Cuff-less Blood Pressure AI Falls Short Across Demographic Groups, Study Warns

Blood pressure is one of the most fundamental vital signs in cardiovascular medicine, and its measurement has remained largely unchanged for more than a century. The standard sphygmomanometer, with its inflatable cuff and stethoscope or oscillometric sensor, delivers reliable readings but is bulky, uncomfortable, and poorly suited to the continuous, real-time monitoring that modern preventive cardiology increasingly demands. A new study published in PLOS Digital Health by Blessing Kolawole, Sinuo Fan, Yingjie Lao, and Valencia J. Koomson examines a promising alternative: estimating blood pressure from photoplethysmography signals, the tiny optical waveforms that smartwatches and wearables already record every day. The researchers’ central finding is sobering for the wearable industry, because the accuracy of machine-learning blood pressure models depends strongly on who is being measured.

Photoplethysmography, or PPG, works by shining light into the skin and measuring how much is absorbed or reflected as blood pulses through the vessels. Each heartbeat produces a characteristic waveform whose shape encodes information about arterial stiffness, vascular tone, and cardiac output. Because blood pressure and arterial properties are physically linked, engineers have long sought to translate PPG morphology into systolic and diastolic pressure estimates without any cuff at all. The appeal is obvious: a sensor already embedded in millions of consumer devices could become a continuous cardiovascular monitor, flagging hypertension long before a clinic visit would.

The complication, as the new study emphasizes, is that the PPG waveform is not a universal signal. Vascularization changes with individual physiology and with age, and those changes directly alter waveform morphology. Arterial walls stiffen over the decades, peripheral perfusion varies with skin tone, body composition, and vascular health, and the same machine-learning model that performs well in one population can drift substantially in another. A model trained predominantly on young, healthy volunteers may quietly lose accuracy when deployed on older adults or across demographic groups it rarely saw during training. In clinical applications, such silent degradation is not a statistical curiosity but a safety issue, because treatment decisions hinge on the numbers the algorithm produces.

To quantify this problem, the team trained three conventional machine-learning models for blood pressure prediction and then carried out a subgroup-aware analysis, evaluating performance not just across the whole dataset but within individual demographic groups and at their intersections. This intersectional approach matters because aggregate accuracy can mask serious failures in subpopulations. A model whose mean error looks acceptable overall may be systematically biased for a particular age group or for specific combinations of demographic characteristics, and only a breakdown by subgroup reveals where the algorithm is quietly underperforming. The study’s framing reflects a broader movement in medical machine learning toward fairness auditing, in which demographic robustness is treated as a first-class requirement rather than an afterthought.

The second pillar of the work concerns signal quality. PPG recordings are notoriously vulnerable to motion artifacts, the distortions introduced when a sensor shifts against the skin, when a wearer moves, or when ambient light interferes with the optical measurement. These perturbations corrupt the waveform in ways that can mimic or obscure genuine physiological features, and if corrupted segments flow into the feature-extraction stage, the downstream model inherits errors it cannot correct. The researchers show that signal quality assessment is a pivotal step in the signal-processing pipeline before feature extraction, acting as a filter that screens out unreliable segments and mitigates external perturbations before they contaminate the training and inference data.

The payoff for this discipline is concrete. When signal quality assessment was incorporated into the modeling pipeline, the team achieved a mean absolute error of 3.36 millimeters of mercury for systolic blood pressure and 1.76 millimeters of mercury for diastolic blood pressure. To put those numbers in context, the Association for the Advancement of Medical Instrumentation has historically evaluated cuff devices against error standards in the range of roughly 5 millimeters of mercury, so errors of this magnitude from a cuff-less optical approach represent competitive performance, achieved here with conventional machine-learning models rather than opaque deep architectures. The result suggests that careful signal hygiene, not exotic model complexity, may be the decisive factor in making PPG-based estimation clinically credible.

The study’s demographic findings carry particular weight for the fast-growing market of consumer wearables that promise blood pressure tracking from the wrist. Devices sold in different regions and used by demographically diverse populations will encounter PPG waveforms whose shapes vary with age-related vascular changes and individual physiology. If manufacturers train and validate their models on narrow datasets, the subgroup-aware methodology in this paper offers a template for discovering exactly where those models are weakest. The intersectional analysis, which examines combinations of demographic attributes rather than each attribute in isolation, is especially important because health disparities often compound: the performance gap for a specific combination of characteristics can exceed what single-axis analysis would predict.

Technically, the work illustrates a pipeline that other groups can adopt. Raw PPG signals first pass through quality assessment, which rejects or down-weights segments degraded by motion artifacts and other external perturbations. Clean segments then feed feature extraction, where morphological characteristics of the waveform are distilled into the numerical descriptors that machine-learning models consume. Finally, the trained models map those features to systolic and diastolic pressure estimates, and the subgroup-aware evaluation framework audits performance across demographic strata. Each stage is standard in isolation, but the study’s contribution is to demonstrate that the stages must be engineered and evaluated together, with signal quality upstream and demographic fairness downstream treated as inseparable parts of a single measurement system.

The broader significance extends beyond blood pressure. PPG sensors are being repurposed for a growing list of physiological estimates, from respiration rate to oxygen saturation to arrhythmia screening, and every one of those applications inherits the same vulnerabilities the study documents: waveform morphology that varies across individuals, sensitivity to motion, and models whose accuracy is unevenly distributed across the population. By showing that signal quality assessment materially improves accuracy and that subgroup-aware evaluation exposes hidden performance gaps, the researchers provide a methodological blueprint that generalizes across the wearable health-monitoring landscape.

For clinicians and patients, the message is one of cautious optimism. Cuff-less blood pressure monitoring is edging closer to clinical usefulness, with error figures that approach the standards expected of traditional devices, but the path there runs through rigorous signal processing and honest demographic auditing rather than through raw model power alone. As continuous monitoring becomes a pillar of preventive cardiology, studies like this one define the standards that cuff-less technologies must meet before their readings can be trusted for anything more consequential than wellness trends. The researchers’ combination of conventional machine learning, systematic signal quality control, and intersectional subgroup analysis offers a realistic, reproducible route to blood pressure estimates that are not only accurate on average but accurate for everyone.

Subject of Research: Demographic subgroup-aware evaluation of machine-learning blood pressure prediction from photoplethysmography signals

Article Title: Demographic subgroup-aware analysis of blood pressure predictive models

Article References: Kolawole, B., Fan, S., Lao, Y., & Koomson, V. J. (2026). Demographic subgroup-aware analysis of blood pressure predictive models. PLOS Digital Health, 5(10), e0001687. https://doi.org/10.1371/journal.pdig.0001687

Image Credits: AI Generated

DOI: 10.1371/journal.pdig.0001687

Keywords: blood pressure, photoplethysmography, machine learning, wearables, signal quality, demographic subgroups, hypertension, digital health, PPG waveform, cuff-less monitoring, health equity, cardiovascular