Story
Holter-to-Sleep: AI-Enabled Repurposing of Single-Lead ECG for Sleep Phenotyping
Key takeaway
Researchers developed an AI method to analyze ECG data and track sleep patterns without expensive sleep tests. This could make it easier to screen for sleep issues that are linked to heart health.
Quick Explainer
The core idea of this study is to repurpose widely available single-lead electrocardiography (ECG) data to jointly characterize a person's sleep architecture and detect clinically relevant sleep disturbances. The approach uses a two-stage deep learning model: first, a CNN-based component learns meaningful ECG representations, which are then refined by a Transformer-based module to capture long-range sleep dynamics. This unified multi-task model can simultaneously predict sleep stages and detect events like arousals and respiratory events, enabling scalable cardio-sleep phenotyping without the burden of dedicated sleep monitoring equipment. The key novelty is leveraging widely accessible ECG data to achieve performance competitive with more specialized sleep-related modalities.
Deep Dive
Holter-to-Sleep: AI-Enabled Repurposing of Single-Lead ECG for Sleep Phenotyping
Overview
This study presents a Holter-to-Sleep framework that uses single-lead ECG as the sole input to jointly characterize sleep architecture and clinically relevant sleep disturbances at scale, with rigorous multi-cohort validation and real-world wearable assessment. The framework also enables Holter-level cardio–sleep association analyses.
Problem & Context
- Sleep disturbances are tightly linked to cardiovascular risk, but polysomnography (PSG) remains resource-intensive and poorly suited for multi-night, home-based, and large-scale screening.
- Single-lead electrocardiography (ECG) is widely available from Holter and patch-based devices, supports comfortable long-term recordings at home, and enables continuous, low-burden monitoring across multiple nights.
- Leveraging ECG for sleep monitoring can enable concurrent sleep and cardiac phenotyping, facilitating scalable investigation of cardio–sleep associations.
Methodology
Model Architecture
- Two-Stage deep learning architecture: CNN-based morphological representation learning combined with Transformer-based temporal refinement to capture long-range sleep dynamics and improve cross-cohort generalization.
- Unified multi-task learning: Simultaneous modeling of sleep staging and detection of key sleep disturbances (arousals, respiratory events).
Evaluation
- Multi-cohort training/internal validation (MESA, MrOS, SHHS) and independent external validation (CFS).
- Real-world wearable validation using patch-based ECG with objective–subjective consistency assessment.
Cardio–Sleep Association Analysis
- Linked ECG-derived sleep phenotypes with Holter-grade cardiac phenotypes (arrhythmia burden, heart rate variability) to enable scalable analyses.
Data & Experimental Setup
- Total of 10,439 PSG studies from four public cohorts (MESA, MrOS, SHHS, CFS).
- Extracted single-lead ECG signals and corresponding sleep staging, arousal, and respiratory event labels.
- Additional 26 overnight wearable ECG recordings with subjective sleep quality (PSQI) data.
Results
Sleep Staging & Event Detection Performance
- Consistent performance with strong cross-cohort generalization:
- 4-class sleep staging: κ = 0.701 (internal), κ = 0.744 (external)
- Arousal detection: AUC = 0.855 (internal), 0.864 (external)
- Respiratory event detection: AUC = 0.787 (internal), 0.785 (external)
- ECG emerged as a competitive modality for sleep staging, approaching performance of more direct sleep-related channels.
Real-World Wearable Validation
- Model-derived metrics (wake probability, sleep efficiency) showed significant correlations with subjective PSQI scores.
Cardio–Sleep Associations
- Participants with arrhythmia phenotypes or higher ectopy burden exhibited poorer sleep, characterized by longer time in bed, longer sleep latency, lower sleep efficiency, and higher oxygen desaturation.
Interpretation
- The Holter-to-Sleep framework provides a scalable paradigm that integrates nocturnal sleep assessment and cardiac risk characterization within the same data source and workflow.
- Leveraging single-lead ECG enables competitive sleep staging without the additional acquisition burden of dedicated neurophysiological electrodes, supporting longitudinal follow-up and population-scale cardio–sleep phenotyping.
- Incorporating sleep phenotypes alongside Holter-level cardiac metrics can reveal clinically meaningful cardio–sleep interactions and risk signals that would be missed by conventional Holter analyses alone.
Limitations & Uncertainties
- The wearable validation relied on PSQI as a subjective reference, whereas the model-derived metrics reflected a single night.
- Cohort-specific training showed reduced performance, potentially due to more complex ECG morphologies and rhythms in high-risk populations.
- Further work is needed to improve robustness against arrhythmias and noise, and to incorporate multi-night wearable recordings with sleep diaries.
What Comes Next
- Explore comorbidity-aware stratification, domain adaptation, and explicit robustness mechanisms to improve performance in high-risk populations.
- Expand real-world validation with multi-night wearable recordings, sleep diaries, and additional modalities to strengthen ecological validity.
- Further develop the integrated web dashboard to facilitate practical adoption by researchers and clinicians.
