Can single-lead ECG enable continuous sleep depth profiling and digital biomarkers?

A continuous depth score replaces five stage labels

Conventional scoring assigns each 30-second epoch one of five labels: Wake, N1, N2, N3 or REM. That scheme is coarse. Two epochs both scored N2 can sit at different points on the way into or out of deep sleep, and the label hides that movement. The paper's premise is that sleep depth is better treated as a continuum, and that the continuum can be estimated from the heart's electrical signal rather than from brain waves.

ECG-SDI is that estimate. For every 30-second ECG epoch the model outputs a scalar depth value, later normalized within each recording to a 0–1 scale using the 5th and 95th percentiles of reference Wake and non-REM epochs. Higher values mean deeper sleep. Because the normalization uses reference stage labels in addition to the model output, the reported scale is a research construct, not a label-free clinical reading.

The model is a two-stage network. A convolutional encoder turns each epoch's ECG, plus 39 handcrafted frequency and time–frequency features, into a representation; a Transformer then reads 15-epoch windows so each epoch is scored with surrounding context. After sleep-staging pretraining, the backbone was frozen and two new heads were trained: one for REM classification, one for continuous depth. Supervision came from ordinary stage labels, combined as REM cross-entropy plus a pairwise ranking loss over Wake and non-REM stages. No manually annotated continuous depth targets were needed.

Depth ordering holds up in an independent cohort

The headline validation metric is a Spearman correlation between the raw depth score and ordered Wake/N1/N2/N3 labels, with REM excluded because REM sits between wakefulness and deep sleep rather than at one end of the ordering. In the pooled internal test set drawn from SHHS, MESA and MrOS, that correlation was 0.828; REM classification reached an AUROC of 0.954.

Cohort-level results varied more than the pooled figure suggests: correlations were 0.730 in internal SHHS, 0.797 in MESA and 0.882 in MrOS, with REM AUROCs of 0.908, 0.934 and 0.987. The external Cleveland Family Study, never used for training or parameter selection, reached 0.848 and 0.977. The weakest setting was the Binzhou CPAP cohort, at 0.456 before and 0.546 after treatment, with REM AUROCs of 0.803 and 0.830.

That gap matters for interpretation. The high internal and CFS numbers describe how well ECG encodes stage order in those cohorts; the BZ-CPAP numbers show the same model transferring less cleanly to a clinical treatment cohort. The paper reports the stage-wise distribution as physiologically sensible, with Wake lowest, values rising from N1 to N3, N3 highest, and REM intermediate — but it also notes that Wake distributions in SHHS and MESA had an unexpectedly broad upper tail, which it attributes only tentatively to cardiovascular burden in those cohorts.

Within-stage variation is visible in the ECG itself

The claim that the index captures something stage labels miss rests on epoch-level examples from CFS. In one participant's N1 epochs, ECG-SDI values of 0.08, 0.36 and 0.46 paired with heart rates of 98, 86 and 79 beats per minute; in N2 epochs from another participant, values of 0.30, 0.59 and 0.80 paired with 91, 75 and 69 beats per minute. Deeper scores went with slower hearts in these non-REM examples.

The REM examples break that pattern: values of 0.32, 0.46 and 0.56 paired with 64, 73 and 58 beats per minute, so heart rate did not fall monotonically as depth rose. The authors use this to argue that ECG-SDI is not simply an inverted heart-rate scale, and that it reflects stage-dependent autonomic regulation rather than one cardiac variable.

Whole-night trajectories make the same point at a larger scale. In a lower-AHI CFS recording (total sleep time 406.0 minutes, sleep efficiency 81%, AHI 4.88 per hour), depth repeatedly approached the top of its range during N3 and fluctuated within single-stage intervals; its normalized area under the depth curve was 0.59. An OSA recording (303.0 minutes, 53% efficiency, AHI 40.19 per hour) showed more sustained low-depth stretches and recurrent interruptions, with an area of 0.38. These are descriptive illustrations from selected recordings, not a quantified comparison across the cohort.

A disturbed phenotype tracks quality and mortality

Eight biomarkers were computed from each night's depth trajectory: cumulative depth area, shallow-sleep burden below 0.2 and below 0.3, coefficient of variation, skewness, approximate entropy, detrended fluctuation analysis, mean REM depth and REM proportion. A two-component Gaussian mixture model was fitted separately in SHHS, CFS and MrOS, and the component with the higher mean apnea–hypopnea index was designated the disturbed subtype.

The features separating the two clusters differed by cohort. In SHHS the disturbed subtype showed higher shallow-sleep burden and overnight variability (Cohen's d = 2.00 and 2.05) and lower cumulative depth (d = −1.15); in CFS the largest separation was approximate entropy (d = 1.71), which was also higher in the disturbed subtype in MrOS. Despite that heterogeneity, the disturbed subtype had lower sleep efficiency in all three cohorts and was associated with poorer subjective sleep quality after adjustment for age, BMI, sex, sleep efficiency and sleep latency — adjusted odds ratios of 1.33 (95% CI 1.18–1.52) in SHHS among the reported values.

The outcome result is the one to read carefully. In SHHS, the disturbed subtype was associated with higher all-cause mortality after adjustment that included AHI and total sleep time: hazard ratio 1.20 (95% CI 1.05–1.38). This is an association between a cluster assignment and mortality in one cohort, not evidence that the depth profile causes the risk, and the confidence interval's lower bound sits close to 1.

Paired CPAP recordings shift depth in the expected direction

The strongest test of whether the index tracks something real is whether it moves when sleep improves. In BZ-CPAP, 507 participants had polysomnography before and after continuous positive airway pressure treatment, giving 1,014 paired recordings analysed with the participant as the unit. After treatment, normalized sleep-depth area rose by 6.1% and shallow-sleep burden below 0.3 fell by 16.0% — relative changes in cohort means, distinct from absolute point changes.

Within a fixed stage, the shift was small but present: among participants with at least 10 minutes of N2 at both visits, mean ECG-SDI within N2 increased by 0.017 (95% CI 0.004–0.030; paired t-test P = 0.013). That is the paper's clearest demonstration that the index responds to change inside a stage category that conventional scoring would call unchanged.

Exploratory analyses linked these shifts to standard PSG outcomes. Reductions in approximate entropy correlated with reductions in the oxygen desaturation index (Spearman rho = 0.24), and reductions in depth variability correlated with gains in sleep efficiency (rho = 0.12). Both associations stayed positive after adjustment, with standardized coefficients of 0.068 (95% CI 0.028–0.108) and 0.190 (95% CI 0.079–0.301) in the model that also adjusted for the baseline biomarker, across 395 paired participants per model. The authors describe the magnitudes as modest and explicitly do not treat the measures as interchangeable.

The home workflow is a demonstration, not a validation

A chest-worn single-lead ECG device in a modified Lead II configuration was connected to a smartphone interface, and the acquired signal was passed through the trained model and biomarker module to produce an overnight depth trajectory with summary fields for area, REM proportion, shallow burden and variability. This shows the acquisition-to-report chain can be assembled.

It does not show that the chain works at home. The paper states that the displayed interface, trajectory and numerical values are illustrative, that no home-based accuracy analysis was performed, and that no individual treatment-benefit classification was validated. The research analyses themselves used polysomnographic recordings with PSG-derived labels for normalization and for the REM-related biomarkers, so the demonstrated accuracy belongs to that setting.

Two further boundaries follow from the design. The model was supervised entirely by conventional stage labels, so it inherits whatever those labels encode and whatever inter-scorer variability they carry; the paper frames the continuous index as an addition to staging, not a replacement validated against an independent physiological reference. And the mortality and quality-of-life findings come from cross-sectional phenotyping in public cohorts, while the responsiveness evidence comes from a separate clinical cohort — the two strands support different claims.

Depth area and shallow burden are the markers to watch

If one result should guide how this paper is read, it is the pairing of cumulative depth area and shallow-sleep burden. They moved in opposite directions after CPAP — area up 6.1%, burden below 0.3 down 16.0% — and they were also the features that most sharply separated the disturbed subtype in SHHS, alongside overnight variability. REM proportion and mean REM depth, by contrast, depend on reference PSG labels in the public-cohort implementation, which makes them less portable to a label-free wearable setting.

The practical priority the evidence supports is therefore narrow: ECG-SDI is a candidate for repeated, low-burden tracking of sleep continuity, with depth area and shallow-sleep burden as the trajectory summaries most worth following, and with the BZ-CPAP correlation of 0.456 a reminder that transfer to a new clinical cohort is not automatic. Turning that into home monitoring requires the accuracy analysis the proof-of-concept did not perform.

Sources used in this answer