SleepISI: A Neuromorphic Time-Domain EEG Classification Framework Using Inter-Spike Intervals From AdEx Encoding of C4-A1 Signals

Accurate sleep stage classification is vital in ICU (neuro-ICU) for assessing consciousness, monitoring sedation, and informing treatment strategies. Most automated approaches rely on frequency-domain analysis and deep neural networks-requiring substantial preprocessing and computation, which limits their suitability for real-time deployment in hardware-constrained ICU environments. This paper presents SleepISI, a neuromorphic, preprocessing-free, time-domain framework for sleep staging using single-channel (C4-A1) raw EEG signals. The signal is encoded into spike trains via a low-power Adaptive Exponential Integrate-and-Fire (AdEx) neuron model, where inter-spike intervals (ISIs) are inversely proportional to signal amplitude. Ten ISI-based features-mean, median, Q1, Q3, IQR, ZCR, CSDM $_{\delta _{\rm mean}}$ and CSDM$_{\text{activity}}$, CSDM$_{\text{mobility}}$, CSDM$_{\text{complexity}}$ - are extracted over fixed 10 s windows and 30 s values were evaluated by aggregating three successive 10 s values. Feature relevance is verified via the Kruskal-Wallis test and classification is performed through a four-level hierarchical Random Forest model to differentiate Wake, REM and Non-REM stages (S1, S2, SWS). The proposed system achieves 86.23% accuracy on the CAP database, 82.13% on the SleepEDF-20 database, and 91.87% on the ISRUC-S3 database using a single EEG channel and no preprocessing. FPGA/45-nm ASIC prototypes confirm real-time feasibility at 70.77 $\mu$W. SleepISI offers a compact, interpretable, and energy-efficient solution for continuous EEG-based sleep monitoring in ICUs and resource-constrained healthcare settings.