A Clinically Informed Cross-Attention Deep Learning Framework for Automated Diagnosis of Obstructive Sleep Apnea

AI Summary

Manual PSG AHI annotation is time-consuming and costly; traditional models use handcrafted features and may not capture complex physiological signal dynamics.
Proposed framework: deep encoder for epoch-level PSG feature learning and classification, then multi-head cross-attention integrates clinical variables for patient-level AHI prediction.
Validated across four large real-world datasets, demonstrating improved robustness and generalisability compared with single-dataset deep learning approaches.

TL;DR (AI Generated)

Deep learning models can learn nonlinear representations directly from raw PSG signals, but many existing works study PSG independently and give less consideration to clinical variables that are known to influence OSA risk, e.g., age, sex.

To tackle these limitations, this project proposes a clinical knowledge-guided cross-attention deep learning framework for automated PSG annotation and AHI prediction from overnight PSG recordings, combined with clinical variables.

The proposed framework employs a deep encoder to learn epoch-level features from PSG data and classify epochs based on whether an adverse respiratory event occurs, providing epoch-level PSG annotations that mimic manual evaluation in current practice.

Basic summary

IEEE Int Conf Automation Sci Eng (CASE). 2026;2026:1057-1064. doi: 10.1109/case69030.2026.11704404. Epub 2026 Oct 6.

ABSTRACT

Obstructive sleep apnea (OSA) is a highly prevalent breathing-related sleep disorder associated with a variety of cardiovascular comorbidities. Polysomnography (PSG) is the clinical gold standard for OSA diagnosis, with disease severity quantified by the apnea hypopnea index (AHI). However, deriving AHI requires manual annotation of overnight PSG recordings by certificated technicians, making the OSA diagnostic process costly and time-consuming. While it is highly desirable to automate this process with a reliable machine learning pipeline, conventional models often rely on handcrafted feature engineering and may therefore not sufficiently capture the complex dynamics of physiological signals. Deep learning models can learn nonlinear representations directly from raw PSG signals, but many existing works study PSG independently and give less consideration to clinical variables that are known to influence OSA risk, e.g., age, sex. More importantly, most prior research has focused on a single dataset, raising concerns about the robustness and generalizability of these models. To tackle these limitations, this project proposes a clinical knowledge-guided cross-attention deep learning framework for automated PSG annotation and AHI prediction from overnight PSG recordings, combined with clinical variables. The proposed framework employs a deep encoder to learn epoch-level features from PSG data and classify epochs based on whether an adverse respiratory event occurs, providing epoch-level PSG annotations that mimic manual evaluation in current practice. Then, the epoch-level embeddings will be integrated with clinical variables using a multi-head attention mechanism for patient-level AHI prediction. The framework is validated across four large real-world datasets.

PMID:43023152 | PMC:PMC13759220 | DOI:10.1109/case69030.2026.11704404

Document this CPD