Smartphone Blood Pressure Claims: What Cuffless PPG Validation Must Prove

If your team is deciding whether a smartphone photoplethysmography (PPG) blood-pressure feature can carry the word “accurate,” the evidence bar is now set by the papers making the claims themselves. The strongest 2026 result in the record, the Attractor-Vascular Coupling Theory (AVCT) model, reports an author-reported mean absolute error (MAE) of 1.98 mmHg systolic and 0.96 mmHg diastolic. Yet its own authors state that prospective evaluation with repeated reference measurements is required before any claim of continuous monitoring, device-level equivalence, or standards compliance is warranted. Meanwhile, an independent benchmark of PPG-based models found that none of the models it evaluated in their original form met the AAMI/ISO 81060-2 accuracy requirements (source), and the one study in this record that used a smartphone camera reports MAE of 3.08–16.64 mmHg and concludes that it “did not meet AAMI and BHS accuracy standards” (source). The practical consequence: retrospective error numbers from convenience cohorts, however small, are not clinical-accuracy evidence, and the burden now sits on repeated reference measurements, subject-level held-out evaluation, and per-user calibration-stability data.

Every accuracy figure in this article is author-reported from a single study with no independent replication, and the cohorts, datasets, and protocols differ across papers, so cross-paper comparison is indicative rather than equivalent.

The 2026 claims: sub-2.5 mmHg errors and a “BHS Grade A” from retrospective cohorts

The AVCT paper, Attractor-Vascular Coupling for Theory-Guided Interpretable Cuffless Blood Pressure Estimation from Photoplethysmography, is the most interesting claim because it is also the most honest. Its flagship result on a 217-subject ambulatory PPG-BP cohort reports SBP/DBP MAE of 1.98/0.96 mmHg, with subject-level MAE staying below 5 mmHg for 200 of 217 participants on systolic pressure and 215 of 217 on diastolic (source). A subject-level distribution like that is more informative than a single aggregate number, and the theory-guided, interpretable framing lowers the plausibility barrier: the model’s error can be decomposed into calibration bias and within-subject noise rather than being treated as one opaque residual. The paper is published in Expert Systems with Applications, and its calibration analysis reports an anchoring-effect ratio of εcal = 0.846 under leave-one-subject-out cross-validation; that ratio carries the sharpest protocol lesson in the record, and the protocol section below takes it up in full.

A September 2026 preprint, ExpertoRhythm, goes further. It reports author-reported SBP/DBP MAE of 2.46/1.46 mmHg on the UCI cuff-less BP dataset of 942 subjects (source) and attaches standards language directly: “our evaluation (Table 4) meets AAMI mean/SD requirements for both SBP and DBP and achieves BHS Grade A for both pressures.” That is a retrospective evaluation of a convenience dataset carrying compliance vocabulary. This is exactly the claim practice clinical-affairs and regulatory reviewers will start seeing in slide decks, and it is the practice the rest of this article interrogates.

Why one cuff reference per subject cannot validate a continuous claim

The decisive detail in the AVCT paper is not the MAE. It is the reference design: “This dataset provides 651 PPG windows but only one cuff BP reference per subject” (source). Three PPG windows per subject, one cuff measurement to anchor all of them.

A single reference measurement per subject cannot expose the two failure modes that matter most for a continuous-monitoring claim. First, within-subject noise: blood pressure varies with posture, activity, caffeine, and time of day, and a model evaluated against one snapshot per person never has to demonstrate that it tracks any of that variation. Second, calibration decay: cuffless PPG models are typically calibrated per user against a cuff reading, and the open question is how long that calibration holds. A cohort with one reference per subject is structurally incapable of measuring drift, because there is no second reference to drift away from.

The AVCT authors say as much themselves: “Prospective evaluation with repeated contemporaneous reference BP, prespecified equivalence or noninferiority margins for representation comparisons, and smartphone-acquired PPG is required before claims of continuous monitoring, device-level equivalence, or standards compliance are warranted.” Read that sentence as a checklist the authors have handed to reviewers. Their own headline number does not satisfy it, and they do not pretend otherwise.

The adversarial benchmark: no unmodified model met AAMI/ISO 81060-2

The counterweight to the optimistic single-paper results is a 2026 benchmarking study, Benchmarking and Enhancing PPG-Based Cuffless Blood Pressure Estimation Methods, built on a standardized subset of 101,453 high-quality PPG segments from 1,103 adults derived from MIMIC-III and VitalDB (source). Its author-reported finding: “none of the evaluated models met the AAMI/ISO 81060-2 accuracy requirements (mean error < 5 mmHg and standard deviation < 8 mmHg)” (source).

One scope condition matters when quoting that sentence: it covers the models as originally published. The same benchmark reports that its demography-enhanced MInception variant, which adds age, sex, and BMI as inputs, reaches 4.75/2.90 mmHg MAE with an error standard deviation of 6.12/3.84 under calibration-based evaluation, and the authors describe it as meeting AAMI/ISO criteria (source). The defensible summary is that no unmodified evaluated model passed, while one modified, demography-conditioned, calibrated variant reached the numerical limits on this dataset; a vendor deck can quote the second sentence under the shadow of the first.

That threshold pair is worth understanding, because aggregate MAE conceals half of it. The AAMI/ISO 81060-2 requirement as the benchmark states it constrains both the mean error and the standard deviation of the error. A model can be unbiased on average and still fail the spread requirement, and this is not a hypothetical. A study of ECG-plus-PPG systolic estimation in surgical patients reports an author-reported mean difference of 0.08 mmHg alongside a standard deviation of difference of 7.97 mmHg, within a hair of the 8 mmHg limit. A marketing page quoting “average error under 0.1 mmHg” from that study would be technically accurate and clinically misleading. When a vendor cites MAE, the first question is what the error standard deviation was, and the second is how many subjects sat above 5 mmHg.

Protocol sensitivity: how 2 mmHg becomes 14.2 mmHg

Reported accuracy in this literature is heavily protocol-dependent rather than an intrinsic property of PPG. The sharpest quantification is in-paper, not cross-paper: the AVCT study reports MAE of 12.85 mmHg unanchored and 1.98 mmHg after one-point reference anchoring, on the same cohort, the same model, and the same leave-one-subject-out split (source). The authors express that as ρanchor = 1 − 1.98/12.85 = 0.846, a descriptive 84.6% MAE change, and warn that it “must not be interpreted as evidence of recovery of within-subject BP changes.” The cross-paper version of the same dependence appears in a multi-wavelength PPG study using curriculum-adversarial learning, which reports author-reported SBP/DBP MAE of 14.2/6.4 mmHg under strict subject-level splitting, roughly a sevenfold gap versus the same-modality headline numbers above.

The mechanism is leakage. When PPG windows from the same person appear in both training and test sets, a model can learn subject-specific waveform signatures rather than a generalizable relationship between waveform and pressure. Random record-level splitting flatters every model that undergoes it. Leave-one-subject-out cross-validation and strict subject-level splits are the minimum conditions under which a held-out error estimate says anything about a new user. Any accuracy claim that does not state its splitting protocol should be read as record-level until proven otherwise.

Calibration protocol is the second lever, and the AVCT anchoring gap above is its clearest face. The DMT transformer paper reports an author-reported MAE of 4.56 mmHg systolic and 2.62 mmHg diastolic under calibration-based evaluation on the PulseDB dataset, materially worse than the zero-shot-looking headlines elsewhere, and arguably more honest about what deployment involves: most cuffless systems need a per-user cuff calibration, so their real-world accuracy is the calibrated accuracy, including whatever happens as that calibration ages. Evidence on calibration decay over weeks or months is thin across the entire record reviewed here, and single-reference-per-subject cohorts cannot supply it.

For comparison across the record:

These numbers come from different cohorts, sensors, and protocols. The table is a map of how much the answer moves with the method, not a leaderboard.

What to demand before shipping, approving, or advertising

The admissibility checklist below converts the evidence above into gate criteria. It is an inference from the cited papers, anchored on the AVCT authors’ own stated requirements; it is not a regulator’s published guidance, since no FDA, notified-body, or standards-body documents were reviewed for this article.

1. Reference density. Repeated, contemporaneous cuff (or better) reference measurements per subject, not one. Without them, within-subject noise and tracking ability are unmeasured.
2. Subject-level held-out evaluation. Leave-one-subject-out or strict subject-level splits, stated explicitly. Record-level splits disqualify the number.
3. Mean and standard deviation of error, plus subject-level distribution. The AAMI/ISO 81060-2 pair (mean error under 5 mmHg, SD under 8 mmHg) and the fraction of subjects above 5 mmHg. Aggregate MAE alone is insufficient; the surgical study’s 0.08/7.97 pairing shows why.
4. Calibration protocol and decay. Whether the reported accuracy is calibrated or calibration-free, and per-user calibration-stability evidence over a stated interval. A single-reference cohort cannot answer this.
5. Sensor and cohort realism. Smartphone-acquired PPG in an ambulatory population resembling the intended users, not hospital-derived waveforms (MIMIC-III, VitalDB) standing in for phone-camera signals. The AVCT authors state that their “present experiments used existing contact-PPG datasets and therefore do not validate smartphone-camera performance,” and they name smartphone-acquired PPG as a prospective requirement.
6. Prospective design with prespecified margins. Equivalence or noninferiority margins fixed before the evaluation, reported against them.

A claim that clears all six is admissible evidence for a clinical-accuracy statement. A claim that clears none, which describes the current smartphone-PPG record as supplied here, is a research result.

What AVCT shows, what standards require, and what the record cannot settle

The fair reading of the AVCT paper is that a mechanistic, interpretable model with a plausible error decomposition makes the cuffless PPG problem look tractable, and its subject-level error distribution is the right way to report results. But its authors correctly refuse to convert a retrospective, single-reference evaluation into a compliance claim. ExpertoRhythm’s choice to attach “meets AAMI” and “BHS Grade A” language to a retrospective dataset evaluation is the counterexample. The AVCT authors decline the same vocabulary, stating that their “reference-measurement design does not follow the required validation protocol,” and the benchmark authors report that prior evaluations “often lack the rigorous experimental controls necessary for valid clinical assessment.” Applying that vocabulary to a convenience-cohort retrospective is a category error on the papers’ own terms, whatever the MAE.

The limits of this article deserve equal clarity. Every number cited is author-reported, single-study, and unreplicated; the cohorts differ enough that no two results above are directly comparable. The benchmark finding that no unmodified model met AAMI/ISO 81060-2 applies to the models that benchmark evaluated, on its particular subset, and does not prove no model ever could. No regulator documents were reviewed, so nothing here asserts what the FDA or a notified body will accept; the checklist is the evidentiary bar the literature itself now implies.

The decision rule, then: treat any smartphone-PPG blood-pressure accuracy claim as inadmissible until a prospective study demonstrates AAMI/ISO 81060-2 mean/SD compliance with repeated contemporaneous references per subject, subject-level held-out evaluation, and per-user calibration-stability data. If a vendor’s evidence pack is a retrospective MAE from a single-reference cohort, the burden of proof has not moved an inch, whatever the number says.

Frequently Asked Questions

What accuracy requirements must a cuffless PPG model meet under AAMI/ISO 81060-2?

The AAMI/ISO 81060-2 requirement as the benchmark states it constrains both the mean error and the standard deviation of the error. A model can be unbiased on average and still fail the spread requirement, and this is not a hypothetical. A study of ECG-plus-PPG systolic estimation in surgical patients reports an author-reported mean difference of 0.08 mmHg alongside a standard deviation of difference of 7.97 mmHg, within a hair of the 8 mmHg limit.

Why is a single cuff reference per subject insufficient for validating continuous monitoring?

A single reference measurement per subject cannot expose the two failure modes that matter most for a continuous-monitoring claim. First, within-subject noise: blood pressure varies with posture, activity, caffeine, and time of day, and a model evaluated against one snapshot per person never has to demonstrate that it tracks any of that variation. Second, calibration decay: cuffless PPG models are typically calibrated per user against a cuff reading, and the open question is how long that calibration holds. A cohort with one reference per subject is structurally incapable of measuring drift, because there is no second reference to drift away from.

What validation protocol is required before claiming standards compliance for smartphone PPG?

Prospective evaluation with repeated contemporaneous reference BP, prespecified equivalence or noninferiority margins for representation comparisons, and smartphone-acquired PPG is required before claims of continuous monitoring, device-level equivalence, or standards compliance are warranted.