Loading...
0379 Inferring Possible Causal Relationships among Polysomnography Signals using Multi-Modal Causal Representation Learning
Sun, Yuewen Sun, Boyang Mazumder, Anirudh Dai, Haoyue Mazumder, Aneesh Westover, M Brandon Thomas, Robert Zhang, Kun Sun, Haoqi
Sun, Yuewen
Sun, Boyang
Mazumder, Anirudh
Dai, Haoyue
Mazumder, Aneesh
Westover, M Brandon
Thomas, Robert
Zhang, Kun
Sun, Haoqi
Files
Loading...
zsag091.0379.pdf
Adobe PDF, 74.25 KB
Supervisor
Department
Machine Learning
Embargo End Date
Type
Journal article
Date
License
http://creativecommons.org/licenses/by/4.0/
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
Introduction Sleep contains rich physiology that reflects the underlying multi-organ network biology. For example, obstructive sleep apnea leads to hypoxemia, cortical arousal, and sympathetic activation. However, inferring how physiologic signals influence each other remains challenging. We applied a novel causal representation learning (CRL) framework to uncover directed physiologic dependencies in sleep, providing data-driven insight into multi-organ dynamics. Methods We analyzed 2,576 participants from the Sleep Heart Health Study visit 2, which is a multi-center community cohort study with at-home polysomnography (PSG). We segmented the signals into 16-second epochs. Within each epoch, we extracted band powers averaged between C3-M2 and C4-M1 from electroencephalography (EEG); standard deviation (SD) of left minus right electrooculography (EOG); average envelope of chin electromyography (EMG); average and SD of heart rate from electrocardiography (ECG); average respiration rate from the airflow thermistor; and average SpO2. For each participant, CRL estimated a directed interaction graph among signals. This approach identifies plausible directionality (e.g., airflow→EEG) while permitting physiologic cyclic patterns. Associations between interaction frequency and age, body mass index (BMI), and apnea-hypopnea index (AHI) were tested separately in NREM and REM. Results The average age was 67.5 years, and 54% female. The relationships observed in more than half of the participants were airflow↔EEG (52%) and airflow→EMG (50%). During NREM sleep, older age was associated with a greater presence of airflow→EEG (t-test p< 0.05); higher BMI with more EEG→EMG, SpO2→EMG (both p < 0.01); higher AHI with more bidirectional SpO2↔ECG and EEG→SpO2 (all p < 0.001). During REM sleep, older age with more SpO2→airflow, EEG→ECG, and EOG→EEG (all p< 0.05); higher BMI with more airflow→SpO2, airflow→EMG, and ECG→EMG (all p< 0.001); higher AHI with more airflow↔SpO2 and SpO2→EMG (all p< 0.001). Conclusion This novel and pioneering application of CRL in extracting plausible physiological interactions from polysomnography signals, including cyclic relationships. The findings align with known interactions in apnea and show differences between NREM and REM. A current limitation is that we lack the description of whether a signal exerts a positive or inhibitory influence on another signal. Validation in interventional settings, such as CPAP titration, is needed to confirm causal mechanisms. Support (if any) AASM Foundation 370-SR-25.
Citation
Y. Sun, B. Sun, A. Mazumder, H. Dai, A. Mazumder, M.B. Westover , et al., "0379 Inferring Possible Causal Relationships among Polysomnography Signals using Multi-Modal Causal Representation Learning," Sleep, vol. 49, no. Supplement_1, pp. a168-a168, 2026, https://doi.org/10.1093/sleep/zsag091.0379.
Source
Sleep
Conference
Keywords
32 Biomedical and Clinical Sciences, 3201 Cardiovascular Medicine and Haematology, 3208 Medical Physiology
Subjects
Source
Publisher
Oxford University Press