A normative model-based assessment framework for large-scale, multi-site EEG data.

Electroencephalography (EEG) overcomes the subjectivity inherent in questionnaire-based and observational assessments. However, most existing EEG-based evaluation methods still impose discrete categorical states onto continuously varying neural dynamics, thereby neglecting the continuity of states. With the rise of neuroscience alliances, challenges such as batch-effects across datasets and inconsistencies introduced by diverse EEG electrode montages have become increasingly prominent. Therefore, a robust assessment framework that accommodates large‑scale, multi‑site EEG data is expected. A normative model-based assessment framework was developed for large-scale, multi-site EEG data, with attention assessments used as illustrative examples. Normative models are first constructed using EEG features from 1212 young individuals, and quantile ranks are computed. Next, feature selection is performed, and elastic net regression and support vector regression are used to model distributed attention (DA) and focused attention (FA). The results from normative model-based features are compared with original features to demonstrate the advantage of quantile rank features. Finally, the model's test-retest reliability and generalizability are assessed. The framework identifies statistical differences (q < 0.05) in attention performance between the top and bottom 20 % participants on attention scales. EEG features demonstrated specific patterns related to accuracy and reaction time in both DA and FA tasks. The normative model outperformed in predictive tasks, showing enhanced stability and interpretability. Additionally, the framework demonstrates strong test-retest reliability and robust generalizability (ICC > 0.9). In conclusion, we proposed a normative model-based framework that harmonizes large‑scale, multi‑site EEG data, enabling efficient and reliable attention assessment while demonstrating promise for broader EEG‑based applications.