NARCOX: Non-stationary auto-regressive transformers for cognitive modeling with ocular exogenous input

Jayanth Shreekumar, Sabine Brunswicker

Abstract:

The ability to accurately predict mental effort required for processing information by engaging cognitive resources is crucial in safety critical domains where timely intervention is needed to prevent cognitive overload and ensure human safety. Established approaches for cognitive load measurement face two key limitations: (1) they typically rely on discrete and sparsely sampled measures, overlooking the continuous, non-linear, and non-stationary nature of working memory engagement over time, and (2) most methods depend on EEG recordings for ground-truth labels, which is impractical for real-world deployment. To address these limitations, this paper frames cognitive load prediction as a time-series regression task and proposes NARCOX, a transformer-based framework for Non-linear Autoregressive time-series regression with eXogenous inputs (NARX) that processes ocular inputs to predict TAR, a validated, EEG-derived, continuous marker for cognitive load. We propose
NARCOX transformer variants operating in the frequency and wavelet domains and demonstrate on a synchronized EEG–gaze dataset that they outperform the time-domain LSTM and transformer baselines by capturing the evolving spectral characteristics of gaze and EEG signals.