants matched by risk-taking tendency. Methods 24 participants were paired in the game according to their risk-taking tendency scores and measured using 32-channel EEG. Our multi-modal EEG analysis includes event-related potential (ERP), event-related spectral (ERS), and deep-learning-based single-trial decoding. Results In ERP, two early components (P200 and N200) distinguished instructed truth-telling from other conditions, with a late component (N300) separating instructed lies from spontaneous conditions. Moreover, a late positive potential was indicative of spontaneous lying versus spontaneous truth-telling and was correlated with participants' risk-taking tendencies. In ERS, alpha and low beta were found to discriminate conditions. Importantly, we observed robust single-trial decoding performance under different conditions for the 1D-CNN. A gradient-based analysis further identified significant time-periods compatible with the ERP results from the classifier. Conclusions Our study represents the first effort to analyze EEG data not only through statistical properties but also with out-of-sample, deep-learning-based classification to decode deceit in the context of an iterative, game-theoretic experiment.