What the Cuffless Device Is Actually Measuring
Cuffless blood-pressure monitoring estimates BP from physiological signals that are related to pressure but are not pressure itself. This chapter follows that inference from the body to the displayed systolic or diastolic value, so the reader can identify where a signal, a model, or a testing condition may change what the number means. The key judgment is not whether a wearable contains an impressive sensor or a sophisticated algorithm; it is whether the relationship between the observed signal and arterial pressure remains sufficiently stable for the intended use.
The number on the screen is an inference
A conventional cuff measures pressure through a physical interaction with the artery. As the cuff inflates and deflates, the device observes oscillations associated with arterial pulsation and uses an established measurement procedure to estimate systolic and diastolic values. An arterial blood-pressure waveform, when available, is closer still to the target because it records pressure changes directly through an invasive or specialized reference arrangement.
Most cuffless devices take a different route. They record a signal that changes as the cardiovascular system changes, then infer pressure from its shape, timing, or relationship to other signals. The distinction is easy to state but important to preserve: a signal can carry information about BP without being a unique representation of BP. The same visible waveform change may result from a pressure change, a change in vascular tone, a movement artifact, a different sensor contact, or several of these at once.
This is why a cuffless estimate needs both a model and a reference. The model supplies the relationship between the signal and a pressure target. The reference supplies the values against which that relationship is learned or tested. Without the reference, the device may display a plausible number, but there is no basis for knowing whether that number is calibrated to the person's pressure or merely resembles a familiar pattern.
Different signals expose different pieces of physiology
Photoplethysmography, or PPG, is the most common signal in the reviewed literature. A PPG sensor shines light into tissue and measures changes in the light returned as blood volume in peripheral vessels varies during the cardiac cycle. That makes PPG useful for identifying pulse timing and waveform morphology. It does not make PPG a pressure sensor. Its amplitude and shape depend on where the sensor sits, how firmly it touches the skin, the optical path through tissue, and the condition of the peripheral circulation.
Imaging PPG, sometimes called remote PPG, uses a camera rather than a contact photodetector. It estimates small changes in reflected light from exposed skin. The absence of a strapped sensor can be useful, but the signal is exposed to lighting and movement in a particularly direct way. A small change in the camera, the face, or the illumination can alter the recorded pattern even when arterial pressure has not changed correspondingly.
Electrocardiography, or ECG, records the electrical activity that initiates cardiac contraction. It provides a useful timing reference and information about heart rhythm, but the electrical event is not the same as the pressure wave that reaches a peripheral artery. ECG-only BP estimation therefore relies on a less direct relationship. In the review of 86 studies, PPG appeared in 70 studies, 37 of them using PPG alone; ECG alone appeared in only four. PPG was combined with ECG in 29 studies, but adding a signal changes the available information, not automatically the validity of the resulting estimate.
Impedance plethysmography, or IPG, tracks changes in electrical impedance associated with blood-volume changes. It can provide another view of pulsatile flow, but its signal is sensitive to movement, pressure, temperature, and low signal-to-noise conditions. Tonometry takes a more direct physical approach than optical or impedance sensing: a contact transducer applies controlled force to partially flatten an artery and records the resulting arterial-wall deformation. Its closer contact with the artery can support more accurate estimation, but contact itself becomes a variable that must be controlled and reported.
Some systems combine signals to form timing biomarkers. Pulse arrival time, or PAT, is commonly defined as the interval between a cardiac electrical event, such as the ECG R-peak, and a fiducial point on a peripheral pulse waveform. Pulse transit time, or PTT, measures the travel time between pulse signals recorded at different locations. Pulse-wave velocity can then be related to the path length and transit time. These quantities can correlate with BP changes, which makes them physiologically attractive. A correlation is not a permanent conversion rule, however, because the timing includes more than the pressure wave's travel through the arteries.
How a proxy becomes an estimate
The signal-to-BP pathway usually contains several transformations. Sensors first capture a waveform or image sequence. Processing then removes some baseline drift and noise, aligns channels, segments usable beats or windows, and either extracts features or presents a cleaned signal to a learned model. A model maps those representations to systolic BP, diastolic BP, mean arterial pressure, a change from calibration, or—in another family of methods—a reconstructed arterial BP waveform from which those values are derived.
The review groups algorithms by what they are trying to estimate. Hemodynamics-derived models use physiological quantities such as PAT, PTT, or pulse-wave velocity in equations intended to reflect arterial behavior. Data-driven BP regression models learn a mapping from raw signals, engineered features, images, or demographic variables to scalar pressure values. ABP waveform reconstruction models attempt to recover the complete arterial pressure waveform and derive systolic and diastolic values afterward. The groups can use overlapping sensors, so the target and modeling relationship matter more than a label such as “wearable” or “deep learning.”
Calibration adds another branch to the pathway. A calibration-based model uses one or more conventional cuff readings to personalize its parameters, often estimating the change from a person's baseline rather than discovering an absolute pressure scale from the wearable signal alone. A calibration-free model tries to generalize across people without that personal reference. The signal-processing steps may look similar, but the claim is different. One system is asking whether it can follow an individual's change after calibration; the other is asking whether it can assign a meaningful pressure value to someone it has not calibrated.
At each step, information can be lost or assumptions can be introduced. A filter may remove a useful feature along with an artifact. A beat detector may align two channels incorrectly. A model may use demographic variables that improve prediction in the development cohort without measuring a physiological change in real time. A calibration procedure may work at rest but fail when the person's vascular state changes. The final number inherits all of these choices, even when the display presents it as a simple measurement.
Why the mapping moves when BP does not
One source of instability is the sensor's physical relationship with the body. Reflective PPG is affected by contact pressure and positioning. A strap that shifts, a wrist held at a different angle, or pressure from clothing can change the amount and distribution of light reaching the detector. The result may be a different pulse morphology before the model has considered BP. A camera-based signal adds illumination and framing to the list. IPG can respond to pressure or temperature changes, while contact-based tonometry depends on how the transducer interacts with the artery.
Body position introduces both physical and physiological changes. Raising or lowering an arm changes hydrostatic pressure along the vascular column and can alter the peripheral waveform. Standing, sitting, and lying down also change venous return, vascular tone, and the demands placed on cardiovascular regulation. If a model was trained mostly on seated rest, it may encounter a different combination of signal morphology and BP when the person stands or walks. The model is then being asked to extrapolate, not merely to repeat a familiar measurement.
Autonomic regulation creates a subtler problem. During exercise, stress, sleep, or ordinary transitions between activities, sympathetic and other regulatory responses change vascular resistance and the tone of peripheral vessels. PPG captures microvascular behavior that reflects both local and systemic influences. That behavior can change even when the relationship between central arterial pressure and peripheral volume is not fixed. In other words, the proxy can respond to the body's attempt to regulate circulation, not only to the pressure value the device is meant to report.
Timing biomarkers have their own confounding path. PAT includes the interval between electrical activation and the arrival of the peripheral pulse. That interval contains a pre-ejection period: the time between electrical activation and the heart's mechanical ejection of blood. Sympathetic activity can change this period. A shorter or longer PAT may therefore reflect a change before the pulse travels through the artery, weakening the assumption that PAT moves monotonically with BP. An algorithm trained in one autonomic state may interpret the same timing change differently in another.
The body itself also changes the signal relationship over time. Vascular elasticity and arterial stiffness vary with age and health status. The earlier review notes that vascular resistance, elasticity, and related physiological factors can influence wearable estimates and speculates that age-related stiffness may interfere with signals. That observation is a reason to test older and clinically diverse populations, not a fixed rule that every older person or every wrist device will perform poorly. The size and direction of the effect belong to the device, population, and conditions being studied.
What a more elaborate model can and cannot repair
Data-driven models are valuable precisely because cardiovascular waveforms contain patterns that are difficult to express in a single closed-form equation. A model can combine morphology, timing, frequency information, multiple sensor channels, and demographic context. Deep learning can learn hierarchical representations from raw or minimally processed signals, while waveform reconstruction can preserve more information than a single scalar prediction. These are legitimate technical advantages when evaluated with data that represent the intended use.
They do not eliminate the measurement problem. A model can learn a stable relationship only where the training data contain evidence of that relationship. If the training set links a particular waveform shape to a pressure value because all recordings came from one person, one posture, or one controlled device arrangement, the model may learn that context along with the physiology. If a new environment changes the waveform for reasons unrelated to BP, a more flexible model may fit the new pattern, but it may also produce a more confident extrapolation from the wrong cue.
The same point applies to multimodal sensing. Combining PPG and ECG can add information, but it does not guarantee that the signals remain aligned or that the added channel contributes beyond a population prior. Waveform reconstruction can produce a visually convincing curve while still missing the pressure range or beat-to-beat changes relevant to a clinical decision. The architecture is part of the system, not a substitute for testing the system.
The practical consequence is that a cuffless reading must be interpreted together with the conditions that produced its input. A PPG waveform is not interchangeable across a finger, wrist, camera, or laboratory arrangement merely because each is called PPG. A timing feature is not a pressure measurement merely because it correlates with pressure in a study. The proxy becomes clinically useful only when its relationship to the target is shown to remain sufficiently reliable across the people, states, and environments in which the claim is made.
That is the bridge to the next question. Once the signal-to-pressure mapping is understood as changeable rather than fixed, a small pooled error can no longer be read as a direct readout of sensor quality. The next task is to examine what aggregate error does—and does not—show about agreement with individual BP readings.