NEUROSLEEP

Neuromorphic Sleep Stage Scoring

100-Subject Benchmark

Sleep-EDF Expanded · 92 subjects · 10-fold subject-level CV · 3 seeds (42, 43, 44) · 30 folds per method

Accuracy 87.7% ± 2.7%

Cohen's κ 0.763 ± 0.043

Macro F1 0.730 ± 0.037

Parameters 99,477

Per-Class F1 (Full Fine-Tuning)

Stage	F1	Precision	Recall
Wake	0.966 ± 0.014	0.995	0.939
N1	0.450 ± 0.060	0.335	0.710
N2	0.773 ± 0.038	0.880	0.691
N3	0.683 ± 0.112	0.569	0.889
REM	0.779 ± 0.072	0.775	0.794

Adaptation Method Comparison

Model	Params	Accuracy	κ	Macro F1
Frozen Base	0 (0%)	87.1% ± 3.6%	0.738 ± 0.077	0.673 ± 0.074
LoRA CNN+Head (r=8)	1,448 (1.43%)	83.6% ± 3.7%	0.693 ± 0.057	0.674 ± 0.045
Full Fine-Tuning	99,477 (100%)	87.7% ± 2.7%	0.763 ± 0.043	0.730 ± 0.037

Architecture

PSG Input (Fpz-Cz, Pz-Oz, EOG, EMG)  [B, 10, 4, 3000]
      ↓
Multi-Resolution Stem (2 parallel Conv1d)
      ↓
Depthwise-Separable CNN (2 blocks)
      ↓
Parametric Gabor Feature Extraction (8 filters)
      ↓
Feature Fusion → [B, 10, 272]
      ↓
2-Layer GRU (hidden=64, 300s context)
      ↓
5-Class Softmax → Wake / N1 / N2 / N3 / REM

Quick Start

from huggingface_hub import hf_hub_download
from safetensors.torch import load_file

path = hf_hub_download(
    repo_id="shamique/Light-Weight-Neuromorphic-Sleep-Stage-Model",
    filename="student_full_finetuned.safetensors",
)

Resources

Model Hub GitHub Kaggle