Interpreted EEG-based sleep staging model with explanation of decisions using attention maps and analysis of physiological patterns of sleep spindles and K-complexes

Abstract

The developed interpretable neural network architecture for automatic sleep staging from electroencephalography signals integrates a convolutional encoder with a two-level multi-head attention mechanism. This ensures both high classification accuracy for the five sleep stages and the ability to visualize and quantitatively analyze the foundations of decision-making through attention maps. The model's internal representations, extracted in the form of attention maps at the intra-epoch and inter-epoch levels, demonstrate a pronounced specialization of individual attention heads in relation to the key physiological graphoelements of stage N2: the third head exhibits more than a threefold increase in the normalized energy contribution with a mean value of 3.561 in the vicinity of sleep spindles, which are characterized by oscillations in the 11-16 Hz range, while the first head demonstrates predominant sensitivity to K-complexes with biphasic morphology and a mean contribution of 3.214. This is confirmed by functional dissociation and the statistical significance of the results. Quantitative indicators of consistency between the attention maps and expert-verified positions of these patterns, including an integral measure of exceeding the baseline attention level by a factor of 3.388, sensitivity of 0.741, and specificity of 0.823 for spindles under threshold binarization, indicate that classification relies directly on physiologically grounded features that correspond to the neurophysiological mechanisms underlying the generation of these graphoelements. The inter-epoch attention component reproduces known macrostructural sleep patterns, revealing variable contextual dependence of stages – from a minimal effective context width of 3.12 epochs and a high proportion of attention to the current epoch for N3 to a maximal width of 7.81 epochs with pronounced asymmetry toward preceding epochs for N1 – which precisely reflects clinical practice in differential diagnosis. The overall classification quality metrics on a large array of recordings reach a macro-averaged F1-score of 0.808 and an overall accuracy of 0.876, with the expected reduction in metrics for stage N1 due to its transitional character and class imbalance. The high reproducibility of the identified patterns of attention head specialization across multiple training runs with different initial conditions, a moderate Spearman correlation of 0.437 between spindle amplitude and combined attention contribution, and the stability of results within cross-validation confirm the structural isomorphism between the model's internal mechanisms and expert somnological knowledge. This level of transparency and verifiability of decisions meets stringent regulatory requirements for software as a medical device, opening prospects for integration into decision support systems in sleep laboratories, application in telemedicine conditions for remote monitoring, and the use of binarized attention maps as an auxiliary tool for semi-automatic detection and annotation of physiological events associated with cognitive health assessment. The proposed approach to the quantitative assessment of interpretability through normalized energy contribution and correlation analysis with physiological annotations establishes a methodological basis for standardizing artificial intelligence validation in biomedical time series analysis tasks, substantially increasing confidence in automated tools in neurophysiology and related fields of clinical medicine.