What Would Make Cuffless BP Trustworthy?
The central subject of this series is cuffless blood-pressure monitoring, but the final question is more exact than whether the technology is promising. What evidence would justify treating an estimate as a trustworthy measurement for a stated use? The answer is a connected standard: the physiological proxy must remain interpretable, the reference and calibration pathway must be clear, the test must involve genuinely independent people and relevant conditions, the error distribution must be visible, and the population must be represented well enough to expose important failure modes. This chapter turns that standard into a practical way to audit a claim without treating a small average error as a clinical guarantee.
Trust is a stack, not a score
A cuffless BP result sits at the end of several linked questions. The first concerns the signal. PPG, ECG, imaging, impedance, timing biomarkers, and tonometry provide information related to cardiovascular pressure, but most are indirect proxies. Their morphology or timing can change with contact, motion, posture, hydrostatic effects, ambient light, autonomic regulation, tissue properties, or vascular stiffness. If the relationship between the proxy and arterial pressure changes in the intended setting, a model can remain mathematically consistent while its estimate loses physiological meaning.
The second question concerns the reference. A cuff, an invasive arterial waveform, and a non-invasive arterial reference are not interchangeable descriptions of the target. The reference device, measurement procedure, and timing of the paired observation determine what the model is being trained or judged against. For a calibrated method, the calibration device also establishes the baseline from which later change is interpreted. A low error relative to an incompletely described reference is therefore weaker evidence than the same error relative to a documented, appropriate reference under the intended conditions.
The third question is independence. Testing new windows from people already present in training can reward subject familiarity rather than a transferable signal-to-pressure relationship. Subject-wise evaluation is a minimum generalization check when the claim concerns new users, but it remains internal to the study's devices, recruitment, preprocessing, and environment. External validation tests whether the result travels to a new cohort, site, or arrangement. It is not a ceremonial extra: it tests whether the model learned the subject of interest or the conditions surrounding the original data.
The fourth question is calibration. Calibration-based and calibration-free systems make different claims. A calibrated system may be useful if it can track a person's change from a cuff-established baseline, but it must be challenged after time has passed and while BP changes dynamically. A calibration-free system must demonstrate meaningful absolute estimation across independent people and a broad pressure spectrum. The same mean error cannot carry the same meaning across both pathways because the model has access to different information and faces a different generalization burden.
The fifth question concerns the distribution of error. A small signed mean can result from positive and negative residuals cancelling. It does not show the typical absolute distance from the reference, the width of the agreement limits, the share of readings inside a useful band, or the conditions in which the largest errors occur. A credible result therefore needs a panel that includes signed bias and variability, absolute error, coverage within error bands, relative error where appropriate, and visualizations that reveal pressure-dependent or condition-dependent disagreement. The choice of metric should clarify the decision, not simply produce a favorable headline.
The sixth question is population and context. A large number of windows from an ICU or laboratory database is not the same as a diverse set of independent people living with different pressure ranges, vascular states, medications, body characteristics, and activity patterns. If subgroup descriptors and stratified results are missing, the evidence cannot support a confident fairness judgment.