Overview

Cuffless BP validation can look like generalization when training and test windows share people, calibration makes the task easy, or the pressure range and reference are poorly described. Subject-wise splits test unseen people, but external cohorts or sites test transportability beyond the original study. Calibration-based systems need dynamic and delayed testing, while calibration-free systems need independent subjects and a broad pressure spectrum. Baselines and documented references reveal what the reported error actually measures.

Learning outcomes

Sample-wise splits can reward subject or recording familiarity when the claim concerns unseen users.
Subject-wise evaluation is stronger, but external cohorts or sites are needed to test transportability beyond one study setting.
Calibration pathway determines the burden: test calibrated change dynamically and after delay, and calibration-free estimates across independent pressure ranges.

Keywords

Cuffless BPValidationSubject-wiseCalibrationBaselineTransportability

The Test Can Be Right for the Wrong Reason

For cuffless blood-pressure systems, an attractive error score is only the symptom. The harder question is whether the test separated the model from the people, conditions, and reference information it could have learned in advance. This chapter diagnoses the main ways a validation result can look like generalization when it is actually measuring subject familiarity, a calibration shortcut, a narrow pressure range, or an incompletely described reference. The practical outcome is a way to identify the missing test behind an overconfident claim.

The first diagnostic question is: new records or new people?

A BP recording is usually divided into windows or beats so that a model has many examples to learn from. Those windows are not automatically independent observations. Several can come from the same person, the same visit, or the same physiological episode. If some windows from a person appear in training and other windows from that person appear in testing, the model may use stable features of that person's body, sensor placement, or recording environment. It can then predict familiar subjects well without learning a relationship that transfers to someone new.

A sample-wise split separates windows. A subject-wise split separates people. The latter is a stronger test when the intended claim is that a cuffless system can estimate BP for an unseen user. It asks the model to confront a person's signal, pressure range, and idiosyncrasies without having seen that person's other windows during training. This is not a guarantee of clinical validity, but it removes one important shortcut.

The broader review found 34 subject-wise publications among 86 studies. Hold-out methods appeared in 48 studies and cross-validation in 38, with overlap possible because a study can use more than one approach. Among the 73 studies that used a cross-validation or hold-out method, 48—65.8%—did not employ subject stratification. The review also identified 13 sample-wise splits, four random splits, 12 studies that did not report the train-test ratio, and four that did not provide enough detail to judge the strategy.

These classifications do not prove that every non-subject-wise study leaked information. A sample-wise split may be appropriate for a narrowly defined within-subject task. The problem arises when the result is presented as generalization to new people. The test then answers a weaker question than the headline implies. The model may be learning a person-specific signature, and the reported error may reward it for doing so.

The review's pooled comparison reinforces the concern without resolving it. Subject-wise studies had a pooled systolic mean error of approximately 0.81 mmHg, whereas record-wise studies had approximately 1.50 mmHg and substantially wider dispersion. The exact difference should not be treated as a universal leakage penalty: the subgroups differ in many other ways, and heterogeneity remained high. It is evidence that validation design is related to the results being reported, not proof of one single cause.

A good internal split is not external validation

Subject-wise testing improves the independence of the evaluation, but it still uses the same study's devices, protocol, preprocessing choices, and recruitment environment. External validation goes further by testing an independently assembled cohort, site, or setting. It can reveal that a model learned features of a particular device, laboratory, population, or data-cleaning process rather than a broadly stable signal-to-pressure relationship.

Published by DeepMox

FAQ

Why is a subject-wise split stronger than a sample-wise split?

It keeps a person's windows from being shared across training and testing, so the evaluation better tests an unseen user rather than familiarity with that person's signal or recording.

What does external validation add to cuffless BP testing?

It tests an independently assembled cohort, site, or setting, revealing whether performance travels beyond the original devices, protocol, population, and environment.

How should calibrated and calibration-free systems be tested differently?

Calibrated systems should face dynamic BP changes and delayed testing after the cuff baseline, whereas calibration-free systems need independent people and a broad BP spectrum.

Previous9 minCuffless BP Agreement Metrics

Next9 minCuffless BP in Everyday Life

Cuffless BP Validation Design

The Test Can Be Right for the Wrong Reason

The first diagnostic question is: new records or new people?

A good internal split is not external validation

FAQ