Understanding Cognitive Load in Digital Cultural Heritage: An Explainable AI-Driven Multimodal Analysis

Digital cultural heritage (DCH) platforms deliver complex multimedia narratives, but their cognitive demands remain insufficiently evaluated. We recorded electroencephalography (EEG) and eye tracking (ET) from 19 valid participants (of 30 recruited) interacting with seven authentic material formats from a national public service platform. Labels were three tiers derived from material-level NASA-TLX means; modeling used leave-one-subject-out validation (LOSO) with data-driven steps fitted within training folds. The primary three-tier model, using ET with random forest, reached 0.669 accuracy (permutation p = 0.0033; 95% CI [0.595, 0.743]; macro-F1 0.66). A complementary two-tier analysis retaining all trials reached 0.737 accuracy (macro-F1 0.73). Nested LOSO with inner-fold model selection yielded three-tier accuracy 0.632 (95% CI [0.564, 0.707]; macro-F1 0.630) and two-tier accuracy 0.684 (95% CI [0.632, 0.744]; macro-F1 0.676). EEG differentiated material formats at the group level (frontal theta, p = 0.0023) but did not improve trial-level classification. Fold-internal TreeSHAP and permutation importance converged on attentional narrowing under high load: shorter saccades, longer dwell, and fewer unique gaze regions. Objective information-density measures attributed the highest load to cross-channel integration demand, not information amount. These findings support design principles for cognitively sustainable digital cultural platforms and delineate the evidentiary boundaries of small-sample multimodal load modeling.