When a Small Error Looks Like Good News
Cuffless blood-pressure monitoring is an attempt to estimate systolic and diastolic pressure without repeatedly inflating a conventional arm cuff. This opening focuses on a deceptively simple question about that technology: what does a small average error actually tell us? The answer matters because a reader may use a number reported in a study to decide whether a wearable can merely observe trends, support a clinician, confirm hypertension, or influence treatment. Those are not interchangeable claims.
The subject is therefore not a particular smartwatch or one preferred algorithm. It is the measurement system formed by a sensor, a physiological signal, a model, a reference device, a population, and a testing protocol. The practical judgment this series supports is whether the evidence for such a system justifies the use being proposed. The central conclusion is deliberately narrower than either enthusiasm or dismissal: a small average discrepancy can be encouraging, but it is not, by itself, clinical reliability.
The opportunity is real because blood pressure is not static
A conventional cuff remains inconvenient precisely where repeated information could be most useful. Blood pressure changes with sleep, waking, posture, activity, stress, medication, and the body's autonomic responses. A reading taken in a clinic is a snapshot, and the clinical setting can itself alter the value. Home and ambulatory measurements can therefore add information beyond a single office reading; the newer review cites guideline support for using them in confirming hypertension and guiding therapy.
Cuffless devices are attractive because they could make observation more frequent and less disruptive. A device worn through ordinary activity might reduce the discomfort, sleep disturbance, and adherence burden associated with repeated inflation. That possibility is especially relevant to longitudinal or nocturnal monitoring, and to people who do not tolerate standard procedures well. But convenience describes the access to measurements, not their validity. A stream of estimates that is easy to collect is not automatically a stream of measurements accurate enough to diagnose disease or alter medication.
This distinction is easy to lose when a study compares every wearable reading with a cuff reading and then reports one summary number. The comparison may answer whether the two devices have similar average values under those study conditions. Clinical use asks a larger question: whether the estimate remains sufficiently accurate for the relevant person, pressure range, activity state, and decision.
“Comparable” depends on what was compared
An earlier systematic review and meta-analysis illustrates why the evidence can initially look reassuring. It combined 21 prospective studies involving 897 participants. Across those selected comparisons, the pooled systolic difference between wearable and cuff readings was 0.58 mmHg, with a 95% confidence interval from −0.35 to 1.51 mmHg. The pooled diastolic difference was −0.33 mmHg, with a 95% confidence interval from −1.50 to 0.85 mmHg. The authors found no statistically significant overall difference and concluded that wearable readings were comparable to cuff sphygmomanometers, while also saying that accuracy still needed improvement.
That finding is meaningful within its scope. It suggests that, in the included studies and mostly resting comparisons, the average values did not show a large systematic offset. It does not show that every individual estimate was close, that performance would persist during movement, or that a device could safely replace a validated cuff for diagnosis. The review itself was based on small prospective, non-randomized studies and noted that standards for cuffless measurement and hypertension diagnosis remained unsettled.
A newer systematic review examined a broader literature: 86 original studies of cuffless BP estimation published through March 2025, with a quantitative analysis of 46 effects reporting systolic mean error. Its pooled signed error was 1.12 mmHg, a small positive value indicating slight average overestimation. But the heterogeneity was extremely high: I² was 98.8%. The review explicitly warns that this pooled error should not be treated as a performance benchmark for a particular clinical use case. Its conclusion is that the current literature does not support strong claims of consistent or clinically meaningful performance.
These findings are not necessarily contradictory. The earlier review asks whether wearable and cuff readings differ on average in a selected set of paired, resting studies. The newer review asks how much reported performance varies across sensing methods, models, populations, calibration strategies, validation schemes, and conditions. A small average difference can be present within a controlled comparison while the broader evidence remains too inconsistent to support a clinical claim.
The average can be small while the person-level error is not
Mean error keeps the direction of a discrepancy. If one estimate is 8 mmHg above the reference and another is 8 mmHg below it, the two errors contribute zero to their average. That cancellation is not a statistical trick; it is the reason the mean is useful for detecting systematic shift. But it also means that the mean alone says little about how far individual estimates may be from the reference.
Consider an illustrative set of three residuals: +8, −8, and 0 mmHg. The mean error is 0 mmHg, yet two of the three readings are 8 mmHg away from the reference. A real study contains many more observations and more complicated dependencies, but the logic is unchanged. A near-zero signed average can coexist with a broad error distribution, clinically important outliers, or errors concentrated in one posture or pressure range.
The newer review reports that the individual study results it collected covered a wide range: systolic mean error values ran from −3.38 to 15.42 mmHg, while reported mean absolute differences ran from 0.018 to 11.3 mmHg. These are not interchangeable summaries, and they should not be compressed into the pooled mean. The high heterogeneity says that the published studies are not estimating one stable, common level of performance. They are drawing attention to different results produced by different systems under different conditions.
A small mean is therefore neither meaningless nor sufficient. It can indicate that a device has little average directional bias in a given comparison. It cannot, by itself, establish the spread of errors, the proportion of readings within a useful band, the limits of agreement, or the stability of performance for the people and circumstances that matter. Those questions require complementary measures and plots, and they require the reader to know how the data were produced.
The number sits on top of hidden conditions
The first hidden condition is what the sensor actually measures. A cuff or arterial waveform is a reference for pressure. Many cuffless systems instead observe a proxy: PPG records peripheral blood-volume changes, ECG records electrical activity, and other devices use optical imaging, bioimpedance, timing between signals, or contact with an artery. The model must infer pressure from that proxy. If the proxy changes because of contact, motion, vascular tone, or another factor that is not a matching change in systemic pressure, the same model input no longer has the same meaning.
The second is who and what was tested. A quiet seated participant, an intensive-care recording, and a person walking through a day do not expose a device to the same problem. The newer review describes motion, posture, hydrostatic effects, behavior, and autonomic fluctuations as central challenges. It reports that truly ambulatory evaluations remain scarce and that controlled data can lose performance when transported to ambulatory cohorts. “Validated” must therefore be followed by “under which conditions?”
The third is whether the test involved new people or merely new windows from familiar people. A model can learn stable individual characteristics when training and test records from one person are mixed. A randomly separated collection of samples may look independent while preserving information about the same subject on both sides of the split. Subject-wise evaluation is a stronger test of generalization, although it is still internal evidence rather than proof that the system will work at another site or in another population.
The fourth is calibration. A calibration-based device uses a cuff measurement to personalize the estimate and often aims to follow change from that personal baseline. A calibration-free device claims to estimate across people without that step. A static test soon after calibration and a delayed test during changing activity do not measure the same capability. Nor does a broad population BP estimate provide the same evidence as tracking a person's change after calibration. A reported error cannot be interpreted until this distinction is clear.
The fifth is the proposed decision. Evidence sufficient to study a research signal is not automatically sufficient to confirm hypertension, titrate medication, or replace a validated cuff. The stronger the decision consequence, the more the evaluation must address individual agreement, dynamic and long-term behavior, population diversity, and clinical outcomes. A convenient trend signal may have a legitimate role before an absolute reading is ready for medical decision-making.
The useful question is consequently not “Is the average error small?” It is “For the intended use, whose pressure is being estimated, from what signal, against which reference, under what conditions, with what subject separation, calibration, error distribution, and uncertainty?” This series follows that question from the physiology of the proxy to the statistics of agreement, the design of validation, and the gap between a laboratory result and everyday care.
The article-in-press review concludes that no generalizable, clinically validated cuffless BP system currently exists across the evidence it examined. That is not a claim that the technology has no value or that improvement is impossible. It is a claim about what the current evidence can responsibly support. A small average error may be the beginning of a trustworthy argument. It is not the argument's conclusion.