Prediction results

Can individually calibrated sensors track blood pressure?

Three personally calibrated approaches estimate systolic and diastolic pressure: a four-feature BCG reproduction, a 42-feature personal ridge regression, and the same 42 features with shared cohort slopes and personal deviations. Features and reference BP use trailing five-beat means. Models learn from windows ending through cycle 70 and are evaluated on five-beat averages ending at cycles 75–100, after a four-cycle gap. Calibration and evaluation methods.

Blood-pressure prediction across three calibrated models

40 participants · 1,030 matched held-out five-beat windows · pressure errors in mmHg.

Table 1. Personally calibrated models evaluated on the same later five-beat windows.
Pressure target	Model	Mean participant MAE	Bias	Error SD
Systolic (SBP)	Four-feature OLS reproduction	5.0895% interval 3.85–6.55	-2.15	7.23
Systolic (SBP)	Personal ridge · 42 features	3.8495% interval 3.09–4.68	-1.04	5.61
Systolic (SBP)	Cohort partial pooling · 42 features	3.5395% interval 2.85–4.30	-0.89	5.18
Diastolic (DBP)	Four-feature OLS reproduction	3.0295% interval 2.38–3.75	-1.10	4.15
Diastolic (DBP)	Personal ridge · 42 features	2.5695% interval 2.00–3.19	-0.61	3.75
Diastolic (DBP)	Cohort partial pooling · 42 features	2.4195% interval 1.83–3.05	-0.51	3.79

Mean participant MAE: calculate each person's mean absolute error, then average across people so each contributes equally. Bias and error SD combine all held-out five-beat windows. Errors are prediction minus reference. The MAE intervals resample whole participants. Definitions and aggregation.

Main finding

Sharing a common coefficient pattern while retaining personal deviations gave mean participant MAE of 3.53 mmHg for SBP and 2.41 mmHg for DBP, compared with 3.84/2.56 for independently fitted personal ridge. The pooled model had lower MAE for 28 / 28 of 40 participants (SBP/DBP). These observed reductions have paired uncertainty intervals spanning zero in this small cohort. Cohort-sharing method.

Sharing improved the average error but did not eliminate individual failures. For example, X1035's DBP MAE was 8.52 mmHg with pooling versus 5.13 with personal ridge. The figure below retains all participants and allows their pressure trajectories to be inspected.

Relation to the source publication

Carlson et al. (2021) introduced the dataset and a four-feature, personally fitted BCG regression. My approximate reproduction scores 5.08/3.02 mmHg (SBP/DBP) on the same later windows. These are my held-out reproduction results, not accuracy values quoted from the publication. The extensions change features and estimator; the comparison does not isolate feature count. Source method and reproduction.

A reference-only sanity check that repeats each person's average calibration BP had mean participant MAE of 4.96/2.95 mmHg (SBP/DBP). It is not a second sensor model.

Reference versus sensor-predicted pressure

Colors identify participants. Click a point or select a participant to inspect their held-out five-beat windows.

ModelParticipantViewCoordinates

Figure 1. All model options use the same smoothed references and windows ending at cycles 75–100. The reference-versus-prediction view includes an exact-agreement diagonal. Calibration-relative coordinates subtract the same person's early reference average from both axes. Error agreement shows prediction minus reference against their average, with bias and descriptive limits (±1.96 error SD) for the displayed points. Full ranges retain large failures. Table 1 summarizes all three models over the full cohort, independently of figure selection. Interpretation.

How does this pilot compare with validation guidance?

The FDA's January 2026 draft guidance for cuffless BP devices recommends, as its first accuracy criterion, signed mean error (bias) within −5 to +5 mmHg and error SD below 8 mmHg. These are two different measures; the 5 mmHg limit is not an MAE target. The document remains nonbinding draft guidance.

Pilot bias / error SD, in mmHg, using the same held-out five-beat windows as Table 1.
Model	Systolic: bias / SD	Diastolic: bias / SD
Four-feature reproduction	-2.15 / 7.23	-1.10 / 4.15
Personal ridge · 42 features	-1.04 / 5.61	-0.61 / 3.75
Cohort partial pooling · 42 features	-0.89 / 5.18	-0.51 / 3.79

All three approaches fall numerically within those reference limits in this pilot, which is encouraging for short-term prediction after personal calibration. This is not a clinical-validation pass: the 40-person, previously explored cohort uses smoothed targets and brief later excerpts against Finometer-reconstructed pressure. Formal evaluation also requires a second participant-level criterion and appropriate study design, including accuracy after pressure changes and across the claimed calibration interval; for continuous devices the draft recommends an arterial-line reference. Guidance context · Study limitations.