A BP prediction pilot versus clinical validation AI
The FDA’s January 23, 2026 draft guidance for cuffless BP devices remains nonbinding and marked not for implementation. Its first recommended accuracy criterion uses mean signed error within −5 to +5 mmHg and error SD below 8 mmHg. These limits apply to different statistics; 5 mmHg is not an MAE target. A second criterion depends on participant-level error behavior and is not implemented here.
The draft also discusses static accuracy, stability over the labeled calibration interval and accuracy after pressure changes. For continuous measurement it recommends arterial-line reference measurements. Meeting numerical reference limits in another dataset is not equivalent to meeting that validation design.
What the pilot establishes
The accepted model results describe short later excerpts from the same personally calibrated recordings. They use a previously explored 40-person cohort, approximate automatic landmarks, offline filtering and smoothed targets. The reference is Finometer-reconstructed brachial pressure, not an arterial line. There is no controlled pressure-change test, long-term follow-up or uncalibrated new-person test.
Population recruitment and demographic and pressure-range coverage were not designed or verified for regulatory validation. Overlapping five-beat windows are not 1,030 independent participants. These observations constrain interpretation even when pooled bias and SD fall inside the first numerical reference limits.
Calibration burden and practical form factor
The models learn from 70 detected heartbeats with simultaneous waveform and reference-pressure measurements. Across these excerpts, that calibration spans approximately 46–100 seconds (median 62 seconds), before the later test period. Its burden is more than elapsed time: each person supplies a dense sequence of reference BP labels while lying on an instrumented bed, with ECG and finger PPG also recorded. This is substantially more information than a single initialization pressure value.
The results therefore demonstrate short-term prediction after a substantial personal calibration procedure, rather than a convenient replacement for an arm-cuff measurement. They do not show that a brief setup, one cuff reading, a wearable sensor or a different posture would achieve the same accuracy. Nor do they establish how often calibration would need repeating.
A bed-based system might serve a different purpose, such as repeated measurements during rest. Its practical value would depend on reliable sensing, tolerable calibration and useful accuracy over time. A device intended to offer cuff-like convenience would need separate evidence that a simpler setup and feasible calibration preserve performance under its intended conditions. These are usability and deployment questions in addition to the numerical agreement criteria.
The calibration protocol, error definitions and source-study comparison explain the specific evidence. The report is a research pilot, not evidence of device clearance or a clinically validated monitoring system.