AIM-STAGE: a self-supervised framework via adaptive interval-aware multi-kernel calibration and spectro-temporal exploration for few-shot EEG staging

Objective: Learning discriminative and generalizable representations from electroencephalography (EEG) signals with limited annotations remains a key challenge in automatic sleep staging. Supervised deep models depend on large-scale expert labels, yet sleep annotation is costly and time-consuming. Existing self-supervised methods often underuse EEG spectral-temporal structure and adopt fixed negative-sample strategies, which are poorly suited to the continuous transitions and ambiguous boundaries between sleep stages. This study proposes AIM-STAGE, a self-supervised representation learning framework for limited-label EEG sleep staging. Approach: AIM-STAGE comprises three modules. (1) Spectral-temporal exploration: maps EEG segments into the time-frequency domain through the short-time Fourier transform (STFT), jointly extracting log-power and cross-channel phase difference to characterize frequency-band activity and phase dynamics. (2) Local-global hybrid encoding: uses a CNN-Transformer encoder, in which convolutional residual modules capture local spectro-temporal patterns and a bottleneck Transformer models long-range dependencies. (3) Adaptive interval-aware multi-kernel calibration: employs multiple Gaussian kernels and dynamic weighting to synthesize informative negative representations, improving view consistency and inter-sample separability. Main results: AIM-STAGE was evaluated on SLEEP-2013, SLEEP-2018, SHHS-D, and SHHS-H against ContraWR, DFformer, EEG2Rep, EEGPT, Superlet-MAE, and a supervised baseline across 1% to 15% labels. Accuracy improved from 88.67% to 92.04% on SLEEP-2013, and AIM-STAGE outperformed all baselines on all four datasets with higher kappa, macro-F1, and balanced accuracy, including on the pathological SHHS-D cohort. Ablation confirmed the contributions of the bottleneck Transformer,positional embeddings, physiology-informed augmentation, and calibration loss. Per-stage analysis showed that N1 is the most difficult stage, with F1 ranging from 0.003 to 0.046, and adding EOG and EMG markedly improved its precision, recall, and F1. Significance: AIM-STAGE offers an interpretable and label-efficient self-supervised solution for EEG sleep staging, improving robustness and stage discriminability under scarce-label conditions while reducing dependence on manual annotation.