PaperlayerView exact revision on arXiv

Source-linked AI summary

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, Jingyu Wang, Yuxi Zhou, Thomas Penzel, Shenda Hong

arXiv:2610.23727v1eess.SPcs.LG

TL;DR

Conventional staging provides a coarse-grained account of sleep and does not capture progressive depth variation within stages. This paper develops ECG-SDI, a single-lead ECG framework for continuous sleep-depth profiling and biomarker extraction. It reports that the approach preserves whole-night organization while revealing within-stage heterogeneity, and that derived phenotypes are associated with adverse sleep-health profiles and mortality.

  • Problem

    Fixed 30-second sleep-stage epochs do not adequately capture progressive within-stage depth variation, while EEG- or PSG-dependent continuous phenotyping remains difficult to scale to home settings.

    Source: Introduction · p. 2

  • Method

    ECG-SDI uses a two-stage single-lead ECG model trained with conventional sleep-stage labels to generate a continuous sleep-depth value for each 30-second epoch.

    Source: Introduction · p. 3, Data curation and deep learning model… · p. 3

  • Results

    ECG-SDI preserves whole-night sleep organization while revealing within-stage heterogeneity, and its biomarker clustering identifies phenotypes associated with adverse sleep-health profiles and increased all-cause mortality.

    Source: Introduction · p. 3

  • Takeaways & Limitations

    The authors conclude ECG-SDI may support treatment-response monitoring and longitudinal home-based follow-up.

    Source: Introduction · p. 3

Abstract

from arXiv · show

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.

Introduction

Conventional staging misses progressive sleep-depth variation within fixed 30-second epochs, while continuous depth and digital biomarkers offer a more detailed framework. ECG-SDI applies single-lead ECG to continuous sleep-depth profiling, phenotype discovery, and treatment-response assessment.

Introduction

  • Fixed 30-second staging does not adequately capture progressive within-stage depth variation or directly quantify sleep fragmentation and recovery dynamics.If continuous phenotyping still depends on EEG or full PSG, its scalability in lightweight home and real-world settings remains limited.Source: Introduction · p. 2
  • The authors propose ECG-SDI, a scalable single-lead ECG framework for continuous sleep-depth profiling, validated in public sleep cohorts and an external clinical CPAP cohort.Source: Introduction · p. 3
  • ECG-SDI preserves whole-night sleep organization while revealing physiological heterogeneity within conventional stage labels.This continuous view captures variation that conventional labels alone do not show.Source: Introduction · p. 3
  • Clustering overnight ECG-SDI biomarkers identifies sleep phenotypes associated with adverse sleep-health profiles and increased all-cause mortality.Source: Introduction · p. 3
  • ECG-SDI biomarkers were sensitive to CPAP-related sleep improvement, supporting longitudinal follow-up and home-based monitoring.The framework is presented as a treatment-responsive digital biomarker, not only a static sleep descriptor.Source: Introduction · p. 3

RESULTS

ECG-SDI was evaluated across internal and independent external cohorts, where it recovered ordered sleep-depth structure and REM classification while representing variation within conventional stages. Biomarker phenotyping and paired CPAP analyses linked its overnight profiles to clinical differences and treatment-associated changes; a wearable workflow was demonstrated as a proof of concept.

Data curation and deep learning model development

  • The analysis included 12,392 overnight PSG recordings; 10,652 from SHHS, MESA, and MrOS were split by participant for model development, while CFS and BZ-CPAP supplied independent external recordings.A frozen pretrained backbone and task-specific heads generated an ECG-SDI value per 30-second epoch using conventional stage labels and neighboring-epoch context.Source: Data curation and deep learning model… · p. 3

Single-lead ECG robustly recovers continuous sleep depth and REM structure across internal and external cohorts

  • Sleep-depth correlations excluding REM were 0.828 in the pooled internal test set and 0.848 in external CFS, with REM classification AUROCs of 0.954 and 0.977, respectively.In BZ-CPAP, corresponding correlations were 0.456 before and 0.546 after CPAP, with REM AUROCs of 0.803 and 0.830.Source: Single-lead ECG robustly recovers… · pp. 3, 5, Table 1 · p. 5

Whole-night ECG-SDI trajectories reveal continuous variation within conventional sleep stages

  • ECG-SDI produced continuous trajectories that fluctuated within fixed conventional stage labels, capturing variation not represented by staging alone.Illustrated same-stage N1 and N2 epochs showed differing ECG-SDI values and heart rates; heart rate did not vary monotonically with ECG-SDI across all illustrated stages.Source: Whole-night ECG-SDI trajectories… · pp. 5–6, Figure 2 · p. 6

ECG-SDI-derived sleep phenotypes show distinct clinical profiles and an association with all-cause mortality

  • Gaussian-mixture clustering identified normal and disturbed sleep phenotypes across SHHS, CFS, and MrOS, with cohort-specific biomarker profiles and consistently lower sleep efficiency in the disturbed subtype.In SHHS, disturbed-subtype shallow-sleep burden and variability were higher, while cumulative sleep depth was lower.Source: ECG-SDI-derived sleep phenotypes show… · p. 6, Figure 3 · p. 8
  • The disturbed sleep subtype was associated with higher all-cause mortality in SHHS after full adjustment (hazard ratio, 1.20; 95% CI, 1.05–1.38).The two biomarker-derived phenotypes also differed consistently in subjective sleep quality across SHHS, CFS, and MrOS.Source: ECG-SDI-derived sleep phenotypes show… · p. 7, Figure 3 · p. 8

ECG-SDI biomarkers are treatment-responsive and track CPAP-associated restoration of sleep continuity

  • Among 507 paired BZ-CPAP recordings, normalized sleep-depth area increased 6.1% and shallow-sleep burden below 0.3 decreased 16.0% after treatment.RB decreased 11.4%, CV decreased 4.9%, and REM proportion increased 44.4%; MDR change was not statistically significant.Source: ECG-SDI biomarkers are… · p. 7, Figure 4 · p. 9

Exploratory analyses link CPAP-associated ECG-SDI changes to PSG improvement and within-stage variation

  • Greater APPe reduction was associated with greater ODI3 reduction (Spearman ρ = 0.24), while greater CV reduction was associated with increased sleep efficiency (ρ = 0.12).Both associations remained positive after covariate adjustment in models with 395 paired participants; mean N2 ECG-SDI increased by 0.017 after CPAP among participants with at least 10 minutes of N2 at each visit.Source: Exploratory analyses link… · pp. 7, 9, Figure 5 · p. 10

A wearable single-lead ECG workflow enables proof-of-concept home sleep-depth reporting

DISCUSSION

ECG-SDI recovered ordered sleep-stage and REM structure without EEG at inference while revealing variation within conventional stages. Its derived phenotypes tracked sleep-health differences, and paired CPAP analyses showed treatment-associated changes; the wearable workflow remains a proof of concept.

Translational scope

  • A wearable acquisition-to-report proof of concept accompanied the cross-sectional phenotyping and treatment-associated analyses.The study presents this workflow as a basis for developing repeated, lower-burden monitoring.Source: DISCUSSION · p. 11

Performance and single-signal scope

  • 0.828 and 0.848 were the correlations with ordered sleep stages excluding REM in the pooled internal test set and external CFS, respectively; REM AUROCs were 0.954 and 0.977.These results show that substantial sleep-related structure can be recovered without EEG input at inference.Source: DISCUSSION · p. 11

Continuous trajectories

  • Overnight trajectories retained recognizable sleep architecture while showing variation within stage categories; higher ECG-SDI accompanied lower heart rates in selected N1 and N2 examples, but not monotonically in REM examples.The examples support within-stage heterogeneity, not interpreting ECG-SDI as a simple inverse heart-rate scale.Source: DISCUSSION · p. 12

Clinical phenotypes

  • The disturbed subtype had poorer subjective sleep quality across SHHS, CFS, and MrOS, and its SHHS all-cause mortality association persisted after adjustment (hazard ratio, 1.20; 95% confidence interval, 1.05–1.38).The mortality model included AHI and total sleep time; cohort-specific biomarker patterns support interpreting the subtypes as multidimensional profiles.Source: DISCUSSION · p. 12

Treatment response

  • In paired CPAP recordings, cumulative sleep depth and REM proportion increased while shallow-sleep burden and overnight variability decreased, forming a pattern consistent with improved sleep continuity.The paired design extends the evidence from between-person differences to within-person changes relevant to candidate treatment monitoring.Source: DISCUSSION · p. 12

Exploratory treatment analyses

  • Spearman ρ was 0.24 for APPe reduction versus ODI3 reduction and 0.12 for CV reduction versus increased sleep efficiency; mean ECG-SDI within N2 increased by 0.017.Both associations remained positive after covariate adjustment; their modest magnitudes do not imply interchangeability.Source: DISCUSSION · p. 12

METHODS

The methods develop ECG-SDI from single-lead ECG and PSG-derived labels, then evaluate stage structure, phenotypes, CPAP-associated changes, and a wearable reporting workflow.

Study design, cohorts, and split strategy

  • The retrospective multi-cohort study included 12,392 PSG recordings; 10,652 public-cohort recordings were split by participant, while CFS and BZ-CPAP recordings were held out for external evaluation.The 507 BZ-CPAP participants contributed paired pre- and post-treatment recordings.Source: Study design, cohorts, and split… · p. 13

ECG preprocessing and PSG-derived labels

  • ECG was resampled, filtered, standardized, and segmented into 30-second epochs aligned with PSG labels; legacy stages 3 and 4 were combined as N3.Thirty-nine time- and frequency-domain features were computed per epoch, and discrete stage labels supervised both REM classification and depth ranking.Source: ECG preprocessing and PSG-derived labels · p. 13

ECG-SDI framework and model development

  • A two-stage model combined convolutional ECG representations and handcrafted features with Transformer temporal context, then trained separate REM-classification and continuous-depth heads.The depth head produced a scalar normalized within each recording using reference Wake/NREM epochs and clipped to [0, 1].Source: ECG-SDI framework and model development · p. 13

ECG-SDI-derived biomarkers and phenotype discovery

  • Overnight trajectories were summarized with distributional, temporal-complexity, and REM-related biomarkers; eight standardized biomarkers were clustered separately in SHHS, CFS, and MrOS.The cluster with higher mean AHI was designated the disturbed subtype.Source: ECG-SDI-derived biomarkers and… · p. 14

Clinical definitions and statistical analysis

  • Clinical analyses defined sleep apnea as AHI ≥5, assessed ordered-stage correlation and REM AUROC, and adjusted subtype and mortality models using prespecified covariates.Mortality was assessed with Cox models, including a model additionally adjusted for AHI and total sleep time.Source: Clinical definitions and statistical… · p. 14

CPAP treatment-response and exploratory analyses

  • CPAP response was assessed in paired recordings, distinguishing relative from absolute changes; exploratory analyses tested biomarker–PSG associations and mean ECG-SDI change within PSG-defined N2.The association models included 395 paired participants per model; the within-N2 analysis required at least 10 minutes of N2 at both visits.Source: CPAP treatment-response and… · pp. 14–15

Wearable proof-of-concept implementation

  • The wearable proof of concept linked chest-worn single-lead ECG to smartphone analysis and a report showing an overnight trajectory and AP, PR, RB, and CV summaries.Displayed values were illustrative; home accuracy and individual treatment-benefit classification were not evaluated.Source: Wearable proof-of-concept implementation · p. 15

Appendix A Baseline characteristics of the study cohorts

The appendix table describes baseline characteristics across study cohorts and clarifies how values, including BZ-CPAP AHI, are reported.

Cohort characteristics

  • The cohort characteristics include sample size, age, female proportion, BMI, AHI, and acquisition timeframe.Continuous variables are reported as mean ± standard deviation except female proportion.Source: Appendix A · p. 22, Table 2 · p. 22

BZ-CPAP reporting

Appendix B Box plot of stage-wise distribution of ECG-SDI

The appendix figure caption describes ECG-SDI distributions across conventional sleep stages and four cohorts.

Stage-wise distribution

  • The boxplots show ECG-SDI distributions for wake, N1, N2, N3, and REM in SHHS, CFS, MESA, and MrOS.Source: Figure 7 · p. 23

Boxplot guide

  • Compare stage-wise medians and spread: the center line marks the median, boxes the interquartile range, and whiskers the distribution spread.Source: Figure 7 · p. 23

Appendix C Detailed paired comparison of ECG-SDI-derived biomarkers before and after CPAP treatment

The paired comparison summarizes ECG-SDI biomarkers before and after CPAP in 507 PSG pairs, with most measures shifting in directions interpreted as greater sleep-depth burden and reduced shallow-sleep dominance. MDR showed no statistically significant pre–post change.

Appendix C Detailed paired comparison of ECG-SDI-derived biomarkers before and after CPAP treatment

  • After treatment, AP, DETRf, and PR increased, while RB, RB0.3, CV, SK, and APPe decreased.The table reports paired means, changes, significance levels, and test results.Source: Appendix C · p. 25
  • These shifts indicate greater cumulative sleep-depth burden, reduced shallow-sleep dominance, attenuated overnight variability, and changes in complexity-related features.Source: Appendix C · p. 25
  • MDR showed no statistically significant pre–post change.Source: Appendix C · p. 25

Appendix D Detailed architecture of the ECG-SDI framework

The ECG-SDI architecture adapts a frozen, pretrained two-stage ECG sleep-staging framework with task-specific outputs for REM classification and continuous sleep-depth estimation. The accompanying appendices describe its biomarkers, preprocessing, handcrafted features, and illustrative examples.

Appendix D Detailed architecture of the ECG-SDI framework

  • ECG-SDI replaces the staging head of a frozen two-stage ECG backbone with separate REM-classification and continuous sleep-depth heads.Stage 1 forms epoch embeddings from a CNN and handcrafted features; Stage 2 uses a Transformer to model inter-epoch context.Source: Figure 8 · p. 26

Appendix E Brief description of ECG-SDI-derived digital biomarkers

  • The biomarker descriptions prioritize intuitive physiological interpretations over formal mathematical definitions; RB0.3 and SampEn are supplementary descriptors in paired CPAP analyses.Source: Appendix E · p. 27, Table 5 · p. 27

Appendix F ECG preprocessing, handcrafted features, and illustrative cases

  • Preprocessing maps ECG to normalized amplitude rather than calibrated voltage, segments signals into 30-s epochs, and excludes invalid labels and specified recording boundaries.The retained sequence is evaluated with centered 15-epoch contexts, so its first and last seven epochs receive no predictions.Source: Appendix F · p. 28
  • Each epoch contributes 39 handcrafted features: nine spectral descriptors and 30 summaries of four-level Daubechies-4 wavelet coefficients.Source: Appendix F · p. 28
  • Whole-night visualizations use unsmoothed normalized SDI values, while selected same-stage epochs are arranged by increasing SDI for descriptive comparison.Heart-rate estimates are derived from displayed waveforms, and the examples are not intended for population-level inference.Source: Appendix F · p. 28

Appendix G Reproducible model specification and evaluation procedures

The released specification combines waveform and handcrafted-feature representations for each epoch, then models temporal context with Transformer layers. Its centered input windows and recording-level sequence attention operate at different temporal scales.

G.1 Epoch representations and temporal context

  • A 1024-dimensional waveform representation is concatenated with a 128-dimensional handcrafted-feature embedding to form 1152 features per epoch.The handcrafted branch uses 39 spectral and wavelet descriptors.Source: §G.1 · p. 29
  • Each target receives a centered 15-epoch window projected to 512 dimensions, then processed by three Transformer encoder layers with eight attention heads each.Source: §G.1 · p. 29
  • Transformer attention spans the recording-level sequence beyond each local window, while the first and last seven retained epochs lack centered-window predictions.Source: §G.1 · p. 29

G.2 Training stages and task-specific outputs

Training proceeds from five-class CNN sleep staging to a Transformer trained on fixed representations, followed by ECG-SDI adaptation with frozen temporal features. The adapted branches predict REM status and an unrestricted depth score using a combined classification and ranking objective.

G.2 Training stages and task-specific outputs

  • The CNN is trained for five-class sleep staging, its fixed representations train the staging Transformer, and adaptation freezes the temporal backbone while optimizing separate REM and depth branches.The REM branch produces two logits; the depth branch produces one unrestricted real-valued score.Source: §G.2 · p. 29
  • The ranking objective orders raw depth scores across differing Wake and NREM stage labels, with REM-containing pairs excluded from its numerator.Specified Wake–NREM margins range from 0.5 to 3.0, and the ranking loss is zero when fewer than two valid epochs are available.Source: §G.2 · p. 29
  • The joint objective combines REM cross-entropy with a ranking loss weighted by 0.5; backbone parameters remain frozen during adaptation.Transformer dropout remains active during training, and no padding attention mask is supplied.Source: §G.2 · p. 30
  • The released CNN training scripts use ReduceLROnPlateau with factor 0.3 and patience 10, and select checkpoints by minimum validation loss.The source-script Transformer batch-size default is 64, whereas the example JSON configuration sets it to 1.Source: §G.2 · p. 30

G.3 Depth normalization and performance evaluation

This section specifies retrospective depth normalization and evaluation, then defines the whole-night biomarkers used for phenotyping. It also describes how REM-related summaries combine depth with PSG reference labels.

G.3 Depth normalization and performance evaluation

  • Per-recording ECG-SDI is normalized using the fifth and 95th percentiles of raw scores among reference wake/NREM epochs; the evaluation is not a label-free pipeline.If reference wake/NREM epochs are absent, all epochs estimate the percentiles; when q95 ≤ q5, the implementation clips the shifted score to [0, 1].Source: §G.3 · p. 30
  • Depth–stage correlation and pairwise concordance use pooled raw wake/NREM scores, while REM metrics pool epoch-level probabilities across recordings.Percentile confidence intervals use 200 pooled-epoch bootstrap resamples, not participant resampling; CFS uses the same pipeline separately.Source: §G.3 · p. 30

Appendix H Operational definitions of ECG-SDI biomarkers

  • Whole-night biomarkers were calculated from unsmoothed normalized depth_score_norm across all retained prediction epochs, including wake and REM; AP used trapezoidal integration at unit epoch spacing, normalized by N rather than PSG total sleep time.Source: Appendix H · p. 31
  • Approximate entropy (APPe) and detrended fluctuation analysis (DETRf) characterize temporal complexity using specified embedding and fluctuation procedures.APPe uses embedding dimension 2, a one-epoch delay, Chebyshev distance, and self-matches; DETRf is the log fluctuation-versus-window-size slope.Source: Appendix H · p. 31
  • REM summaries use PSG reference labels: MDR combines ECG-derived depth with reference REM intervals, whereas PR is a reference-stage summary.These calculations use the exported reference depth_label rather than predicted REM class.Source: Appendix H · p. 31

Appendix I Clinical coding, phenotype estimation, and survival analysis

The appendix describes cohort-specific phenotype construction, clinical-outcome coding, and association and survival analyses. It also specifies paired-change analyses and a participant-level comparison of mean ECG-SDI within PSG-defined N2.

Appendix I Clinical coding, phenotype estimation, and survival analysis

  • Eight standardized biomarkers entered separate two-component Gaussian mixture models by cohort; components were oriented as disturbed or normal using mean AHI, not as a clustering input.Records missing any selected feature were excluded, and the component with higher mean AHI was labelled disturbed.Source: Appendix I · p. 32
  • Questionnaire indicators were operational symptom or perceived-quality definitions rather than diagnoses; sleep apnea was defined as AHI ≥5 h−1.The cohort-specific source fields for questionnaire indicators and clinical conditions are summarized in Table 7.Source: Appendix I · p. 32, Table 7 · p. 32
  • Subtype differences used Cohen’s d for continuous measures and logistic regression for binary associations, with age, BMI, subtype, and sex included when available.Confidence intervals used 500 bootstrap replicates; subtype assignments remained fixed, and no multiplicity adjustment was applied.Source: Appendix I · p. 32
  • SHHS mortality analyses compared subtypes with Kaplan–Meier estimates and log-rank tests, then fitted Cox models with progressively expanded covariate adjustment.The final model added AHI and available total sleep time; missing values for several clinical covariates were coded zero.Source: Appendix I · pp. 32–33

J.1 Participant-level changes and association models

  • Paired-change analyses related reductions in APPe to reductions in ODI3 and reductions in CV to increases in sleep efficiency; adjusted models treated PSG changes as outcomes.Both adjusted analyses included 395 paired participants and characterized associations between paired changes, not individual treatment-response predictions.Source: §J.1 · p. 34

J.2 Mean depth within PSG-defined N2

  • The N2 analysis averaged ECG-SDI across each participant’s PSG-defined N2 epochs at each visit and compared paired participant-level means.Eligibility required at least 10 minutes of N2 at both visits; a two-sided paired t-test evaluated the mean paired difference.Source: §J.2 · p. 34

Related papers

This paper cites

Loading 2610.23727v1…