Holter-to-Sleep: AI-Enabled Repurposing of Single-Lead ECG for Sleep Phenotyping

Donglin Xie; Qingshuo Zhao; Jingyu Wang; Shijia Geng; Jiarui Jin; Jun Li; Rongrong Guo; Guangkun Nie

eess.SPMar 19, 2026

Holter-to-Sleep: AI-Enabled Repurposing of Single-Lead ECG for Sleep Phenotyping

Authors: Donglin Xie, Qingshuo Zhao, Jingyu Wang, Shijia Geng, Jiarui Jin, Jun Li, Rongrong Guo, Guangkun Nie,

Organizations: National Institute of Health Data Science, Peking University, Beijing, China · Institute of Medical Technology, Peking University Health Science Center, Beijing, China · Department of Computer Science, Tianjin University of Technology, Tianjin, China · Department of Respiratory and Critical Care Medicine, Binzhou Medical University Hospital, Binzhou, China ·

Abstract

Sleep disturbances are tightly linked to cardiovascular risk, yet polysomnography (PSG)-the clinical reference standard-remains resource-intensive and poorly suited for multi-night, home-based, and large-scale screening. Single-lead electrocardiography (ECG), already ubiquitous in Holter and patch-based devices, enables comfortable long-term acquisition and encodes sleep-relevant physiology through autonomic modulation and cardiorespiratory coupling. Here, we present a proof-of-concept Holter-to-Sleep framework that, using single-lead ECG as the sole input, jointly supports overnight sleep phenotyping and Holter-grade cardiac phenotyping within the same recording, and further provides an explicit analytic pathway for scalable cardio-sleep association studies. The framework is developed and validated on a pooled multi-center PSG sample of 10,439 studies spanning four public cohorts, with independent external evaluation to assess cross-cohort generalizability, and additional real-world feasibility assessment using overnight patch-ECG recordings via objective-subjective consistency analysis. This integrated design enables robust extraction of clinically meaningful overnight sleep phenotypes under heterogeneous populations and acquisition conditions, and facilitates systematic linkage between ECG-derived sleep metrics and arrhythmia-related Holter phenotypes. Collectively, the Holter-to-Sleep paradigm offers a practical foundation for low-burden, home-deployable, and scalable cardio-sleep monitoring and research beyond traditional PSG-centric workflows.

Figures & tables

Figure 1: Overview of the Holter-to-Sleep study design and analytic workflow. a, Data collection. Single-lead ECG was extracted from multi-center PSG cohorts. b, Model development. ECG-based deep learning model for sleep staging and sleep disturbance detection. c, Wearable data validation. Participants underwent overnight recordings using a patch-based ECG device, accompanied by PSQI-based subjective sleep quality assessment; model-derived sleep-consistency metrics were used to evaluate real-world agreement. d, Holter-derived cardiac metrics and sleep associations. Model outputs were used to align long-duration nocturnal ECG with sleep, enabling joint extraction of sleep metrics and Holter-grade cardiac phenotypes, followed by association analyses to support scalable cardio–sleep phenotyping.

Metric	Training ( $n=6{,}811$ )	Train-val ( $n=1{,}460$ )	Internal val ( $n=1{,}459$ )	External val ( $n=709$ )
Demographics
Age (years)	68.48 $\pm$ 10.89	68.38 $\pm$ 11.20	68.72 $\pm$ 10.56	41.34 $\pm$ 19.45
Female, $n$ (%)	2,354 (34.56%)	496 (33.97%)	495 (33.93%)	388 (54.72%)
BMI (kg m -2 )	27.94 $\pm$ 4.84	28.04 $\pm$ 4.87	28.02 $\pm$ 4.91	32.12 $\pm$ 9.26
Event burden and oxygen desaturation
AHI (events h -1 )	20.21 $\pm$ 17.02	20.39 $\pm$ 17.41	20.22 $\pm$ 16.98	12.30 $\pm$ 16.75

Table 1: Baseline characteristics and PSG-derived sleep metrics across data splits. Continuous variables are reported as mean

\pm

s.d.; categorical variables are reported as

n

(%). AHI, apnea-hypopnea index; ArI, arousal index; ODI, oxygen desaturation index; TST, total sleep time; TIB, time in bed; WASO, wake after sleep onset.

Class setting	Accuracy $\uparrow$	Weighted F1 $\uparrow$	Kappa $\uparrow$	Macro AUC $\uparrow$
Internal validation
5-class	0.773 [0.772, 0.773]	0.761 [0.760, 0.762]	0.670 [0.668, 0.670]	0.930 [0.930, 0.930]
4-class	0.804 [0.803, 0.804]	0.801 [0.801, 0.802]	0.701 [0.700, 0.702]	0.944 [0.944, 0.944]
3-class	0.870 [0.869, 0.870]	0.869 [0.869, 0.870]	0.776 [0.775, 0.777]	0.962 [0.962, 0.962]
2-class	0.926 [0.926, 0.926]	0.926 [0.926, 0.926]	0.834 [0.834, 0.835]	0.974 [0.974, 0.975]
External validation (CFS)

Table 2: Cross-cohort performance for ECG-based sleep staging at multiple granularities. Values are reported as point estimates with 95% CIs in brackets. For 4-class staging, Light merges N1+N2 and Deep corresponds to N3; for 3-class staging, NREM merges N1–N3; for 2-class staging, Sleep merges N1–N3 and REM.

Figure 2: Cross-cohort generalization of single-lead ECG sleep staging on internal and external validation. The top row shows internal validation and the bottom row shows external validation; from left to right are confusion matrices for 5-class, 4-class, 3-class, and 2-class staging. Each cell reports the epoch count, with the percentage in parentheses normalized by the predicted class. For 4-class staging, Light merges N1+N2 and Deep corresponds to N3; for 3-class staging, NREM merges N1–N3; for 2-class staging, Sleep merges N1–N3 and REM.

Metric	Arousal	Respiratory
Internal validation
Accuracy $\uparrow$	0.852 [0.852, 0.853]	0.798 [0.798, 0.799]
Precision $\uparrow$	0.518 [0.517, 0.520]	0.368 [0.366, 0.369]
Recall $\uparrow$	0.593 [0.591, 0.595]	0.534 [0.532, 0.536]
Specificity $\uparrow$	0.899 [0.899, 0.900]	0.843 [0.843, 0.844]
F1-score $\uparrow$	0.553 [0.552, 0.555]	0.435 [0.434, 0.437]

Table 3: Cross-cohort performance for ECG-based sleep event detection. Values are reported as point estimates with 95% CIs in brackets. AUC, area under the receiver operating characteristic curve.

Figure 3: Cross-cohort generalization of sleep event detection. ROC curves are shown for arousal and respiratory event detection; the left panel corresponds to internal validation and the right panel to external validation. The area under the ROC curve (AUC) is reported in each panel.

Channel	Method	Accuracy $\uparrow$	Weighted F1 $\uparrow$	$\kappa$ $\uparrow$	Macro AUC $\uparrow$	Relative $\kappa$ increase
C4+LOC+EMG1	Net1D	0.843	0.842	0.782	0.962	–
C4+LOC+EMG1	Net1D+Trans	0.886	0.881	0.838	0.977	+7.16%
EEG-C4	Net1D	0.827	0.822	0.759	0.955	–
EEG-C4	Net1D+Trans	0.873	0.869	0.820	0.974	+8.04%
EOG-L	Net1D	0.832	0.829	0.767	0.956	–
EOG-L	Net1D+Trans	0.878	0.875	0.826	0.975	+7.69%

Table 4: Channel-wise comparison on the CFS cohort highlights ECG as a competitive modality for scalable sleep staging. Performance is reported for 5-class sleep staging using individual PSG channels. Net1D denotes the CNN-only baseline, and CNN+Trans denotes the proposed two-stage model with temporal refinement. Relative

\kappa

increase is computed as

(\kappa_{\mathrm{CNN+Trans}}-\kappa_{\mathrm{CNN}})/\kappa_{\mathrm{CNN}}\times 100\%

Figure 4: Real-world wearable validation of ECG-derived sleep consistency metrics against PSQI. Scatter plots show associations between normalized PSQI scores and model-derived overnight metrics. Left: wake probability, defined as the overnight mean of epoch-wise predicted wake probabilities. Right: sleep efficiency estimated from model-inferred sleep–wake information. Solid lines indicate linear regression with 95% confidence bands. Pearson’s correlation coefficient ( $r$ ) and two-sided $p$ values are shown in each panel.

Figure 5: ECG-derived sleep phenotypes differ across Holter arrhythmia-related strata. Violin plots show distributions of time in bed (TIB), sleep latency (SL), sleep efficiency (SE), and oxygen desaturation index at 3% (ODI 3%). Participants are stratified by (i) frequent premature ventricular contractions (PVC burden $\geq 21$ events h -1 ) versus non-frequent PVC, (ii) frequent premature atrial contractions (PAC burden $\geq 30$ events h -1 ) versus non-frequent PAC, (iii) atrial fibrillation (AF) presence during sleep versus absence, and (iv) ventricular tachycardia (VT) presence during sleep versus absence. Embedded boxplots indicate the median and interquartile range. Values below each panel denote mean $\pm$ s.d. $P$ values from between-group comparisons are annotated in each panel. TIB and SL are reported in minutes, SE in %, and ODI 3% in events h -1 .

Figure 6: An integrated web dashboard for ECG-based sleep and Holter cardiac reporting. The dashboard supports raw EDF input, automatically extracts the ECG channel, and integrates the proposed sleep inference pipeline with Holter-grade cardiac analytics. The interface provides (top) an overnight heart-rate trend, (middle) a sleep-stage timeline, and (bottom) summary sleep metrics and cardiac metrics (including HRV and arrhythmia-related summaries) in a consolidated report.

Figure 7: Study dataset construction and screening flow. Public PSG cohorts included CFS (ECG2, $N=735$ ), MESA (EKG, $N=2056$ ), SHHS (ECG, $N=5793$ ), and MrOS (ECGR, $N=3924$ ). Exclusion 1: recordings with total monitoring duration $<7$ h or $>14$ h; remaining samples were CFS $N=720$ , MESA $N=2028$ , SHHS $N=4346$ , and MrOS $N=3528$ . Exclusion 2: recordings with abnormal sleep-stage distributions; remaining samples were CFS $N=709$ , MESA $N=2007$ , SHHS $N=4228$ , and MrOS $N=3495$ . CFS was used as an independent external validation set ( $N=709$ ). MESA, SHHS, and MrOS were pooled and split into a training set ( $N=6811$ ), a training-validation set ( $N=1460$ ), and an internal validation set ( $N=1459$ ). For real-world wearable ECG, $N=70$ recordings were collected; Exclusion 3: failure to complete the PSQI questionnaire; the final real-world validation subset comprised $N=26$ .

Figure 8: Two-stage single-lead ECG framework for sleep staging and event detection. Stage 1 takes 30-s single-lead ECG epochs as input. After preprocessing, a 1D CNN (Net1D) learns deep representations. Frequency/time–frequency features are also extracted and embedded, and then fused with CNN features to produce a shared representation for five-class sleep staging and multi-task sleep event detection, yielding initial predictions. Stage 2 constructs a multi-epoch window from neighboring epochs and feeds the sequence of Stage 1 features into a Transformer. The Stage 1 feature extractor is frozen, and the Transformer performs temporal refinement by predicting the center epoch, producing refined sleep staging and event detection outputs.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Study	Classes	Dataset(s)	$N$	$\kappa$	Acc
Wei et al. (2018) 46	3	MIT–BIH Polysomnographic Database 11	18	0.560	0.770
Urtnasan et al. (2022) 38	3	Private dataset	112	–	0.860
Li et al. (2018) 24	4	MIT–BIH Polysomnographic Database 11	18	0.540	0.756
Wei et al. (2019) 47	4	Xijing Hospital dataset (FMMU)	238	0.550	0.778
Radha et al. (2019) 30	4	Siesta database 17	588	0.610	0.770
Sun et al. (2020) 35	5	MGH Cohort	8682	0.490	–

Appendix

Table 14: Representative single-lead ECG-based sleep-staging studies and their self-reported metrics. Reported values are taken from the original publications and correspond to each study’s internal test-set evaluation under the label set indicated (3-/4-/5-class).

N

denotes the number of PSG recordings unless otherwise specified. Because cohorts, label definitions, and evaluation protocols differ across studies, this table is provided for qualitative context rather than head-to-head benchmarking. Missing entries indicate that the corresponding metric was not reported in the original publication.

Sleep metrics (16)	Holter-grade cardiac metrics (32)
AHI (events h -1 ) ArI (events h -1 ) N1 / TST (%) N2 / TST (%) N3 / TST (%) NREM / TST (%) REM / TST (%) SE (%) SL (min) TIB (min) TST (min) TTSP (min) WASO (min) Wake / TIB (%) ODI 3% (events h -1 ) ODI 4% (events h -1 )	Total_valid_beats Avg_HR Min_HR Max_HR Snt_max_beat_count Snt_duration Snb_max_beat_count Snb_duration Total_PAC Single_PAC Paired_PAC Bigeminy_PAC Triad_PAC Total_PVC Single_PVC Paired_PVC Bigeminy_PVC Triad_PVC Total_VT Total_SVT SDNN SDANN SDANNIndex RMSSD pNN50 HRV_tri LF HF LFNU HFNU LF/HF AF_duration

Appendix

Table 15: Summary of sleep-related metrics and Holter-grade cardiac metrics used in this study.

Frequency Domain Features	Time-Frequency Features
1. Delta Power	Energy Features : Energy calculated for each wavelet decomposition level (5 levels, first-level coefficients for each level)
2. Theta Power	Statistical Features : Statistical features computed for each wavelet decomposition level (5 $\times$ 5 matrix, 5 features per decomposition level), including:
3. Alpha Power	1. Mean
4. Total Power	2. Standard Deviation
5. Peak Frequency	3. Variance
6. Spectral Centroid	4. Maximum Value

Appendix

Table 16: Handcrafted spectral and time–frequency descriptors (39 features) used for feature-level fusion.

Wearable Health Monitoring

Wearable health monitoring uses body-worn sensors and AI to track physiological and behavioral signals for continuous, personalized health assessment.

See all 109 papers

ECG-SDI: Single-lead Electrocardiogram Enables Continuous Sleep Depth Index Profiling and Scalable Digital Biomarkers for Sleep Health
Oct 16, 2026Donglin Xie, Qingshuo Zhao, Songchi Zhou +5Sleep Stage Classification Electrocardiography
Institute of Medical Technology, Peking University Health Science Center, Beijing, China · National Institute of Health Data Science, Peking University, Beijing, China · Department of Computer Science, Tianjin University of Technology, Tianjin, China +6
Comparative Analysis of State-of-the-Art Foundation Models for Sleep Analysis Under Channel Reduction
Aug 14, 2026Hassan Mehdi, Riku Klen, Ayse Kosal Bulbul +5Sleep Stage Classification EEG Foundation Models
Institute of Biomedicine, University of Turku, Finland · Turku PET Centre, University of Turku and Turku University Hospital, Finland · University College Cork, Ireland +3
AnySleep: a channel-agnostic deep learning system for high-resolution sleep staging in multi-center cohorts
Dec 16, 2025Niklas Grieger, Jannik Raskob, Siamak Mehrkanoon +1Electroencephalography Sleep Stage Classification
Department of Medical Engineering and Technomathematics, FH Aachen University of Applied Sciences, 52428 Jülich, Germany · Department of Information and Computing Sciences, Utrecht University, Utrecht, The Netherlands · Institute for Data-Driven Technologies, FH Aachen University of Applied Sciences, 52428 Jülich, Germany

Oct 16, 2026J/K move · Enter open · S save

ECG-SDI: Single-lead Electrocardiogram Enables Continuous Sleep Depth Index Profiling and Scalable Digital Biomarkers for Sleep Health

Donglin Xie, Qingshuo Zhao, Songchi Zhou, Haoyu Wang+4

Institute of Medical Technology, Peking University Health Science Center, Beijing, China · National Institute of Health Data Science, Peking University, Beijing, China · Department of Computer Science, Tianjin University of Technology, Tianjin, China+6

Sleep Stage Classification Electrocardiography

Open paper AskPDF arXiv

Conventional sleep staging provides a coarse-grained representation that obscures within-stage depth variation. We developed ECG-SDI, a single-lead electrocardiography (ECG) framework for continuous sleep-depth profiling and digital biomarker extraction. Using 12,392 overnight polysomnographic recordings, we trained and validated a two-stage deep learning model supervised by conventional sleep-stage labels. Correlations with ordered stages excluding rapid eye movement (REM) sleep were 0.828 in the pooled internal test set and 0.848 in the external Cleveland Family Study; corresponding REM classification AUROCs were 0.954 and 0.977. Overnight trajectories captured within-stage variation. Biomarker-based phenotyping identified a disturbed subtype associated with poorer subjective sleep quality and higher all-cause mortality in the Sleep Heart Health Study (adjusted hazard ratio, 1.20; 95% CI, 1.05-1.38). Among 507 participants with paired recordings before and after continuous positive airway pressure (CPAP), normalized sleep-depth area increased by 6.1% and shallow-sleep burden below 0.3 decreased by 16.0%. Exploratory analyses linked reductions in approximate entropy and depth variability to reduced oxygen desaturation and increased sleep efficiency, respectively (Spearman rho = 0.24 and 0.12); both associations remained positive after covariate adjustment. Mean ECG-SDI within N2 increased by 0.017 (95% CI, 0.004-0.030). A wearable proof-of-concept illustrated the acquisition-to-report workflow. These findings support ECG-based sleep phenotyping and further evaluation of CPAP response monitoring at home.

Abstract

Wearable Health Monitoring

Explore similar work

ECG-SDI: Single-lead Electrocardiogram Enables Continuous Sleep Depth Index Profiling and Scalable Digital Biomarkers for Sleep Health

Comparative Analysis of State-of-the-Art Foundation Models for Sleep Analysis Under Channel Reduction

AnySleep: a channel-agnostic deep learning system for high-resolution sleep staging in multi-center cohorts