Change Point--Aware Evaluation and Re-Calibration of PPG-Based Blood Pressure Estimation

Yunwon Tae; Minje Park; Gyunho Rho; Dongjoon Yoo; Sunghoon Joo

eess.SPAug 19, 2026

Change Point--Aware Evaluation and Re-Calibration of PPG-Based Blood Pressure Estimation

Authors: Yunwon Tae, Minje Park, Gyunho Rho, Dongjoon Yoo, Sunghoon Joo

Organizations: VUNO Inc. · University of California, Irvine · Research Institute for Future Medicine, Samsung Medical Center · Department of Critical Care Medicine and Emergency Medicine, Inha University Hospital

Abstract

Non-invasive continuous blood pressure (BP) monitoring using photoplethysmography (PPG) is a promising alternative to cuff-based measurements. However, existing PPG-based BP estimation studies predominantly rely on aggregated performance metrics (e.g., mean absolute error) computed over entire evaluation intervals, which can obscure model failures during rapid BP fluctuations and limit clinical relevance. In this work, we propose a fluctuation-aware evaluation framework for PPG-based BP estimation based on time-series change point detection. Instead of heuristic BP thresholding (e.g., $Δ\mathrm{BP} > 10\mathrm{mmHg}$ ), we identify BP change points by capturing abrupt distributional shifts in BP trajectories and evaluate estimation performance specifically during these fluctuation periods. Our analysis shows that several state-of-the-art models exhibit substantial performance degradation around BP change points, and that periodic test-time calibration is insufficient to handle such dynamic BP variations. To address this limitation, we introduce a targeted re-calibration framework triggered by detected BP change points, improving robustness without modifying model architectures. To the best of our knowledge, this is the first systematic evaluation of PPG-based BP estimation from a BP change point perspective, highlighting the importance of fluctuation-aware evaluation and calibration for real-world continuous BP monitoring.

Figures & tables

Figure 1: Limitations of existing PPG-based blood pressure (BP) estimation approaches under BP fluctuations. (a) Calibration-free methods track the overall BP trend but fail to capture rapid fluctuations (red region), resulting in large errors during clinically critical periods. (b) Calibration-based methods with periodic re-calibration (orange) struggle when BP shifts occur between calibration points. Targeted re-calibration at BP fluctuation points (green) enables timely adaptation and maintains accurate tracking throughout.

Figure 2: Overview of the proposed targeted re-calibration framework. (a) Ground-truth blood pressure (BP) change points are obtained from the reference BP sequence. (b) Supervised contrastive learning trains a PPG encoder to attract features within the same segment (defined by change points) and repel those from different segments. (c) Online change point detection via sequential clustering: PPG features are sequentially compared against cluster centroids; low similarity to the current cluster triggers change point detection. (d) Detected change points supplement periodic calibration, enabling timely re-calibration upon physiological shifts.

Figure 3: An illustration of a sequential clustering algorithm for PPG-based blood pressure change point detection and re-calibration. (1) The PPG feature ( $\mathrm{z}_{t}$ ) is extracted sequentially from the incoming PPG signal ( $\mathrm{x}_{t}$ ). (2) Similarity between incoming features and the current cluster centroid ( $\mathrm{c}_{curr}$ ) determines whether the physiological state changes. (3) The change point is detected when similarity remains below the threshold for $K$ consecutive frames, triggering cluster update and re-calibration.

	MIMIC-III		VitalDB
	Train	Test	Train	Test
N. Subjects/Cases	1,213/2,095	135/236	1,293/1,347	144/146
N. 10s Segments	436,680	54,000	465,480	57,600
N. Unstable Segments	53,518	6,085	54,507	7,132
SBP (mean ± SD)	121.94 ± 22.58	122.47 ± 21.83	115.48 ± 18.93	115.43 ± 18.72
DBP (mean ± SD)	60.73 ± 13.14	60.91 ± 12.58	62.92 ± 12.08	63.02 ± 11.87

Table 1: PulseDB Dataset Statistics for MIMIC-III and VitalDB.

Models	Whole		Unstable BP Intervals
Models	SBP ↓	DBP ↓	SBP ↓	DBP ↓
Baseline (Median)	16.48 ± 0.04	9.75 ± 0.02	21.10 ± 0.07	11.85 ± 0.04
Vanilla
Inception1D	13.79 ± 0.03	8.36 ± 0.02	17.04 ± 0.07	8.86 ± 0.04
PaPaGei-S	13.68 ± 0.03	8.66 ± 0.02	15.81 ± 0.07	9.12 ± 0.04
Pulse-PPG	13.72 ± 0.03	8.37 ± 0.02	17.84 ± 0.07	9.38 ± 0.04

Table 2: Comparison of BP estimation performance on whole time-series segments and clinically unstable BP intervals. Mean absolute error in mmHg with 95% confidence intervals are reported. Baseline (Median) represents the result of predicting the training set median for all samples.

Calibration Points	TTC		PPG2BP-Net		Points / Case
Calibration Points	SBP ↓	DBP ↓	SBP ↓	DBP ↓	Points / Case
Every 120 minutes	11.28 ± 0.06	6.22 ± 0.03	9.33 ± 0.05	5.21 ± 0.03	3.30 ± 1.52
+ $\Delta$ BP	9.89 ± 0.05	5.34 ± 0.03	8.05 ± 0.05	4.53 ± 0.03	17.36 ± 22.57
+ Random	9.27 ± 0.05	5.08 ± 0.03	7.44 ± 0.04	4.01 ± 0.03	9.91 ± 4.02
+ PELT	8.97 ± 0.05	4.94 ± 0.03	5.84 ± 0.04	3.85 ± 0.02	9.82 ± 4.02

Table 3: Impact of targeted re-calibration using ground-truth (offline) blood pressure change points. Mean absolute error in mmHg with 95% confidence intervals is reported. The bold represents the significant difference (

p<0.05

) against other baseline methods. Points / Case denotes the average number of re-calibration points per case (mean ± std). The plus sign (“

+

”) keeps the 120-minute schedule as the base and adds the change points of strategies such as

\Delta

BP , Random , and PELT . Random uses the per-case point count of PELT with uniformly sampled locations. Additional results for the unstable subgroup are provided in Appendix F.1 .

Models	AUROC	AUPRC	Weighted Accuracy	IoU
Inception1D	0.7275	0.3079	0.6457	0.6149
PaPaGei-S	0.7223	0.3070	0.6412	0.6062
Pulse-PPG	0.7308	0.3066	0.6112	0.6099
AnyPPG	0.7286	0.3075	0.6176	0.6106

Table 4: Online Change Point Detection Performance.

Calibration Points	TTC		PPG2BP-Net		Points / Case
Calibration Points	SBP ↓	DBP ↓	SBP ↓	DBP ↓	Points / Case
Every 120 minutes	11.28 ± 0.06	6.22 ± 0.03	9.33 ± 0.05	5.21 ± 0.03	3.30 ± 1.52
+ Random	9.52 ± 0.05	5.20 ± 0.03	7.60 ± 0.04	4.07 ± 0.03	8.77 ± 3.34
+ Online CPD	9.24 ± 0.05	5.12 ± 0.03	7.27 ± 0.04	3.87 ± 0.02	8.77 ± 3.33

Table 5: Performance of targeted re-calibration using the proposed online change point detection. Mean absolute error in mmHg with 95% confidence intervals is reported. The bold represents the significant difference (

p<0.05

) against other baseline methods. Points / Case denotes the average number of re-calibration points per case (mean ± std). The plus sign (“

+

”) keeps the 120-minute schedule as the base and adds the change points of strategies such as Random and Online CPD . Random uses the per-case point count of Online CPD with uniformly sampled locations. Both Classification and Online CPD use Inception1D as the CPD backbone. Additional results for the unstable subgroup are provided in Appendix F.3 .

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Parameter	Value
model	rbf
min_size	1
jump	1
pen	5

Appendix

Table 6: Hyperparameters used for ground-truth BP change point detection using the PELT implementation in ruptures .

PPG Preprocessing
Parameter	Value
Filter type	Third-order Butterworth
Filter bandwidth	0.5–8.0 Hz
Normalization	Z-normalization
PPG Encoder Training
Epoch	50

Appendix

Table 10

Figure 4: Effect of periodic calibration interval on BP estimation error, with and without PELT-based change point triggering.

Figure 5: Qualitative results for calibration-based methods using periodic re-calibration (orange) and targeted re-calibration at BP fluctuation points (green).

Figure 6: Qualitative results for calibration-based methods using periodic re-calibration (orange) and targeted re-calibration at BP fluctuation points (green).

Figure 7: Qualitative results for calibration-based methods using periodic re-calibration (orange) and targeted re-calibration at BP fluctuation points (green).

Parameter	Value	Points/Case	$\|\Delta_{\mathrm{Seg}}\|$ SBP	$\|\Delta_{\mathrm{Seg}}\|$ DBP
Penalty	1	26.75	11.69	6.16
	5 ∗	9.09	13.18	6.81
	10	6.00	12.04	6.27
Model	l1	56.72	7.52	3.94
	l2	121.70	6.32	3.33
	rbf ∗	9.09	13.18	6.81

Appendix

Table 7: Sensitivity analysis of PELT hyperparameters. The selected default values are marked with an asterisk.

Parameter	Value	AUROC	Accuracy	IoU
$\eta_{\mathrm{base}}$	0.999	0.7271	0.6412	0.6140
	0.99 ∗	0.7275	0.6457	0.6149
	0.9	0.7190	0.7214	0.5656
$\lambda$	1.0	0.7444	0.2391	0.2793
	0.1	0.7401	0.5271	0.5758
	0.05 ∗	0.7275	0.6457	0.6149

Appendix

Table 8: Sensitivity analysis of online CPD hyperparameters. The selected default values are marked with an asterisk.

Calibration Points	Whole		Unstable BP Intervals		Points / Case
Calibration Points	SBP ↓	DBP ↓	SBP ↓	DBP ↓	Points / Case
Every 120 minutes	9.33 ± 0.05	5.21 ± 0.03	10.93 ± 0.17	5.78 ± 0.09	3.30 ± 1.52
+ $\Delta$ BP	8.05 ± 0.05	4.53 ± 0.03	9.46 ± 0.15	3.88 ± 0.08	17.36 ± 22.57
+ Random	7.44 ± 0.04	4.01 ± 0.03	9.10 ± 0.15	4.63 ± 0.09	9.91 ± 4.02
+ PELT	5.84 ± 0.04	3.85 ± 0.02	6.73 ± 0.13	4.11 ± 0.07	9.82 ± 4.02

Appendix

Table 9: Impact of targeted re-calibration using ground-truth (offline) blood pressure change points, evaluated with PPG2BP-Net . Mean absolute error in mmHg with 95% confidence intervals are reported over the whole recording ( Whole ) and over unstable BP intervals defined in Section 3.2 . Points / Case denotes the average number of re-calibration points per case (mean ± std). The plus sign (“

+

”) keeps the 120-minute schedule as the base and adds the change points of strategies such as

\Delta

BP , Random , and PELT . Random uses the per-case point count of PELT with uniformly sampled locations.

CPD Backbone	CPD Method	CPD		BP Estimation
CPD Backbone	CPD Method	AUROC ↑	IoU ↑	SBP MAE ↓	DBP MAE ↓
Inception1D	Classification	0.5332	0.3919	8.31 ± 0.04	4.56 ± 0.03
Inception1D	Ours	0.7275	0.6149	7.27 ± 0.04	3.87 ± 0.02
PaPaGei-S	Classification	0.5321	0.4552	7.79 ± 0.04	4.13 ± 0.02
PaPaGei-S	Ours	0.7223	0.6062	7.30 ± 0.04	3.87 ± 0.02

Appendix

Table 10: Comparison with a naive binary classification baseline for online change point detection. Detection performance and the corresponding downstream BP estimation error using PPG2BP-Net are reported.

Calibration Points	Whole		Unstable BP Intervals		Points / Case
Calibration Points	SBP ↓	DBP ↓	SBP ↓	DBP ↓	Points / Case
Every 120 minutes	9.33 ± 0.05	5.21 ± 0.03	10.93 ± 0.17	5.78 ± 0.09	3.30 ± 1.52
+ Classification	8.31 ± 0.04	4.56 ± 0.03	8.73 ± 0.09	4.72 ± 0.05	12.31 ± 22.59
+ Random	7.60 ± 0.02	4.07 ± 0.01	8.06 ± 0.09	4.33 ± 0.05	8.77 ± 3.34
+ Online CPD	7.27 ± 0.02	3.87 ± 0.01	7.76 ± 0.08	4.15 ± 0.05	8.77 ± 3.33

Appendix

Table 11: Performance of targeted re-calibration using the proposed online change point detection. Mean absolute error in mmHg with 95% confidence intervals are reported. Points / Case denotes the average number of re-calibration points per case (mean ± std). The plus sign (“

+

”) keeps the 120-minute schedule as the base and adds the change points of strategies such as Random , Classification , and Online CPD . Random uses the per-case point count of Online CPD with uniformly sampled locations. Both Classification and Online CPD use Inception1D as the CPD backbone.

CPD Backbone	Params (M)	FLOPs (M)	Forward (ms)	Cluster (ms)
Inception1D	0.5	591.8	21.9	0.03
PaPaGei-S	5.8	76.6	9.2	0.03
Pulse-PPG	28.5	1140.5	31	0.03
AnyPPG	5.9	245.5	26.2	0.04

Appendix

Table 12: Analysis on computation cost and latency of online change point detection

Context	Unstable Segments (%)
Vasopressor	6,094 (85.4%)
No vasopressor	1,038 (14.6%)

Appendix

Table 13: Distribution of unstable BP segments in VitalDB with respect to vasopressor administration.

Explore similar work

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
ExpertoRhythm: Morphology-Aware Learning for Waveform Reconstruction and Cuffless Blood Pressure Estimation from Single-Channel PPG
Sep 28, 2026Amir Arjomand, Kenneth B. Kent, Georgiy KrylovPhotoplethysmography Wearable Health Monitoring
University of New Brunswick, Fredericton, Canada
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

Explore similar work

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

ExpertoRhythm: Morphology-Aware Learning for Waveform Reconstruction and Cuffless Blood Pressure Estimation from Single-Channel PPG

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