REF-1699 · Xu et al., 2026 · Sensors · Paper · PubMed

Optical (PPG) pulse-rate variability differed from ECG heart-rate variability at rest by a fraction of the ECG standard deviation on average

In a meta-analysis of 43 validation studies in healthy people, 8 of which could be pooled per measure, the gap between study means was 0.188 of the ECG standard deviation for RMSSD (95% CI 0.066 to 0.309) and 0.134 for SDNN (0.014 to 0.255, PPG reading higher); only resting or controlled recordings were pooled, no limits of agreement could be computed, no accepted threshold says how small an error is small enough, and agreement looked more variable in sleep, exercise and stress

Main caveat: The authors': only 10 of 43 studies could be pooled; representative-condition selection favoured rest.
moderate evidenceMeasured on non-athletes to trained athletesRead in full

Xu and colleagues asked how far heart-rate variability measured from an optical pulse signal (PPG, as in watches, rings, arm bands, finger clips and phone cameras) departs from the same variability measured from an ECG, in healthy non-clinical people.

They searched three databases from 2017 to May 2026 and kept 43 studies that reported RMSSD or SDNN against an ECG reference; only 10 gave the paired means and SDs needed for pooling, and one representative condition per study was chosen by a fixed hierarchy that preferred resting, seated or supine, short-term recordings.

Because Bland–Altman limits of agreement were not reported consistently, the pooled outcome was the absolute difference between the PPG and ECG group means divided by the ECG SD. For RMSSD this was 0.188 SD (95% CI 0.066 to 0.309; I² 11.69%) over 8 studies and for SDNN 0.134 SD (0.014 to 0.255; I² 0%) over 8 studies.

With direction kept, PPG read higher than ECG for SDNN (0.134, p = 0.029) and showed a non-significant upward trend for RMSSD (0.110, 95% CI −0.054 to 0.275). The 33 studies that could not be pooled suggested better agreement in controlled resting or paced-breathing recordings and more variable agreement in sleep, exercise and recovery, stress tasks and field recordings.

The authors conclude that PPG pulse-rate variability may serve for within-device trend monitoring under controlled conditions, that the pooled figures do not extend to sleep, exercise, stress or daily life, and that it is not interchangeable with ECG HRV.

Participants

43 studies in the qualitative synthesis: healthy or apparently healthy non-clinical participants — adults, athletes, physically active people, students and healthy adolescents; one mixed-population study kept only for its able-bodied subgroup.
10 pooled studies (Table 1), sample sizes 18 to 94; RMSSD analysis 362 participants and SDNN analysis 278 (our sums of the Table 1 n values; 429 across the 10 unique studies).
Sex, age, training level and competition level of the pooled samples: not in the main text (Table S1 not read).

Protocol

Search: PubMed, Web of Science, Scopus, records from 1 January 2017 to 1 May 2026 (final search 1 May 2026), plus reference lists and citation checking; 779 records, 581 screened, 295 full texts, 43 included.
Inclusion: PPG-derived (or other pulse-derived) beat-to-beat variability against an ECG-derived reference, reporting RMSSD and/or SDNN, in healthy non-clinical people. References included laboratory ECG, Holter, chest-strap ECG (Polar H10 in three pooled studies) and wearable ECG.
Selection of one condition per study: total sample over subgroups, healthy over clinical, rest or baseline over exercise, recovery, stress, posture change, free living or sleep, standard short-term over ultra-short, supine or seated over standing.
Analysis: absolute standardized error (ASE) as primary outcome, signed standardized error as secondary; random effects with REML (metafor); sampling variance assumed with a within-person correlation of 0.50 (sensitivity 0.25 and 0.75); leave-one-out, criterion-ECG-only and paced-breathing sensitivity analyses; Pearson correlations pooled as exploratory evidence.
Quality: a custom reporting-completeness checklist scored by two reviewers; not used as a risk-of-bias judgement. Certainty not graded; small-study effects not tested.

Key Results

Analysis (PPG vs ECG, resting or controlled)StudiesEstimate (units of ECG SD)95% CII²
RMSSD, absolute standardized error80.1880.066 to 0.30911.69%
SDNN, absolute standardized error80.1340.014 to 0.2550%
RMSSD, signed (positive = PPG higher)80.110, p = 0.188 (not significant)−0.054 to 0.27551.14%
SDNN, signed80.134, p = 0.0290.014 to 0.2550%
RMSSD, criterion ECG only60.1780.029 to 0.32722.9%
RMSSD, without the paced-breathing study70.1920.048 to 0.33523.85%
SDNN, criterion ECG only (chest-strap references removed)50.103 (not significant)−0.048 to 0.2540%
SDNN, without one study at a time70.108 to 0.156; CI crossed 0 after removing either of two studies——
RMSSD Pearson correlation (exploratory)6r = 0.9800.847 to 0.99897.33%
Single-study ASE for RMSSD ranged from 0.004 (a laboratory BioSemi PPG channel) to 0.524 (Empatica E4 wristband); Polar OH1 arm sensor 0.190, Kairos wristband 0.106, a phone-camera SDK 0.169.
Reporting completeness of the 10 pooled studies: mean 78.5% (range 70-85%); artefact correction or PPG signal-quality handling was not reported in 9 of 10, and preprocessing was clearly described in 1 of 10.
Narrative synthesis (33 studies, not pooled): agreement looked better under controlled rest, standardized posture or paced breathing; more variable in the 10 sleep or nocturnal studies, the 2 exercise or recovery studies, the 3 stress-task studies and field recordings. No numbers are given for these.

Practical Application

At rest, an optical sensor's HRV differed from an ECG reference by a fraction of the ECG standard deviation on average, but this review cannot say how far one morning's reading can be off. Across 8 studies of short resting or controlled recordings (one with paced breathing), the PPG mean sat 0.188 of the ECG standard deviation from the ECG mean for RMSSD and 0.134 for SDNN; PPG read higher for SDNN (signed 0.134, p = 0.029), while for RMSSD the upward lean was not significant (0.110, 95% CI −0.054 to 0.275); there are no individual limits of agreement, so a single reading's error is unknown.
Keep the measurement the same every time: same device, same position, same time, same duration, quiet breathing, and read the trend against the athlete's own baseline — the authors' own recommendation, consistent with SYNTH-91's same-device rule.
Overnight HRV from a ring or watch is not covered by these pooled numbers. The authors say sleep needs its own validation and found agreement more variable there; SYNTH-91's overnight-HRV row rests on device-specific night studies (REF-1251, REF-364), which this review does not replace. HRV taken during or right after exercise is not supported either.
Do not compare an optical HRV value with a chest-strap value, or across devices: the pooled correlation was high (r = 0.98) but the authors stress that correlation is not agreement and that results belong to the whole system (sensor site, sampling, filtering, artefact correction).
Inference, not tested here: for a morning HRV check, a chest strap remains the safer choice when a strap is at hand; an optical or camera reading taken the same way each day is usable as a within-person trend, which is what SYNTH-91 already allows for measured readings read as a trend.

Limitations

The authors': only 10 of 43 studies could be pooled; representative-condition selection favoured rest; heterogeneous devices, sensor sites, references and processing; artefact and signal-quality handling rarely reported; findings limited to RMSSD and SDNN in healthy people at rest or under control.
The pooled outcome compares group means scaled by the ECG SD: it has no accepted threshold for "good enough", depends on how spread out the ECG values were, and folds over- and under-reading together; it does not describe one person's error.
No risk-of-bias tool and no certainty grading; publication bias not tested.
Several pooled PPG sources are laboratory or in-house sensors (BioSemi, a finger plethysmograph, a "PPG sensor"), not consumer wearables; one pooled study (Schoffl 2023) lists the same laboratory system as PPG source and ECG reference.
The strict exploratory correlation analysis included a coronary-bypass patient study that the eligibility criteria excluded; the authors removed it only in the expanded analysis.
Sampling variances rest on an assumed within-person correlation (0.50); results were similar at 0.25 and 0.75, though the SDNN interval crossed zero at 0.25.
No external funding; the authors declare no conflicts of interest.
Supplementary tables (study characteristics, completeness scores) were not read, so the populations and devices of the 33 narrative studies are known only from the main text.
How this entry was checked and changed
Why this level

read in full, but the review applied no risk-of-bias tool and no certainty grading (only a descriptive reporting checklist, 70-85% per study), so the rubric's condition for A- is not met; in addition the pooled metric compares group means and gives no individual limits of agreement, and each pooled analysis rests on 8 small studies (HRV is the measure under validation, so no marker step)