Multimodal neurocognitive assessment of internet gaming disorder using ERP and fNIRS during an oddball paradigm: A machine learning-based classification

Excessive internet gaming has been associated with social, cognitive, and behavioral impairments; however, the neural mechanisms underlying Internet Gaming Disorder (IGD) remain insufficiently explored. This study examined neurophysiological differences among individuals with IGD, recreational game users (RGU), and healthy controls (HC). A total of 125 male undergraduate participants (54 IGD, 29 RGU, 42 HC; aged 18–23 years) completed a visual/auditory oddball task while simultaneous event-related potentials (ERP) and functional near-infrared spectroscopy (fNIRS) signals were recorded. ERP features included P300 and N200 amplitudes and latencies, whereas fNIRS features consisted of General Linear Model (GLM) beta coefficients and statistical descriptors (e.g., mean, standard deviation, maximum peak, kurtosis, skewness, slope). Following ReliefF-based feature selection, multiple machine learning algorithms were trained using feature subsets ranging from 10 to 100 features. ReliefF identified P300 amplitude and latency, N200 latency, and several fNIRS-derived statistical features (standard deviation, maximum peak, skewness, GLM beta values) as the most discriminative markers. The highest accuracy (88.17%) was achieved with the top 80 multimodal features and a Random Forest classifier, while ERP-only features reached 82.57%. To our knowledge, this is the first study to classify IGD, RGU, and HC using multimodal ERP–fNIRS recordings during an oddball task assessing divided attention. The findings demonstrate that multimodal integration substantially improves classification performance, underscoring its potential to develop objective neurobiological markers of problematic gaming behavior