Skip to main content
2024 BMES Annual Meeting Main banner
Icon Legend
This session is not in your schedule.
This session is in your schedule. Click again to remove it.
Back

Bioinformatics, Computational and Systems Biology

Bioinformatics, Computational and Systems Biology - Poster Session E

Poster L26 - Evaluating Inter-Beat Intervals for Sleep Staging: A Comparative Study with Electrocardiogram Data

Saturday, October 26, 2024
10:00 AM – 11:00 AM EST
Location: Exhibit Hall E, F & G

Introduction:  Sleep staging, a critical component of sleep studies, has traditionally relied on polysomnography, utilizing primarily electroencephalogram (EEG) and electrooculogram (EOG) signals. Recent advances have shown that sleep staging can be effectively achieved using only electrocardiogram (ECG) signals, highlighting their lower obtrusiveness for collection and suitability for continuous monitoring due to the location of the electrodes. Building on this, our study explores the feasibility of monitoring sleep stages using only cardiovascular measures, with the aim of translating ECG monitoring capabilities to wrist-worn wearable devices that collect continuous photoplethysmography (PPG) signals.

This research investigates the translation of sleep staging techniques from ECG to PPG via the calculation of inter-beat intervals (IBIs), a common metric derived from cardiovascular signals. Utilizing data from the first visit of the Sleep Heart Health Study (SHHS1), which includes 5786 records (N=5786 patients), we propose a model combining residual networks (ResNet) for feature extraction with temporal convolution and Long Short-Term Memory (LSTM) networks for 3-class sequence-to-sequence predictions (Wake, NREM, REM). We assessed the model’s performance on both ECG and IBI data. Our model achieved a Cohen’s Kappa of 0.788 and a Macro F1 Score of 0.880 on the ECG data. For the IBI data, the performance was comparable with a Cohen’s Kappa of 0.751 and a Macro F1 Score of 0.859. These results underscore the potential of IBI data in containing useful information for accurate sleep staging, serving as a potential intermediary for translating these findings from ECG-based sleep staging models to PPG-based wearable devices for sleep monitoring.

Materials and

Methods:  The SHHS1 dataset comprises 5786 ECG records (one record was excluded due to inconsistent sampling frequency) from individual participants monitored over a single night’s sleep study, with data sampled at 125 Hz. Each 30-second segment is categorized into one of six sleep stages by a sleep technician: Wake, Non-REM 1, Non-REM 2, Non-REM 3, REM, and Movement. We consolidated the Non-REM stages into a single label to simplify classification, and the Movement labels are encoded as -1. IBI were extracted from the ECG signals by identifying the time intervals between consecutive heartbeats (R-peaks). Both the IBI and the original ECG data were downsampled to 25 Hz to streamline the analysis and reduce dataset size. Although this downsampling may lead to the loss of some features, we hypothesized that it would retain sufficiently important waveform information to enable effective sleep staging.

Our model architecture uses a ResNet to extract 256 feature channels as an initial feature extraction step. This is complemented by a Temporal Convolutional Network (TCN) which processes the data though dilation layers, enhancing the model’s capability to capture long-range sleep stage transition patterns. Sequence prediction is further refined by 3 bi-directional LSTM layers with 8 attention heads. The dataset was divided into 70% for training, 10% for validation, and 20% for testing. Cohen’s Kappa, Macro and REM F1 Scores, and accuracy are used to assess model performance. Training is conducted over 200 epochs using a cross-entropy loss function and the Adam optimizer on a Nvidia RTX 4090 GPU.

Results, Conclusions, and Discussions: The comparative analysis of our model trained on either IBI or ECG data offers revelations about the practical utility of cardiovascular signals in sleep staging. Notably, the model trained with IBI data attained an accuracy of 0.861, a Macro F1 score of 0.859, Cohen’s Kappa of 0.751, alongside a REM F1 score of 0.776. In comparison, the ECG-trained model achieved an accuracy of 0.880, a Macro F1 of 0.880, Cohen’s Kappa of 0.788, and an REM F1 of 0.802.

While the ECG data provides richer information, enabling higher performance on all metrics due to its unfiltered waveform, the performance of the IBI-based model is compellingly close. This closeness is particularly significant given that IBI data is a more abstracted form of cardiac information and can similarly be extracted from PPG signals from wrist-worn devices and even other types of monitoring devices (e.g. acoustic- or radar-based sensing).

This study demonstrates that while ECG data inherently contains more detailed information for sleep staging, IBI data can achieve comparable performance metrics. This finding showcases the potential for the development of unobtrusive wearable sleep monitoring technologies utilizing easily obtainable cardiovascular signals like IBI. Our future efforts will aim to refine IBI signal quality and filter out artifacts to boost training stability. Additionally, we will employ interpretability techniques such as gradient ascent or Shapley Additive Explanations to identify which aspect of the ECG signals are most crucial for sleep staging. Such insights could facilitate the preservation and effective translation of these critical features into IBI data, enhancing the performance and utility of wearable devices in sleep studies.

Acknowledgements (Optional):