Attention-Based Temporal Resolution Enhancement for Microsleep Forecasting Across Multiple Horizons

Detecting microsleep events, that is, brief involuntary lapses in attention, is critical forsafety in tasks such as driving and industrial control. Current EEG-based methods achieve high accuracy (AUC–ROC ≈ 0.95) when detecting microsleep onset in real-time but show reduced predictive sensitivity (17–35% drop) when forecasting events 1–2 seconds ahead, primarily due to diminishedtemporal resolution of EEG signals. Prediction ahead of microsleep episodes is desirable for utility inoperator assessment scenarios.To address this challenge, we propose an attention-augmented temporal modeling approach that applies a causal multi-head self-attention mechanism before a Gated Recurrent Unit (GRU) network to preserve fine-grained temporal features. In addition, we incorporate a focal loss function to emphasize minorityclass microsleep events and mitigate class imbalance. Experiments on the multi-channel MWT dataset demonstrated consistent and statistically significantimprovements over both CSP+SBL and conventional GRU baselines, with sensitivity gains of up to 42 percentage points and AUC–ROC improvements of up to 22 points (Wilcoxon test, p = 0.0003). An ablation study confirmed that both focal loss and attention contribute complementary benefits, withtheir combination yielding the best overall results. To evaluate robustness, we further tested on the single-channel Sleepy Driver dataset, where the proposed model improved specificity and AUC metrics, though gains were more modest than in clinical multi-channel EEG.Overall, our findings demonstrate that integrating causal attention and focal loss sustains forecasting performance across horizons and datasets, offering a reliable and adaptive framework for real-time microsleep prediction in both clinical and consumer-grade settings.