Decoding Cognitive Performance from EEG Using Energy-Based and Biophysical Models

Understanding the neural mechanisms underlying cognitive performance requires models that gobeyond descriptive EEG features toward interpretable representations of brain dynamics. In thisstudy, we applied energy landscape analysis, attractor modeling, and neural mass simulations toEEG data from 36 individuals during rest and a paced counting task, with the goal of identifyingneural markers distinguishing high and low performers.Our energy-based analysis of alpha and beta amplitude envelopes suggested that good per-formers may exhibit broader energy ranges and more local minima, indicating a trend towardincreased metastability. However, none of these comparisons reached statistical significance,and all high-density intervals (HDIs) included zero. Hopfield network-based attractor modelingyielded small, non-significant differences favoring the good group in attractor count and basinentropy across several bands, suggestive of richer but unconfirmed dynamical capacity.In contrast, circuit-level modeling with the Jansen-Rit neural mass model revealed crediblegroup differences in two parameters recurrent connectivity strength (C) and background input(μ) both showing HDIs excluding zero. These effects were consistent across rest and task condi-tions and may reflect enhanced cortical excitability and coupling in high performing individuals,though multiple testing remains a caveat.Finally, we trained an XGBoost classifier on extracted model features using synthetic bal-ancing (SMOTE) to address class imbalance. The model achieved 81% test accuracy (95% CI:[62%, 94%]), indicating some predictive utility. However, this result lacks comparison to simplerbaselines and should be interpreted as preliminary given the small and imbalanced dataset.In sum, while most findings were exploratory and limited by sample size, certain convergingtrends point to the potential of integrative modeling for understanding individual differences inbrain function. Future work should prioritize replication, statistical correction, and comparisonsacross simpler and more scalable methods to validate and refine these candidate neural markers.