SomnoNet

A simple, efficient, and interpretable sleep staging method (SOTA)

SomnoNet is an extremely lightweight yet highly effective deep learning framework for automatic sleep staging.

---

📊 Benchmarking Results

Dataset		Overall			F1 score
		OA	MF1	k	W	N1	N2	N3	R
Physio2018	SomnoNet	80.9	79.0	0.739	84.6	59.0	85.1	80.2	86.3
	SleePyCo	80.9	78.9	0.737	84.2	59.3	85.3	79.4	86.3
	SomnoNet-nano	80.5	78.6	0.734	84.0	57.8	84.8	80.2	86.2
	XSleepNet	80.3	78.6	0.732	-	-	-	-	-
	SeqSleepNet	79.4	77.6	0.719	-	-	-	-	-
	U-time	78.8	77.4	0.714	82.5	59.0	83.1	79.0	83.5
SHHS	SomnoNet	88.1	80.8	0.833	93.1	48.6	88.6	85.0	88.6
	SleePyCo	87.9	80.7	0.830	92.6	49.2	88.5	84.5	88.6
	SleepTransformer	87.7	80.1	0.828	92.2	46.1	88.3	85.2	88.6
	XSleepNet	87.6	80.7	0.826	92.0	49.9	88.3	85.0	88.2
	SomnoNet-nano	87.4	79.0	0.822	92.3	43.0	88.0	83.9	87.9
	IITNet	86.7	79.8	0.812	90.1	48.1	88.4	85.2	87.2
	SeqSleepNet	86.5	78.5	0.81	-	-	-	-	-

Table1. Benchmarking against recent state-of-the-art methods

---

📦 Model Size Comparison

Fig1. EEG data with different adoption rates

Method	parameter
SomnoNet-nano	0.049M
SomnoNet	0.43M
SalientSleepNet	0.9M
U-time	1.1M
TinySleepNet	1.3M
SleepEEGNet	2.1M
DeepSleepNet	21M

Table2. The number of model parameters for different methods

We additionally provide three interpretability visualization notebooks for understanding the learned features of SomnoNet, along with example data and corresponding models.

🔍 New: Interpretability & Visualization

To better understand how SomnoNet extracts discriminative features for sleep staging, we provide three visualization Jupyter notebooks:

✔ Interpretability_voting.ipynb

✔ Interpretability_vector_forward.ipynb

✔ Interpretability_vector_backward.ipynb