MM-Net: A Multimodal Multi-Task Network for Sleep Staging and Sleep Apnea Assessment After Ischemic Stroke

Md Wahiduzzaman Suva; Esm E Moula Chowdhury Abha; Md Imtiaj Alam Sajin; Md Iftekharul Mobin; Jia Uddin

cs.LGOct 27, 2026

MM-Net: A Multimodal Multi-Task Network for Sleep Staging and Sleep Apnea Assessment After Ischemic Stroke

Authors: Md Wahiduzzaman Suva, Esm E Moula Chowdhury Abha, Md Imtiaj Alam Sajin, Md Iftekharul Mobin, Jia Uddin

Organizations: Department of Computer Science and Engineering, American International University-Bangladesh, Dhaka, 1229, Bangladesh · AI and Big Data Department, Woosong University, Daejeon, Republic of Korea

Abstract

Automated polysomnography after ischemic stroke is challenging because of altered sleep-related neural activity and concurrent respiratory disturbances. Existing stroke-specific approaches mainly address sleep-stage classification, leaving joint analysis of sleep architecture and respiratory events underexplored. Moreover, adapting pretrained EEG encoders introduces a representation inconsistency when downstream models are trained on embeddings from previously seen patients but evaluated on unseen individuals. We propose MM-Net, a multimodal multi-task framework that resolves this inconsistency through cross-fitting and ridge-based embedding alignment. Sleep-adapted EEG-EOG representations and engineered neural descriptors are fused with learned cardiorespiratory features through cross-attention and shared temporal modeling for simultaneous sleep staging and respiratory-event detection. On iSLEEPS (99 patients with subacute ischemic stroke), ten-fold patient-independent evaluation achieved 76.3% staging accuracy (Cohen's $κ$ = 0.656) and 0.825 respiratory-event AUC, outperforming the evaluated baselines under matched protocols. The sleep-adapted encoder improved the downstream model when its embeddings were aligned, raising staging accuracy by 1.4 percentage points over the frozen encoder ( $p$ = 0.020) versus 0.2 points without alignment. MM-Net also achieved 86.4% staging accuracy on Sleep-EDF, above every single-EEG architecture reimplemented on the same recordings, and 80.3% staging accuracy with 0.889 respiratory AUC on ISRUC-I, exceeding its retrained staging baselines. Without retraining, the iSLEEPS-trained model retained respiratory AUCs of 0.833 on ISRUC-I and 0.841 on UCD. These findings support representation-consistent neural adaptation and joint physiological modeling across stroke, healthy-sleep, and sleep-disorder cohorts.

Figures & tables

Dataset	Population	Data used	Role
iSLEEPS [ 31 ]	Subacute ischemic stroke	99 patients	Development and in-domain evaluation
Sleep-EDF [ 27 , 16 ]	Healthy sleep	40 recordings	Reimplementation checks, healthy-sleep reference, and external staging
ISRUC-Sleep I [ 28 ]	Sleep disorders	100 patients	Cross-validation and transfer (staging, respiratory, AHI)
ISRUC-Sleep II [ 28 ]	Sleep disorders	8 subjects, two nights each	Cross-validation and transfer
ISRUC-Sleep III [ 28 ]	Healthy	10 subjects	Cross-validation and transfer
UCD [ 40 ]	Suspected sleep apnea	25 patients	Transfer (staging, respiratory, AHI)

Table 1: Datasets used in the study and their roles in the evaluation.

Stage	W	N1	N2	N3	REM
Share (%)	26.7	10.1	42.3	8.9	12.1

Table 2: Sleep-stage distribution in the analyzed iSLEEPS cohort (

N=99

;

92{,}560

epochs).

Figure 1: Patient-level respiratory-event burden in iSLEEPS, measured as the fraction of 30-s epochs containing the onset of a scored respiratory event. The dashed line denotes the cohort median of

13\%

Figure 2: The two-stream, multi-task model. The neural stream combines $188$ engineered descriptors with the embedding of a LaBraM encoder adapted to sleep from EEG and EOG and integrated by cross-fitting and alignment; the cardiorespiratory stream is a convolutional encoder over eight raw channels. The streams are fused and read by a bidirectional LSTM over a $20$ -epoch window, which feeds a staging head and a respiratory head; the respiratory head also reads the cardiorespiratory embedding directly, and the AHI is computed from the respiratory scores over decoded sleep.

Figure 3: An illustrative held-out night (patient SN90, three-seed MM-Net, staging accuracy $0.83$ ). (A) Predicted and reference hypnograms, REM highlighted. (B) EEG spectrogram, C4 derivation, $0$ – $25$ Hz. (C) Per-epoch stage-probability ribbon; the black line is the maximum posterior.

Figure 4: Row-normalized confusion matrix for MM-Net (three-seed, pooled ten-fold test epochs, hidden-Markov operating point). Errors concentrate on N1, which is absorbed into Wake and N2.

Model	W $\uparrow$	N1 $\uparrow$	N2 $\uparrow$	N3 $\uparrow$	R $\uparrow$
MM-Net	0.81	0.20	0.82	0.72	0.79

Table 5: Per-class staging F1 (Wake, N1, N2, N3, REM), ten-fold patient-independent, at the hidden-Markov operating point (MM-Net: three-seed, pooled test epochs).

Figure 5: Per-fold staging accuracy of MM-Net with the sleep-adapted encoder, relative to the frozen-encoder reference (three-seed, ten patient-independent folds). The training embeddings are cross-fitted in both cases; aligning them to the deployed encoder turns an inconsistent effect into a gain in eight folds of ten.

Figure 6: t-SNE [ 48 ] of

4000

epochs from the ten held-out patients of one fold, colored by scored stage. Left: frozen LaBraM. Right: the sleep-adapted, aligned EEG/EOG encoder used in MM-Net.

Figure 7: Respiratory-event detection by MM-Net (three-seed), pooled over the ten patient-independent test folds. (a) ROC curve. (b) Precision–recall; the dashed line is the

0.160

event prevalence, the no-skill floor.

(a) Detection method
Method	AUC $\uparrow$	AP $\uparrow$
Desaturation rule	0.595	0.207
Logistic regression	0.603	0.211
Gradient boosting	0.673	0.290
Raw effort belt § [ 33 ]	0.602	0.252
Raw seven channels §	0.733	0.366

Table 7: Respiratory-event detection on iSLEEPS. (A) Comparison with classical and raw-signal baselines on the same folds ( § single seed) and (B) MM-Net AUC by respiratory-event type. Bold denotes the best value in each comparison.

Modality removed	Output	Change	Holm $p$
EEG (neural)	Staging accuracy	$-0.050$	$0.018$
All cardiorespiratory	Respiratory AUC	$-0.118$	$0.018$
Respiratory effort	Respiratory AUC	$-0.028$	$0.031$
Airflow	Respiratory AUC	$-0.004$	$0.65$

Table 9: Modality removal and retraining (frozen-encoder configuration; ten folds, three seeds). Each row removes one modality and reports the change in the output that modality serves, relative to the full configuration (staging accuracy

0.736

, respiratory AUC

0.796

). Holm

p

: Wilcoxon on the ten fold means, Holm-corrected across the nine ablations of the grid.

Figure 8: MM-Net AHI estimate against the annotated AHI. Left: iSLEEPS, ten-fold test patients. Right: ISRUC-I, MM-Net trained on iSLEEPS only and applied without retraining (Section 5.10 ). Shading marks the AASM severity bands; the dashed line is identity.

AHI band	AHI (events/h)	Patients	Staging acc $\uparrow$
Normal	$<5$	15	$0.784\pm 0.057$
Mild	$5$ – $15$	23	$0.754\pm 0.092$
Moderate	$15$ – $30$	23	$0.764\pm 0.083$
Severe	$\geq 30$	38	$0.758\pm 0.078$

Table 11: Sleep-staging accuracy of MM-Net across AASM severity groups of the clinical AHI in iSLEEPS. Values are mean

\pm

SD across patients; the ordered-group trend is not significant (Jonckheere–Terpstra

p=0.20

Characteristic	Group	Patients	Staging acc $\uparrow$	Encoder gain
NIHSS	$0$ – $4$ (minor)	86	$0.758$	$+0.004$
	$\geq 5$ (moderate to severe)	12	$0.778$	$+0.028$
Modified Rankin	$0$ – $2$ (independent)	88	$0.757$	$+0.005$
	$3$ – $5$ (dependent)	10	$0.794$	$+0.024$

Table 12: Sleep-staging accuracy of MM-Net (three-seed, per patient) by stroke severity (NIHSS) and functional status (modified Rankin scale) in iSLEEPS. Encoder gain: accuracy difference from the frozen-encoder reference of Table 6 on the same patients.

Sleep Stage Classification

Sleep stage classification assigns sleep epochs to physiological stages from signals such as EEG, helping quantify sleep architecture and support clinical monitoring.

See all 40 papers

Mar 24, 2026

AI Generalisation Gap In Comorbid Sleep Disorder Staging

Accurate sleep staging is essential for diagnosing OSA and hypopnea in stroke patients. Although PSG is reliable, it is costly, labor-intensive, and manually scored. While deep learning enables automated EEG-based sleep staging in healthy subjects, our analysis shows poor generalization to clinical populations with disrupted sleep. Using Grad-CAM interpretations, we systematically demonstrate this limitation. We introduce iSLEEPS, a newly clinically annotated ischemic stroke dataset (to be publicly released), and evaluate a SE-ResNet plus bidirectional LSTM model for single-channel EEG sleep staging. As expected, cross-domain performance between healthy and diseased subjects is poor. Attention visualizations, supported by clinical expert feedback, show the model focuses on physiologically uninformative EEG regions in patient data. Statistical and computational analyses further confirm significant sleep architecture differences between healthy and ischemic stroke cohorts, highlighting the need for subject-aware or disease-specific models with clinical validation before deployment. A summary of the paper and the code is available at https://himalayansaswatabose.github.io/iSLEEPS_Explainability.github.io/

Saswata Bose, Suvadeep Maiti, Shivam Kumar Sharma +4

International Institute of Information Technology (IIIT-H), Hyderabad, India · Queen Square Institute of Neurology, University College London, United Kingdom · National Institute of Mental Health and Neuroscience (NIMHANS), Bangalore, India +1

Sleep Stage Classification Domain Generalization

Ask PDF

Aug 14, 2026

Comparative Analysis of State-of-the-Art Foundation Models for Sleep Analysis Under Channel Reduction

Automatic sleep staging from polysomnography (PSG) is a well-studied task, but PSG itself is expensive, clinic-based, and burdensome to manually score, which limits its use for long-term or at-home monitoring. Most existing sleep-staging foundation models are evaluated using the full PSG montage. We instead ask how much of that montage is actually necessary. We evaluate six sleep staging models on the Multi-Ethnic Study of Atherosclerosis (MESA) PSG dataset across three signal conditions: electroencephalography (EEG), electrocardiography (ECG), and their combination (EEG+ECG). This is motivated by edge-cloud deployment, where EEG requires a clinic-grade scalp electrode, whereas ECG is already captured by consumer wearables. We test state-of-the-art foundation models such as SleepFM with an encoder trained from scratch on MESA, alongside BIOT, MOMENT, LaBraM, a base-scale Vision Transformer (ViT-B) reimplementation of SensorLM trained from scratch, and YASA, spanning EEG-pretrained, general-time-series, from-scratch, and classical non-learned approaches. No model architecture is modified from its original form; SensorLM's encoder is reimplemented only in PyTorch. For EEG-only staging, BIOT achieves the best result with a macroF1 of 0.7237, followed by LaBraM (0.6835) and SleepFM from scratch (0.6582). Across the five models capable of ECG-only staging, switching from EEG to ECG costs between 0.2798 (MOMENT) and 0.4151 (BIOT) macroF1, averaging 0.3531, while cutting the raw channel data rate to a third. Adding ECG to EEG provides no gain for most models. These results show that EEG carries most of the sleep-staging signal, quantify the consistent accuracy cost of the wearable-compatible alternative, and demonstrate that sleep-relevant pretraining transfers well to MESA.

Hassan Mehdi, Riku Klen, Ayse Kosal Bulbul +5

Institute of Biomedicine, University of Turku, Finland · Turku PET Centre, University of Turku and Turku University Hospital, Finland · University College Cork, Ireland +3

Sleep Stage Classification EEG Foundation Models Time-Series Foundation Models

Ask PDF

Dec 16, 2025

AnySleep: a channel-agnostic deep learning system for high-resolution sleep staging in multi-center cohorts

Sleep is essential for health, yet studying its dynamics requires manual sleep staging, a labor-intensive step in research and clinical care. Across centers, polysomnography (PSG) recordings are traditionally scored in 30-s epochs for pragmatic, not physiological, reasons and vary in electrode count, montage, and subject characteristics. These constraints challenge harmonized multi-center studies and the discovery of robust biomarkers on shorter timescales. We present AnySleep, a deep neural network that scores sleep from any electroencephalography (EEG) or electrooculography (EOG) data at adjustable temporal resolutions. We trained and validated the model on over 20,000 overnight recordings (> 200,000 hours of EEG and EOG) from 28 datasets across multiple clinics to promote robust generalization across sites. The model attains state-of-the-art performance and surpasses or equals established baselines at 30-s epochs. Performance improves with more channels, yet remains strong when EOG is absent or only EOG or single EEG derivations (frontal, central, or occipital) are available. On sub-30-s timescales, the model captures short wake intrusions consistent with arousals and improves prediction of pathophysiological conditions (obstructive sleep apnea, narcolepsy type 1, insomnia) over 30-s scoring. We make the model publicly available to facilitate large-scale studies with heterogeneous electrode setups and accelerate biomarker discovery in sleep.

Niklas Grieger, Jannik Raskob, Siamak Mehrkanoon +1

Department of Medical Engineering and Technomathematics, FH Aachen University of Applied Sciences, 52428 Jülich, Germany · Department of Information and Computing Sciences, Utrecht University, Utrecht, The Netherlands · Institute for Data-Driven Technologies, FH Aachen University of Applied Sciences, 52428 Jülich, Germany

Electroencephalography Sleep Stage Classification EEG Decoding Domain Generalization

Ask PDF

Abstract

Sleep Stage Classification

Explore similar work

AI Generalisation Gap In Comorbid Sleep Disorder Staging

Comparative Analysis of State-of-the-Art Foundation Models for Sleep Analysis Under Channel Reduction

AnySleep: a channel-agnostic deep learning system for high-resolution sleep staging in multi-center cohorts