Multi-Horizon Forecasting of Apnea and Hypopnea Events from Polysomnographic Signals

Status: 🚧 Paper under review. Code and data will be released upon publication. 🚧

This repository accompanies the manuscript:

"Multi-Horizon Forecasting of Apnea and Hypopnea Events from Polysomnographic Signals" Gurleen Kaur, Mohammadreza Hajipour, AJ Marcus Hirsch Allen, Najib T. Ayas, and Ghassan Hamarneh Submitted to IEEE Journal of Biomedical and Health Informatics

Authors and Affiliations

• Gurleen Kaur — School of Computing Science, Simon Fraser University, Burnaby, BC, Canada (gka82@sfu.ca)
• Mohammadreza Hajipour — Brigham and Women's Hospital, Harvard Medical School, Harvard University, Boston, MA, USA (mhajipour@bwh.harvard.edu)
• AJ Marcus Hirsch Allen — Department of Medicine, University of British Columbia, Vancouver, BC, Canada (ajhirschallen@gmail.com)
• Najib T. Ayas — Department of Medicine, University of British Columbia, Vancouver, BC, Canada (najib.ayas@vch.ca)
• Ghassan Hamarneh, Senior Member, IEEE — School of Computing Science, Simon Fraser University, Burnaby, BC, Canada (hamarneh@sfu.ca)

Overview

Obstructive sleep apnea (OSA) is a prevalent sleep-related breathing disorder affecting nearly one billion people worldwide. While polysomnography (PSG) is the clinical gold standard for detecting apnea and hypopnea events after they occur, this work explores whether such events can be forecast before onset to support anticipatory intervention.

We formulate apnea/hypopnea prediction as a multi-horizon binary time-series classification task and propose a hybrid deep learning architecture that fuses:

• Learned spectral-temporal EEG representations (via an EEGNet-inspired CNN pipeline), and
• Hand-crafted physiological descriptors (band-power, spectral statistics, and wavelet features)

through a cross-attention fusion mechanism.

Key Contributions

• A rigorous multi-horizon forecasting formulation (Δ ∈ {5, 10, 20, 30} seconds) with explicit post-event exclusion and guard-margin constraints, distinguishing true pre-event forecasting from during-event detection.
• A cross-attention fusion architecture combining CNN-based and analytical (hand-crafted) feature branches.
• Evaluation on 477 full-night PSG recordings (University of British Columbia dataset) using subject-wise cross-validation — a substantially larger and more diverse cohort than prior EEG-inclusive forecasting studies.
• Interpretability analysis via SHAP to characterize channel-level contributions (e.g., nasal pressure, respiratory effort, EEG).
• Uncertainty-based selective prediction (abstention) to improve reliability on high-confidence subsets.

Results Summary

Horizon (Δ)	F1 Score	Accuracy (%)
5 s	0.73	71.44
10 s	0.71	69.56
20 s	0.68	64.06
30 s	0.69	63.50

Performance is highest near event onset and decreases gracefully as the forecasting horizon increases, consistent with the expected reduction in available predictive signal at longer lead times.

Dataset

This study uses a private clinical PSG dataset provided by the University of British Columbia (477 full-night recordings). The dataset is not publicly available due to privacy and institutional data-sharing restrictions.

Repository Structure (planned)

.
├── data/            # Data loading and preprocessing scripts (dataset not included)
├── models/          # Model architecture definitions (EEGNet branch, analytical branch, cross-attention fusion)
├── training/        # Training and cross-validation scripts
├── analysis/        # SHAP interpretability and abstention/calibration analysis
└── README.md

(Structure subject to change as the code release is finalized.)

Citation

A formal citation will be added once the paper is accepted/published. In the meantime, please reach out to the corresponding authors for reference details, or use the placeholder BibTeX entry below (update once accepted):

@article{kaur2026multihorizon,
  title   = {Multi-Horizon Forecasting of Apnea and Hypopnea Events from Polysomnographic Signals},
  author  = {Kaur, Gurleen and Hajipour, Mohammadreza and Hirsch Allen, AJ Marcus and Ayas, Najib T. and Hamarneh, Ghassan},
  journal = {IEEE Journal of Biomedical and Health Informatics},
  year    = {2026},
  note    = {Under review},
}

Contact

• Gurleen Kaur — gka82@sfu.ca
• Ghassan Hamarneh — hamarneh@sfu.ca

Acknowledgments

This work was supported by computational resources from the Digital Research Alliance of Canada.

---

This README is a placeholder and will be updated with full documentation, setup instructions, and code upon completion of the peer-review process.