PaperlayerView exact revision on arXiv

Source-linked AI summary

Context-Augmented, Session-Calibrated Blood Pressure Tracking from Photoplethysmography

Farouk Ganiyu Adewumi, Timothy Oladunni

arXiv:2610.29091v1cs.AI

TL;DR

The paper asks whether PPG contributes to session-calibrated BP tracking beyond the cuff anchor and acquisition context. It uses participant-separated evaluation and matched controls; results support small PPG/context contributions, with the DBP PPG increment sensitive to time removal, while clinical adequacy and transfer remain unestablished.

  • Problem

    The study asks whether PPG adds information about subsequent BP changes beyond an initial cuff anchor and acquisition context, under changing activity and posture.

    Source: §1.1 · p. 2

  • Method

    The study uses repeated participant-level validation and matched analyses to separate context and PPG contributions, remove elapsed-time terms, and evaluate distinct follow-up models.

    Source: §1.1 · p. 2, §4.4.1 · p. 9, §4.4.4 · p. 11

  • Results

    Matched calibrated controls showed small PPG/context contributions, with the DBP PPG increment sensitive to time removal; nonlinear augmentation had no established additional benefit.

    Source: §6 · p. 22, §8 · p. 24

  • Takeaways & Limitations

    The traceable workflow and retained negative findings do not establish clinical adequacy, model transfer, equitable performance, or cuff replacement.

    Source: §8 · p. 24

    AuroraBP is a retrospective internal validation after earlier dataset-guided development, and neither the separate pulse-rate nor VitalDB study establishes prospective wearable BP performance or transfer of one model.

    Source: §1.1 · p. 2, §7 · p. 23

Abstract

from arXiv · show

Objective.To distinguish the contributions of acquisition context and photoplethysmography (PPG) to session-calibrated blood pressure (BP) tracking.Approach. The original comparison used three repeats of five-fold participant-level cross-validation in 653 AuroraBP participants, excluding calibration-labelled prediction targets. Matched analyses separated context and PPG and removed elapsed-time terms. Later exploratory follow-ups retained PPG, accelerometer signals and three initial cuff readings. Separate studies evaluated pulse-rate fusion and an intraoperative VitalDB model. Main results.The original wearable-state model achieved participant-level mean absolute errors (MAEs) of 7.697/6.139 mmHg for systolic/diastolic BP (SBP/DBP). Added context reduced MAE by 0.436/0.126 mmHg (97.5% intervals 0.305-0.573/0.053-0.200). In the matched factorial without the external PPG prior, PPG improved on calibration, time and protocol-provided posture/activity by 0.214/0.135 mmHg; removing time left the DBP interval including zero. Standard PPG descriptors contributed, but additional nonlinear/attractor benefit was not established. In a separate follow-up series, a sensor-history ensemble achieved 7.299/5.445 mmHg versus its static-rich reference's 7.363/5.470, small gains of 0.064/0.025 mmHg. Larger DBP errors did not improve uniformly. Quality-weighted fusion reduced pulse-rate MAE from 3.985 to 3.690 beats per minute. VitalDB MAEs remained 15.375/8.833 mmHg despite small PPG gains. Significance. The contribution is a calibrated evaluation framework with explicit controls, reproducible feature-cache-to-results analyses and retained negative findings. The sensor-history gain is exploratory after repeated dataset reuse. Independently estimated wearable states, complete raw-signal reproduction of the supplied richer features, external transfer and clinical adequacy remain unestablished.

1 Introduction

The study asks whether PPG contributes to session-calibrated BP tracking beyond the initial cuff anchor and acquisition context, and audits context, calibration, and pigmentation interventions. It also reports distinct follow-up analyses while retaining negative findings and acknowledging repeated dataset reuse.

1.1 Motivation

  • The clinically relevant question is whether PPG adds information about subsequent BP changes beyond the initial cuff anchor and acquisition context.The study examines this retrospectively within controlled recording protocols, not treatment decisions or clinical outcomes.Source: §1.1 · p. 2
  • Participant overlap, calibration burden, and protocol shortcuts can obscure attribution, motivating participant-level validation and audits of anchoring, context, and pigmentation interventions.The primary repeated analysis remains internal validation after earlier AuroraBP development; fixed-holdout analyses are retained as historical development evidence.Source: §1.1 · p. 2

1.2 Study Questions and Contributions

  • The study’s primary comparison tests whether wearable-state context improves calibrated SBP/DBP tracking beyond a shared model core.Matched extensions isolate PPG and context and remove elapsed-time terms.Source: §1.2 · pp. 2–3
  • Separate analyses examine pigmentation interventions, recent-cuff anchoring, pulse-rate fusion, VitalDB, and PPG/accelerometer sensor history.Hashes, environments, folds, and numerical checks support computational reproduction, while repeated dataset reuse precludes prospective confirmation.Source: §1.2 · p. 3

2 Related Work

The related work frames cuffless BP tracking as a calibrated or calibration-free PPG problem, alongside concerns about nonlinear features, pigmentation, and evaluation rigor. The paper distinguishes empirical feature tests and signal-quality analyses from claims about physiological laws or equitable performance.

2.1 PPG-Based Cuffless Blood Pressure Estimation

  • Cuffless PPG BP methods include pulse-transit-time, waveform-feature, and end-to-end approaches, while calibrated methods use cuff readings to model subsequent change.Evaluation guidance emphasizes participant separation, calibration burden, and careful use of AAMI/ISO 81060-2 and BHS criteria.Source: §2.1 · p. 3

2.2 Nonlinear and Information-Theoretic Feature Representations

  • The feature bank combines waveform moments, delay-coordinate reconstruction, recurrence analysis, entropy, and spectral measures, but these descriptors are not treated as validated physiological laws.The matched ablation tests the empirical contribution of the combined nonlinear/attractor family rather than using prediction as evidence for its underlying theories.Source: §2.2 · p. 3

2.3 Skin Pigmentation and PPG-Derived Vitals

  • Pigmentation-related optical differences motivate separate evaluation of signal quality and downstream error, but optical interventions do not establish improved BP tracking or subgroup parity.The paper distinguishes pulse extraction from BP prediction and does not interpret nonsignificant subgroup differences as parity.Source: §2.3 · pp. 3–4

2.4 Evaluation Rigor and Data Leakage

  • Participant overlap can inflate reported accuracy, while dataset shortcuts can obscure PPG’s contribution; the audit addresses calibration-row indicators that may reveal near-zero targets.The cited work discusses record- or window-level split risks (González et al., 2023; Sanches et al., 2024) and dataset shortcuts.Source: §2.4 · p. 4

3 Framework and Audit Controls

The framework combines PPG descriptors, calibration anchoring, and acquisition context while distinguishing raw, corrected, and auxiliary simulated channels. Its audits find a context gain under strict calibration-row exclusion, while the tested pigmentation correction has no established overall benefit or equivalence.

3.1 Attractor Feature Bank

  • The feature bank uses eleven empirical PPG descriptors, including waveform statistics, entropy, recurrence measures, CSI, and a band-power quality ratio.It embeds sampled windows with d = 4 and τ = 5 samples; the descriptors are not validated physiological laws.Source: §3.1 · p. 4

3.2 Pigmentation Correction Operator

  • AuroraBP uses measured raw waveforms and heuristic-corrected channels, while the auxiliary simulated stress-test channel is not physiological ground truth and is excluded from the primary core.The correction rescales a median-centered waveform using recorded Fitzpatrick type and blends it with Wiener filtering; it does not estimate melanin or a Beer–Lambert parameter.Source: §3.1 · p. 4, §3.2 · p. 4
  • The tested pigmentation correction produced corrected-minus-raw MAE differences of −0.001 mmHg SBP and −0.031 mmHg DBP, with no established overall benefit or equivalence.The SBP 95% CI was −0.037–0.034 and the DBP interval was −0.061–0.002.Source: §3.2 · pp. 4–5

3.3 Calibration Dose and Anchoring

  • Primary tracking uses an initial three-reading anchor without later cuff updates, with return-visit targets 21.051–32.853 hours after the anchor.Historical prior anchoring can use cuff readings only minutes old within a visit, and neither scheme establishes cuff replacement.Source: §3.3 · p. 5

3.4 Context Operator

  • Wearable-state context uses protocol-provided posture/activity and supplied optical quality, so the no-added-context arm is not a no-PPG condition.The context states were not independently estimated and would require sensing or classification in deployment; the matched factorial supplies optical-input-free controls.Source: §3.4 · p. 5

3.5 Calibration-Row Leakage Audit

  • Strictly excluding calibration-flagged rows from fitting and scoring leaves positive context gains of 0.598 mmHg SBP and 0.557 mmHg DBP.Flagged rows had mean absolute SBP deltas of 6.774 mmHg versus 9.902 mmHg on development rows, so they were smaller but not trivially near zero.Source: §3.5 · p. 6

4 Methodology

The methodology centers on AuroraBP participant-separated evaluation, matched feature and context comparisons, and distinct follow-up analyses. It also documents retrospective-development limits, repeated-measures assumptions, and the different roles of pulse-rate and intraoperative studies.

4.1 Datasets

  • AuroraBP’s primary cohort includes 653 participants with three initial calibration readings, and excluding later calibration-labelled rows leaves 7,879 scored measurements.The dataset spans calibration, static, exercise, and temporal-challenge phases; the fixed 501/168 historical split is treated as development evidence, not untouched validation.Source: §4.1 · p. 6, §4.4 · p. 9

4.2 Feature Extraction and Selection

  • AuroraBP feature extraction uses a central 10-second optical window, a third-order 0.5–8 Hz bandpass, and timestamp-derived 500 Hz sampling.All 11,427 feature rows matched the sampling audit within 2 × 10^-11 Hz, and CSI recomputation agreed within 2.3 × 10^-16.Source: §4.2 · p. 8
  • The primary procedure ranks eligible feature changes, time terms, and the UCI prior within each outer training set, retaining up to 24 candidates plus baseline BP.The UCI prior was retained in 26/30 endpoint/fold cores and uses corrected optical features, so excluding explicit Fitzpatrick terms does not establish absence of indirect Fitzpatrick dependence.Source: §4.2 · p. 8

4.3 Model and Training

  • The primary model uses training-fitted median imputation and squared-loss HistGradient-Boosting with participant weights, but its change gate uses fitted training predictions rather than independent inner validation.The settings were inherited from development; this gate procedure is identified as a development limitation.Source: §4.3 · pp. 8–9

4.4 Evaluation Protocol

  • Three repeats of five participant folds keep feature ranking, imputation, fitting, and gate selection within outer training participants, with calibration targets excluded from fitting and scoring.MAE is averaged within participant and equally across participants; participant bootstraps preserve readings and repeats, but AuroraBP’s earlier development use makes this retrospective internal validation.Source: §4.4 · p. 9

4.4.1 Claim hierarchy and multiplicity

  • Positive gain is an absolute MAE reduction in mmHg, and the primary individual gain intervals use 97.5% levels to target approximate Bonferroni simultaneous 95% coverage for SBP/DBP.The sign-flip tests receive Holm adjustment; these procedures do not remove dependence from shared training sets or retrospective development.Source: §4.4.1 · p. 9

4.4.2 Matched context-by-PPG and time-removal analyses

  • The matched factorial compares calibration/time, posture/activity, PPG, and both together without UCI, corrected features, or enriched protocol metadata; a sensitivity analysis removes elapsed-time terms.The factorial uses supplied optical quality and eleven raw optical descriptors in the PPG block, making its context-only arms optical-input-free; its exploratory marginal intervals do not extend the primary test family.Source: §4.4.2 · p. 10

4.4.3 Signed errors and repeated-measures agreement

  • Agreement statistics are computed per repeat with participant-balanced errors, while approximate limits rely on random-intercept variance components and assume homogeneous variation and conditionally independent residuals.Serial dependence, nonnormal tails, and BP-dependent errors may limit coverage, so the limits are descriptive rather than device certification.Source: §4.4.3 · pp. 10–11

5 Experiments and Results

Across repeated participant-level validation, adding wearable-state context reduced BP error, while matched analyses found smaller and time-sensitive PPG contributions. Agreement, feature-family, calibration-dose, historical-anchor, and pigmentation analyses qualify the scope of those results.

5.1 Primary Repeated Participant-Level Validation

  • Wearable-state MAEs were 7.697/6.139 mmHg SBP/DBP, and added context reduced MAE by 0.436/0.126 mmHg on 653 participants and 7,879 strict-endpoint measurements.The paired-gain 97.5% intervals were [0.305, 0.573]/[0.053, 0.200].Source: §5.1 · p. 12, Table 3 · p. 13

5.2 Matched Context-by-PPG Factorial and Time Sensitivity

  • With time terms, PPG reduced MAE by 0.214/0.135 mmHg with context; without time, its DBP gain was 0.087 mmHg with an interval including zero, and activity annotation drove much of context benefit.Without elapsed-time terms, context gains with PPG remained 1.198/0.299 mmHg; the time-free activity-label gains were 0.823/0.245 mmHg SBP/DBP.Source: §5.2 · pp. 12–13, Table 4 · p. 14

5.3 Primary Signed Errors and Agreement

  • Mean bias was near zero, but participant-balanced error SDs were 10.957/9.641 mmHg SBP/DBP, with approximate repeated-measures limits of [−21.084, 21.159]/[−18.756, 18.574] mmHg.The agreement limits are approximate and should not be assumed to provide uniform 95% coverage across BP levels.Source: §5.3 · p. 13, Figure 1 · p. 15

5.4 Matched Feature-Family Ablation

  • Adding standard PPG descriptors reduced participant MAE by 0.187/0.180 mmHg SBP/DBP, while nonlinear/attractor features added only 0.025/0.009 mmHg and showed no established advantage.The latter contrasts had Holm-adjusted p values of 0.532 and 0.662.Source: §5.4 · p. 14, Table 6 · p. 15

5.5 Calibration Dose

  • Calibration-dose models were refitted on identical participants and future measurements within the same folds; the three-reading setup represents one-session repeated calibration, not one-reading calibration.The reported dose gains belong to the protocol-enriched arm, not the lead wearable-state arm.Source: §5.5 · p. 15, Table 7 · p. 16

5.6 Decomposition of Historical Prior-Anchor Improvement

  • In the historical fixed-feature ladder, the net SBP reduction was 2.069 mmHg, with approximately 84% attributable to replacing the initial anchor with the most recent prior cuff reading.This quantifies dependence on the reference scheme, not cuff replacement; the historical HCA addition worsened SBP MAE by 0.128 mmHg.Source: §5.6 · pp. 15–16, Table 8 · p. 16

5.7 Secondary Pigmentation-Intervention Audit

  • The full historical pigmentation configuration worsened SBP/DBP MAE by 0.300/0.227 mmHg, and six V–VI holdout participants cannot establish fairness or parity.The reconstructed historical evaluation group was reused during development, making the evidence exploratory internal evidence.Source: §5.7 · p. 16, Table 9 · p. 17

5.7.1 Repeated-CV pigmentation-correction intervention

  • In the historical strict-cohort channel ablation, the full corrected-channel-plus-Fitzpatrick configuration changed SBP MAE by +0.003 mmHg (95% CI −0.038 to 0.043) relative to raw; its exploratory DBP decrease was 0.0382 mmHg.The DBP contrast was exploratory and unadjusted across several configurations and does not establish general pigmentation-correction benefit or equitable performance. The historical comparison included only six V–VI participants.Source: §5.7.1 · pp. 17–18

5.8 Independent Quality-Fusion and Guarded-Correction Test

Separate studies found a benefit from quality-weighted pulse-rate fusion and small incremental PPG gains in VitalDB and an exploratory sensor-history series. The gains remain bounded by high VitalDB errors, nonuniform individual accuracy, and comparisons that do not establish clinical adequacy or broad transfer.

5.8 Independent Quality-Fusion and Guarded-Correction Test

  • Quality weighting reduced pulse-rate MAE from 3.985 to 3.690 BPM, while guarded MST correction established no additional benefit.The quality-weighted comparison also improved on either single wavelength; the MST guard activated in 5% of outer folds.Source: §5.8 · p. 18, Figure 2 · p. 19

5.10 State-of-the-Art Comparison

  • Published intraoperative estimates cannot be directly ranked against this VitalDB result because populations, selection, calibration, and averaging differ.No comparator was rerun on the 227-subject holdout, so relative performance under an identical protocol remains untested.Source: §5.10 · pp. 18–19

5.9 VitalDB Context-Plus-PPG Frozen Holdout

  • 15.375/8.833 mmHg VitalDB MAEs fell to 15.375/8.833 mmHg with the fixed blend from context-only values of 15.827/9.189 mmHg.The blend improved on context for 67.0%/70.5% of subjects, but absolute errors remained high and do not establish clinical adequacy or prospective wearable performance.Source: §5.9 · p. 18, Table 12 · p. 19, Figure 3 · p. 20

5.10.1 VitalDB error structure and practical interpretation

  • VitalDB errors were high at elevated pressures and large anchor departures, and no high-error observations were removed or holdout-based retuning performed.For SBP strata below 110, 110 to below 140, and at least 140 mmHg, MAEs were 15.559, 9.436, and 34.104 mmHg; for DBP strata below 60, 60 to below 80, and at least 80 mmHg, they were 7.694, 6.174, and 22.845 mmHg.Source: §5.10.1 · pp. 20–21

5.11 Exploratory Sensor-History Extension

  • 7.299/5.445 mmHg were the sensor-history series’ lowest observed SBP/DBP MAEs, with gains of 0.064/0.025 mmHg over its static-rich reference.The history comparison did not establish a nonlinear/attractor feature effect, and DBP errors above 20 mmHg increased from 2.50% to 2.59%.Source: §5.11 · pp. 21–22, Table 14 · p. 22

6 Discussion

The discussion treats context and PPG findings as conditional predictive information, not evidence of a universal mechanism or deployment-ready system. It emphasizes retained limitations in wearable-state estimation, generalization, and clinical adequacy.

6 Discussion

  • The primary comparison supports adding wearable-state context to a calibrated core, while matched analyses show small conditional optical contributions.Removing clock terms reduces one protocol shortcut, but context still uses supplied posture/activity; independently estimated wearable states were not evaluated.Source: §6 · p. 22
  • The separate studies answer distinct questions: pulse-rate fusion improved with quality weighting, while sensor history modestly improved MAE without uniformly improving DBP tails.Direct BP targets worsened error; the findings identify conditional predictive information, not an attractor-specific mechanism or universally generalizing architecture.Source: §6 · pp. 22–23
  • Near-zero aggregate bias coexisted with broad individual errors, and the results do not establish clinical adequacy, cuff replacement, or clinical outcome benefit.The contribution is a transparent calibrated evaluation with retained negative findings.Source: §6 · p. 23

7 Limitations

The evidence is limited by AuroraBP’s controlled single-site protocol, three-reading calibration burden, and prior dataset-guided choices. Several analyses also have scope, inference, and reproduction boundaries that prevent broader validation claims.

7 Limitations

  • Participant-separated folds do not remove optimism from earlier dataset-guided choices, and the controlled single-site protocol does not establish prospective wearable performance or transfer of one model.AuroraBP return targets occur 21.051–32.853 h after initial calibration; Appel–Theart and VitalDB address separate questions.Source: §7 · p. 23
  • Primary tracking requires three initial cuff readings; the one-reading protocol-enriched arm had higher MAE (8.257/6.280 mmHg), but a matching wearable-state dose analysis is unavailable.Protocol annotations and enriched metadata may encode acquisition structure, while the history ensemble retains time and protocol-dependent sampling.Source: §7 · p. 23
  • Only 25 primary and six historical V–VI participants were represented, so null or adverse results for one correction establish neither equivalence nor impossibility of better interventions.Auxiliary pigmentation ranks are synthetic; pulse fusion does not validate BP equity, and VitalDB has no pigmentation labels.Source: §7 · p. 23
  • Exploratory intervals do not account for adaptive experiments, overlapping training sets, or method selection, and the DBP router’s 2.9% recall precludes a safety claim.BHS/AAMI/ISO summaries are numerical context, not certification.Source: §7 · p. 23
  • Re-extracting 11,427 standard AuroraBP rows checked cache consistency, not independent algorithm validity; rich/beat extraction and complete third-party raw-to-results replication remain outstanding.The newer-runtime bridge verifies one original fold, not every fit.Source: §7 · pp. 23–24

8 Conclusion

Matched calibrated controls and participant-separated evaluation found small PPG/context contributions, while additional nonlinear benefit was not established. Sensor-history gains were small and exploratory, and separate pulse-fusion and VitalDB analyses do not establish clinical adequacy or transfer.

8 Conclusion

  • Matched calibrated controls and participant-separated evaluation showed small PPG/context contributions, with the DBP PPG increment sensitive to time removal.Nonlinear augmentation had no established additional benefit.Source: §8 · p. 24
  • Sensor history gave small exploratory gains relative to its own reference, without uniformly improving larger DBP errors.Source: §8 · p. 24
  • Independent analyses supported pulse-rate fusion and a small PPG contribution in a separate VitalDB model.Source: §8 · p. 24
  • The traceable workflow and retained negative findings do not establish clinical adequacy, model transfer, equitable performance, or cuff replacement.Source: §8 · p. 24

Code and Reproduction Availability

The original companion archive provides the primary analysis and supporting computational materials, while reviewer extensions and follow-up chains are in a separate revision package. Reproduction is scoped to specified inputs and checks; full third-party replication remains outstanding.

Code and Reproduction Availability

  • The versioned original companion archive contains the primary analysis, feature-family ablation, helpers, dependencies, hashes, manifests, and aggregate outputs.Source: Code and Reproduction Availability · p. 24
  • Reviewer extensions, historical reconstructions, and the static-rich/three-follow-up chain are in the accompanying revision package, not the original repository.Restricted feature banks, participant predictions, and contrasts remain local under dataset access conditions.Source: Code and Reproduction Availability · p. 24
  • Feature-cache-to-results reproduction requires hashed inputs and the recorded environment; rich/beat extraction was not independently regenerated, and complete third-party replication remains outstanding.The standard AVCT bank has a separate waveform re-extraction check; the newer follow-up fits used equivalent Python sources and local workers.Source: Code and Reproduction Availability · p. 24

Data Availability

AuroraBP participant-level data are not redistributed by the authors and are available through its provider subject to application and agreement. Other datasets are identified by their respective access routes, with historical MIMIC-II availability limited.

Data Availability

  • AuroraBP is available from its provider upon application through the cited Zenodo record and acceptance of the Data Use Agreement; the authors cannot redistribute participant-level data.Source: Data Availability · pp. 24–25
  • BIDMC, UCI Cuff-Less BP, multi-wavelength data, and VitalDB are available through their named repositories, while the cited MIMIC-II version is no longer distributed by PhysioNet.The multi-wavelength data are cited to Appel and Theart (2025), and VitalDB to Lee and Jung (2022).Source: Data Availability · p. 25

Ethics Statement

This secondary analysis used deidentified data without participant contact or intervention and did not require institutional review or additional consent. It followed the Declaration of Helsinki and applicable local requirements and was not a clinical trial.

Ethics Statement

  • The analysis used previously collected, deidentified human data, with no recruitment, intervention, participant contact, direct identifiers, or attempt to reidentify participants.The authors report that the study was not a clinical trial.Source: Ethics Statement · p. 25
  • Morgan State University IRB review was not required under its published process for deidentified or publicly available data, and no approval number was issued.Ethics approvals and written informed-consent procedures for original collections are reported by dataset providers and source publications.Source: Ethics Statement · p. 25
  • No additional consent or consent for publication was required because the analysis used only deidentified secondary data and reports no identifiable participant information.The analysis followed the Declaration of Helsinki and applicable local requirements.Source: Ethics Statement · p. 25
Loading 2610.29091v1…