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

Farouk Ganiyu Adewumi; Timothy Oladunni

cs.AIOct 21, 2026

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

Authors: Farouk Ganiyu Adewumi, Timothy Oladunni

Organizations: Department of Computer Science, Morgan State University, Baltimore, MD 21251, USA

Abstract

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.

Figures & tables

Stage	Participants	Eligibility or transition
Released dataset	1,221	673 participants have 11,597 auscultatory measurements; 548 released participants have no auscultatory records.
Cached waveform features	672	11,427 extracted rows from 672 participants. One of the 673 auscultatory participants has no available waveform; together with the 548 above, this accounts for the 549 outside this bank.
Historical calibration cohort	669	At least two valid initial calibration readings and future measurements. Three participants with only one initial reading are excluded. The historical anchor averages available initial calibration replicates.
Historical development/evaluation	501 / 168	Fixed split within the preceding 669. Evaluation participants were inspected across development analyses.
Primary common cohort	653	At least three valid initial calibration readings; 19 of the 672 have insufficient initial readings (16 have two; three have one). All 653 also belong to the historical cohort; 489/164 were in its development/evaluation groups.
Primary measurement endpoint	653	9,216 eligible future rows before excluding 1,337 later calibration-labelled rows; 7,879 readings remain. Initial anchor readings do not enter this future-row count or the fitted/scored endpoint.

Table 1: AuroraBP participant flow and analysis populations. Release-to-waveform counts were checked against the original participant and measurement tables; downstream eligibility and cohort overlap were reconciled using the cached feature-bank participant identifiers.

Dataset	Material analyzed	Role	Source
AuroraBP	653 participants; 7,879 scored measurements	Primary calibrated BP evaluation	Microsoft Research / Zenodo ( Mieloszyk et al., 2022a ; Mieloszyk et al., 2022b )
BIDMC	53 subjects; 1,272 windows	Historical PCA representation only	PhysioNet ( Pimentel et al., 2018 ; Pimentel et al., 2017 )
MIMIC-II	1,500 sampled windows	Historical PCA representation only	PhysioNet ( Saeed et al., 2011a ; Saeed et al., 2011b )
Welltory-PPG	106 windows	Historical PCA representation only	Published PPG collection ( Neshitov et al., 2021 )
UCI Cuff-Less BP	425 records; 3,268 windows	Prior training for original comparison	UCI repository ( Kachuee et al., 2015a ; Kachuee et al., 2015b )
Appel–Theart	108 participants; 630 windows	Independent pulse-rate fusion evaluation	Mendeley Data ( Appel and Theart, 2025 )

Table 2: Datasets and analysis roles. AuroraBP access requires provider approval and a Data Use Agreement. Access details are consolidated in Data Availability. UCI-derived inputs are excluded from the clean factorial.

Arm	SBP MAE [95% CI]	DBP MAE [95% CI]
Calibration retention	9.790	6.767
No added context	8.133 [7.824, 8.455]	6.265 [6.019, 6.526]
Optical quality only	8.110 [7.802, 8.431]	6.273 [6.025, 6.535]
Wearable-state context	7.697 [7.407, 8.008]	6.139 [5.899, 6.391]
Protocol-enriched upper bound	7.523 [7.243, 7.815]	5.743 [5.527, 5.977]
Primary gain [97.5% CI]	0.436 [0.305, 0.573]	0.126 [0.053, 0.200]

Table 3: Original repeated participant-level validation (653 participants; three-reading anchor). Arm intervals are marginal 95%; the two gain intervals are 97.5% to target approximate 95% simultaneous coverage by a Bonferroni rationale. Holm adjustment applies separately to paired test

p

values. The shared core can contain PPG descriptors and the UCI-derived prior; “no added context” is not a no-PPG control.

With elapsed-time terms
Arm	SBP MAE [95% CI]	DBP MAE [95% CI]
A: Calibration	8.426 [8.098, 8.768]	6.472 [6.215, 6.751]
B: Calibration + context	7.923 [7.610, 8.243]	6.263 [6.015, 6.532]
C: Calibration + PPG	8.111 [7.806, 8.432]	6.263 [6.015, 6.525]
D: Calibration + context + PPG	7.709 [7.414, 8.015]	6.128 [5.888, 6.383]
Without elapsed-time terms

Table 4: Reviewer-requested matched factorial, with and without time terms. All arms use the same 653 participants, folds, fixed estimator, weights and training-stage gate; no arm uses the UCI prior. A/B contain no optical inputs. Context is protocol-provided posture/activity; supplied optical quality is assigned to the PPG block. Intervals are exploratory marginal 95% participant-bootstrap intervals.

Endpoint	Arm	Bias (mmHg)	Error SD (mmHg)	$\|e\|>15$ mmHg (%)
SBP	No added context	-0.009	11.605	14.75
SBP	Wearable-state context	0.038	10.957	13.35
DBP	No added context	-0.086	9.846	7.19
DBP	Wearable-state context	-0.091	9.641	7.03

Table 5: Signed errors in the original primary comparison. Bias is prediction minus reference; positive denotes overestimation. SD is participant-balanced individual-error dispersion. Statistics are calculated separately per repeat and then averaged; MAEs are reported in Table 3 .

Figure 1: Agreement for the original wearable-state model in the first cross-validation repeat. Hexagonal bins show individual reading density, with prediction minus reference on the vertical axis. Solid lines mark participant-balanced bias; dashed lines mark approximate repeated-measures limits using within- and between-participant error variances. Limits describe individual-reading agreement under variance-model assumptions, not clinical interchangeability.

Arm	SBP MAE [95% CI]	DBP MAE [95% CI]
Calibration, time, wearable-state context	7.920 [7.607, 8.256]	6.318 [6.067, 6.589]
+ Standard PPG descriptors	7.733 [7.441, 8.048]	6.137 [5.898, 6.388]
+ Nonlinear/attractor family	7.709 [7.418, 8.017]	6.128 [5.887, 6.384]
Attractor-family gain [95% CI]	0.025 [ $-0.018$ , 0.069]	0.009 [ $-0.033$ , 0.050]
Holm-adjusted paired $p$	0.532	0.662

Table 6: Secondary Matched AuroraBP Feature-Family Ablation (653 Participants)

Target	Readings	Retention	Model MAE [95% CI]	Gain vs. retention
SBP	1	10.547	8.257 [7.933, 8.602]	2.290 [2.051, 2.525]
SBP	2	10.037	7.750 [7.458, 8.055]	2.286 [2.066, 2.509]
SBP	3	9.790	7.523 [7.243, 7.815]	2.267 [2.044, 2.489]
DBP	1	7.213	6.280 [6.026, 6.548]	0.934 [0.739, 1.129]
DBP	2	6.900	5.873 [5.648, 6.112]	1.027 [0.833, 1.215]
DBP	3	6.767	5.743 [5.527, 5.977]	1.024 [0.832, 1.211]

Table 7: Refitted Protocol-Enriched Models by Calibration Dose (653 Participants)

Step	MAE	Gain [95% CI]	Share
Session anchor (legacy)	7.970	–	–
$\rightarrow$ Prior anchor	6.223	$1.747$ [1.298, 2.200]	84.4%
$\rightarrow$ Absolute-error loss	6.110	$0.113$ [-0.013, 0.249]	5.5%
$\rightarrow$ +Context	5.773	$0.337$ [0.209, 0.467]	16.3%
$\rightarrow$ +HCA features	5.901	$-0.128$ [-0.242, -0.019]	-6.2%
Total (session $\rightarrow$ full)		$2.069$ [1.617, 2.516]	100%

Table 8: Reconstructed historical isolation ladder (SBP; common development-selected feature set; 168 evaluation participants). The fixed evaluation group was inspected during development; intervals are marginal and do not represent untouched independent validation.

Reference minus full MAE [95% CI]; negative favors reference:
Configuration	SBP MAE	DBP MAE
Reference	5.773	4.483
Fitzpatrick inputs only	5.770	4.479
+ Band reweighting and group gating	6.073	4.710
Full Fitzpatrick-input configuration	6.073	4.710
SBP	$-0.300$ [-0.419, -0.182]

Table 9: Reconstructed historical Fitzpatrick-input comparison (168 evaluation participants). The fixed evaluation group was reused during development; this is exploratory internal evidence.

Target	Scope	$n$	Raw	Corrected	$\Delta$ MAE [95% CI]
SBP	Overall	653	7.524	7.523	$-0.001$ [-0.037, 0.034]
SBP	I–II	516	7.487	7.484	$-0.003$ [-0.040, 0.035]
SBP	III–IV	112	7.683	7.645	$-0.038$ [-0.143, 0.062]
SBP	V–VI	25	7.571	7.777	$0.206$ [0.023, 0.386]
DBP	Overall	653	5.745	5.714	$-0.031$ [-0.061, 0.002]
DBP	I–II	516	5.683	5.642	$-0.041$ [-0.077, -0.004]

Table 10: Reconstructed corrected-versus-raw channel ablation in repeated participant CV. This separate historical configuration is not the primary model hierarchy. Intervals are marginal 95%.

Arm	Overall MAE (BPM)	MST 7–10 MAE
Visible only	4.366	5.434
Infrared only	5.037	5.165
Equal-weight fusion (post hoc)	3.985	4.587
Quality-weighted fusion	3.690	4.184
Guarded MST-aware fusion	3.703	4.265

Table 11: Independent Multi-Wavelength PPG Mechanism Test

Figure 2: Participant-balanced pulse-rate MAE with marginal 95% bootstrap intervals. The equal-weight comparison was added post hoc; quality weighting improves overall error relative to equal weighting, while guarded MST correction does not establish additional benefit.

Target	Retention	Context	Fixed blend	Blend $-$ context [95% CI]
SBP	25.291	15.827	15.375	$-0.451$ [ $-0.781$ , $-0.121$ ]
DBP	13.098	9.189	8.833	$-0.357$ [ $-0.512$ , $-0.194$ ]

Table 12: VitalDB frozen holdout (227 subjects). Fixed blend is the arithmetic mean of context and PPG-augmented predictions; negative blend-minus-context differences favor the blend. All entries are participant-balanced MAEs in mmHg.

Figure 3: VitalDB locked-holdout participant-balanced MAE. Calibration-plus-context explains most of the gain over retention; the frozen fixed context/PPG blend adds small CI-supported incremental improvements for both endpoints.

Study	Data and calibration/evaluation protocol	SBP	DBP
Joung et al. ( Joung et al., 2023 )	SNU surgical cohort, partly represented in VitalDB; paired calibration/inference network; participant-independent development/test split. Whole 797-subject test set.	7.991	4.789
Joung et al. ABP-20m subset	629 subjects assessed after 20 minutes following arterial-line insertion; subset of the preceding cohort.	5.525	3.282
Dias et al. ( Dias et al., 2024 )	VitalDB; 1,068 eligible patients; normalized PPG/calibration-window pairs 3 min–2 h apart, $\|\Delta\mathrm{SBP}\|\leq 60$ mmHg; unfiltered PPG result from author-posted Table 3.	9.20	5.68
Kim and Baek ( Kim and Baek, 2026 )	VitalDB target; cross-dataset pretraining followed by target-subject fine-tuning. Personalization uses 50–1,800 five-second segments; main comparison uses 1,800, with 720 validation and 720 test segments per subject.	3.53	1.96
This study	VitalDB; 55 development / 227 disjoint held-out subjects; three early arterial-reference windows; fixed blend evaluated on 5,472 future windows, with equal subject weighting.	15.375	8.833

Table 13: Calibrated intraoperative PPG studies. Errors are MAE in mmHg. Source-specific calibration and evaluation conditions are essential to interpretation; SNU denotes Seoul National University Hospital.

Model	SBP MAE	DBP MAE
Static-rich reference	7.363	5.470
Median/calibration ensemble	7.346	5.462
Direct-BP fixed ensemble	7.662	5.567
Causal sensor-history ensemble	7.299	5.445

Table 14: Later exploratory AuroraBP models: participant-balanced MAE (mmHg), with the same cohort, initial three-reading burden and participant folds within this series. The static-rich reference is separate from the original primary model in Table 3 . Its features and blending differ from that primary model. Specifications were frozen per experiment after earlier dataset-guided development.

Cuffless Blood Pressure Estimation

Cuffless blood pressure estimation infers blood pressure from signals such as photoplethysmography without an inflatable cuff.

See all 15 papers

Blood Pressure Estimation from PPG: A Comparative Study of Direct and ECG-Mediated Deep Learning Pipelines
Jul 26, 2026Bo Wu, Haoling Wang, Zhuodiao Kuang +1Photoplethysmography Wearable Health Monitoring
Department of Informatics and Networked Systems University of Pittsburgh Pittsburgh, USA · Department of Biostatistics and Health Data Science University of Pittsburgh Pittsburgh, USA · Language Technologies Institute Carnegie Mellon University Pittsburgh, USA
Attractor-Vascular Coupling Theory: Formal Grounding and Empirical Validation for AAMI-Standard Cuffless Blood Pressure Estimation from Smartphone Photoplethysmography
May 11, 2026Timothy Oladunni, Farouk Ganiyu AdewumiDynamical Systems Photoplethysmography
Department of Computer Science, Morgan State University, Baltimore, MD 21251 USA
Single-Beat Cuffless Blood Pressure Estimation Using Ear-PPG and ECG with a Lightweight Hybrid Learning Framework
Jul 29, 2026Kindeep K. Dhatt, Tengyue Wu, Hanbang Hua +1Wearable Health Monitoring Cuffless Blood Pressure Estimation
1. Vanderbilt Institute for Surgery and Engineering, Vanderbilt University, Nashville, TN, 37232 · 6. Department of Biomedical Engineering, Vanderbilt University, Nashville, TN, 37235 · 2. Department of Electrical and Computer Engineering, Vanderbilt University, Nashville, TN, 37235 +3

Jul 26, 2026J/K move · Enter open · S save

Blood Pressure Estimation from PPG: A Comparative Study of Direct and ECG-Mediated Deep Learning Pipelines

Bo Wu, Haoling Wang, Zhuodiao Kuang, Kateryna Shapovalenko

Department of Informatics and Networked Systems University of Pittsburgh Pittsburgh, USA · Department of Biostatistics and Health Data Science University of Pittsburgh Pittsburgh, USA · Language Technologies Institute Carnegie Mellon University Pittsburgh, USA

Photoplethysmography Wearable Health Monitoring Cuffless Blood Pressure Estimation

Open paper AskPDF arXiv

Continuous cuffless blood pressure (BP) monitoring is essential for connected health systems and wearable devices, enabling early detection, longitudinal tracking, and personalized management of cardiovascular disease. Many prior approaches attempt to estimate BP indirectly by reconstructing electrocardiography (ECG) from photoplethysmography (PPG), assuming ECG provides a stronger physiological link to BP. However, ECG sensing is less accessible in wearable settings and may introduce unnecessary complexity. In this work, we first perform a large-scale physiological correlation analysis on the MIMIC-III waveform database, revealing that PPG exhibits substantially stronger coupling with arterial blood pressure (ABP) (

|r|=0.247

p<0.001

) than ECG does (

r=0.018

p=0.187

), challenging the assumption that ECG provides a superior intermediate representation. Motivated by this insight, we conduct a systematic comparison between direct PPG-to-BP prediction and ECG-mediated pipelines using multiple state-of-the-art deep learning models. Across 1.74M segments from 3,127 patients, direct PPG-to-BP prediction achieves British Hypertension Society Grade A performance (

\mathrm{MAE}_{\mathrm{SBP}} = 4.82 mmHg

\mathrm{MAE}_{\mathrm{DBP}} = 4.31 mmHg

), outperforming all ECG-mediated approaches, which achieve only Grade B accuracy. Our findings suggest that accurate continuous BP monitoring can be achieved directly from wearable PPG signals, enabling simpler, more efficient pipelines for real-world connected health systems.

Abstract

Cuffless Blood Pressure Estimation

Explore similar work

Blood Pressure Estimation from PPG: A Comparative Study of Direct and ECG-Mediated Deep Learning Pipelines

Attractor-Vascular Coupling Theory: Formal Grounding and Empirical Validation for AAMI-Standard Cuffless Blood Pressure Estimation from Smartphone Photoplethysmography

Single-Beat Cuffless Blood Pressure Estimation Using Ear-PPG and ECG with a Lightweight Hybrid Learning Framework