EEG Decoding of Finger Numeral Configurations With Machine Learning
In this study, we used multivariate decoding methods to study processing differences between canonical (montring and count) and noncanonical finger numeral configurations (FNCs). While previous research investigated these processing differences using behavioral and event-related potentials (ERP) methods, conventional univariate ERP analyses focus on specific time intervals and electrode sites and fail to capture broader scalp distribution and EEG frequency patterns. To address this issue a supervised learning classifier—support vector machines (SVM)—was used to decode ERP scalp distributions and alpha-band power for montring, counting, and noncanonical FNCs (for integers 1 to 4). The SVM was used to test whether the numerical information presented in FNCs can be decoded from the EEG data. Differences in the magnitude and timing of accuracy rates were used to compare the three types of FNCs. Overall, the algorithm was able to predict numerical information presented in FNCs beyond the random chance level accuracy, with higher rates for ERP scalp distributions than alpha-power. Montring had lower peak accuracy compared to counting and noncanonical configurations, likely due to automaticity in processing montring configurations leading to less distinct scalp distributions for the four numerical magnitudes (1 to 4). Paralleling the response time data, the peak decoding accuracy time for montring was earlier for montring (472 ms), compared to counting (577 ms) and noncanonical FNCs (604 ms). The results provide support for montring configurations being processed automatically, somewhat similar to number symbols, and provide additional insights for processing differences across different forms of FNCs. This study also highlights the strengths of decoding methods in EEG/ERP research on numerical cognition.