Q-MeLoNet: A Quantum-Enhanced Transformer Framework for Adolescent Mental Dysfunction Detection from fNIRS Signals

Abstract

Abstract
Timely intervention of affective and cognitive dysfunctions in teenagers necessitates early and accurate identification of mental workload states. However, this has remained a perennial issue owing to subjective diagnostic approaches and limited availability of neuroimaging suites. In this paper, Q-MeLoNet, an advanced quantum-enhanced transformer model, is introduced for multi-class teenage mental workload classification from functional near-infrared spectroscopy (fNIRS) signals. Using the publicly available Tufts fNIRS2MW dataset, which captures prefrontal hemodynamics during a four-level n-back working memory task, our approach uses quantum encoding layers along with transformer-based self-attention to capture nuanced temporal relations in multichannel fNIRS time series. Model performance was strictly evaluated using 10-fold cross-validation. Q-MeLoNet achieved a mean accuracy of 91.2%, well above traditional baselines such as CNNs and BiLSTMs, as well as ablated variants of itself. An extensive ablation study revealed that the transformer and quantum components both significantly affected classification performance. Statistical testing confirmed these gains were significant (p < 0.01). Visualization via confusion matrices highlighted the model's ability to distinguish between fine-grained workload states, particularly towards the extremes (0-back and 3-back). Compared to previous work with limited binary classification, Q-MeLoNet presents a data-driven and scalable approach for multi-class cognitive classification across adolescent groups, advancing non-invasive neurophysiological investigation.