PaperlayerView exact revision on arXiv

Source-linked AI summary

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

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

arXiv:2608.18639v1eess.SPcs.AIcs.LG

TL;DR

PPG blood-pressure estimators are commonly evaluated with aggregate metrics that can miss failures during rapid, clinically important fluctuations. This paper detects BP change points for fluctuation-aware evaluation and targeted re-calibration, revealing degradation during transitions and improved accuracy over periodic calibration alone.

  • Problem

    Aggregate evaluation of PPG-based BP estimation does not explicitly account for non-stationary BP fluctuations or identify when calibration is needed.

  • Method

    The framework detects distributional shifts in BP trajectories, evaluates unstable intervals separately, and triggers targeted re-calibration using online PPG-based change point detection.

  • Results

    Calibration-free models show 2–4 mmHg higher SBP MAE and 1–2 mmHg higher DBP MAE during fluctuations, while change point-triggered re-calibration outperforms periodic-only and ΔBP strategies.

  • Takeaways & Limitations

    Fluctuation-stratified reporting and calibration aligned with physiological transitions are important for assessing and improving continuous BP monitoring under dynamic conditions.

  • Takeaways & Limitations

    The framework is evaluated on ICU and operating-room PulseDB data, so generalizability to ambulatory or wearable monitoring remains unvalidated.

Abstract

from arXiv · show

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.

1. Introduction

PPG-based BP estimation methods often overlook abrupt, clinically important BP fluctuations, which aggregate evaluation and periodic calibration may fail to capture. This work introduces change point–aware evaluation and targeted, online re-calibration to improve robustness during dynamic BP transitions.

  • Motivation: Both calibration-free and calibration-based PPG methods fail to explicitly account for BP’s dynamic, non-stationary fluctuations.Abrupt transitions may arise from physiological stress, posture changes, medication effects, or acute pathological events.
  • Limitations: Whole-interval evaluation can appear accurate while masking substantial prediction errors during BP fluctuation intervals.Periodic calibration may reduce long-term drift yet miss sudden BP drops occurring between scheduled calibration points.
  • Study Design: The study evaluates PPG-based BP estimation specifically within fluctuation intervals and examines fluctuation-aware balanced sampling and event-aligned calibration.The calibration-based analysis tests whether aligning calibration timing with BP fluctuation events improves accuracy beyond fixed periodic schedules.
  • Method: BP fluctuations are defined using PELT-based time-series change point detection rather than heuristic rules such as ΔBP > 10mmHg.PELT identifies distributional transitions in continuous BP signals to provide a data-driven definition of fluctuation intervals.
  • Contributions: Targeted re-calibration at BP change points significantly improves estimation accuracy over periodic re-calibration alone, while online PPG-driven detection enables real-time triggering.The framework is presented as architecture-agnostic and supports adaptive calibration during continuous monitoring.

2. Related Work

Prior PPG-based BP estimation has progressed from handcrafted PWA/PTT features and conventional machine learning to deep learning and subject-specific calibration. However, existing BP-change studies use fixed thresholds, while real-time re-calibration requires online CPD that can operate from PPG alone without reference BP.

  • PPG-to-BP estimation: Calibration-free PPG-to-BP estimation evolved from handcrafted PWA/PTT features with conventional machine learning to deep architectures learning representations directly from raw PPG.Deep learning models have demonstrated competitive performance on benchmark datasets.
  • Calibration-based estimation: Calibration-based approaches use subject-specific references and periodic cuff measurements to personalize predictions and compensate for inter-individual variability.These methods improve estimation accuracy in real-world deployment settings but commonly rely on fixed periodic schedules.
  • BP change detection: Prior BP-change studies primarily track relative fluctuations rather than absolute BP, defining change when variation over a time window exceeds a fixed threshold such as ∆BP > 10mmHg.Explicit change point detection has rarely been explored for BP estimation.
  • Change Point Detection: Change Point Detection partitions time series into statistically homogeneous segments by identifying abrupt shifts in the underlying distribution, with offline and online settings.Offline CPD uses the entire sequence, whereas online CPD detects changes in real time from historical data.
  • Deployment challenge: Real-time BP re-calibration requires online CPD, but deployment lacks reference BP, so the detector must identify distributional shifts using PPG alone.Existing online BP change point methods rely on ground-truth BP measurements as input and are therefore inapplicable in deployment.

3. Problem Formulation

The paper formulates PPG-based blood pressure estimation as regression and evaluates it around distributional change points and clinically unstable intervals. It distinguishes calibration strategies and motivates targeted re-calibration triggered by physiological shifts.

  • 3.1. Calibration-Free PPG to BP Estimation: Standard PPG-to-BP estimation predicts systolic and diastolic blood pressure from fixed-length normalized PPG segments using a regression model.The model maps input x to prediction ŷ = f(x; θ).
  • 3.1. Calibration-Free PPG to BP Estimation: Model parameters are learned by minimizing prediction discrepancy from ground-truth blood pressure, typically with Mean Squared Error loss.The optimal parameters are denoted θ∗.
  • 3.2. Blood Pressure Fluctuations and Change Points: Aggregating errors over entire recordings can hide clinically unacceptable estimation failures during rapid blood pressure transitions.Low aggregated MAE may coexist with substantial errors during critical hemodynamic shifts.
  • 3.2. Blood Pressure Fluctuations and Change Points: A BP change point is defined as a time step where the underlying BP distribution shifts, with ground-truth points derived using PELT on the reference BP sequence.The definition requires p(y_t | t < τ) ≠ p(y_t | t ≥ τ).
  • 3.2. Blood Pressure Fluctuations and Change Points: Unstable BP intervals contain severe hypertension or hypotension, defined by SBP > 180 mmHg or mean arterial pressure < 65 mmHg.The protocol evaluates accuracy on all samples and separately within unstable intervals.
  • 3.3. Calibration-Based PPG to BP Estimation: Calibration-based estimation incorporates subject-specific reference data Dcal acquired before inference to address inter-subject physiological variability.The target is estimated from the current PPG input and calibration set.
  • 3.3. Calibration-Based PPG to BP Estimation: Input-level calibration feeds calibration samples into a fixed-parameter model, whereas parameter-level calibration updates θ to a personalized θ′ during online adaptation.Input-level calibration uses f(x_t, x_cal, y_cal; θ), while parameter-level calibration uses f(x_t; θ′).
  • 3.4. Targeted Re-Calibration: Targeted re-calibration supplements periodic updates with re-calibration triggered by detected BP change points, enabling timely adaptation to physiological shifts.Because BP-based detection requires continuous BP monitoring, the paper proposes online detection using only PPG signals.

4. Targeted Re-Calibration with Online Blood Pressure Change Point Detection

The framework detects BP change points online from PPG alone and uses detected state transitions to trigger targeted re-calibration. Contrastive feature learning and refined sequential clustering separate physiological states while reducing false positives and delayed responses.

  • Online Change Point Detection: PPG-only online change point detection combines contrastive feature learning with sequential clustering to trigger immediate re-calibration.The approach avoids explicit BP tracking by learning discriminative PPG features and predicting change points in real time.
  • PPG Representation Learning: Supervised contrastive learning clusters PPG samples within the same BP segment and separates samples from different segments in feature space.The encoder maps each PPG signal to a normalized embedding and uses reference BP change points to define positive samples.
  • Sequential Clustering: Sequential clustering tracks an active physiological-state centroid, updating it when cosine similarity remains above threshold η and transitioning or creating a state when similarity falls below η.A transition to an existing or newly initialized centroid flags a change point and triggers re-calibration.
  • Robustness Refinements: Fixed-threshold clustering can produce false positives from outliers or miss subtle shifts when the threshold is raised.PPG features may also remain unstable immediately after a genuine state change.
  • Robustness Refinements: Duration-dependent thresholding starts the threshold low after transitions and increases it toward ηbase as the current segment duration grows.The variant is η = ηbase ·(1−exp(−λn)), where n is elapsed frames and λ controls the rate of increase.
  • Robustness Refinements: Delayed prediction confirms a change point only when a deviation persists for K consecutive frames, buffering features instead of reacting to short-term outliers.The buffer is reset when no deviation is detected, and confirmed buffered features update the clusters.

5. Cohort

Experiments use PulseDB, which combines ECG, PPG, and ABP waveforms from MIMIC-III and VitalDB. The study uses quality-controlled, subject-independent splits and evaluates multiple preprocessing, calibration, and BP change point configurations.

  • Dataset: PulseDB integrates ECG, PPG, and ABP signals from the MIMIC-III and VitalDB datasets for cuff-less BP estimation experiments.The dataset is publicly available and serves as a large-scale benchmark.
  • Dataset: 10-second segments sampled at 125 Hz are filtered by cardiac-cycle completeness, PPG skewness, and PPG–ABP correlation quality criteria.Segments with incomplete cardiac cycles, negative PPG skewness, or PPG–ABP correlation below 0.9 are excluded.
  • Dataset: Subject-independent MIMIC-III and VitalDB training and test splits summarize subjects, cases, segments, unstable-BP intervals, and SBP/DBP distributions.Ground-truth SBP and DBP values are derived directly from ABP waveforms, and retained segments have irregular temporal spacing across subjects.
  • Preprocessing: PPG signals undergo band-pass filtering and z-normalization, with resampling to 50 Hz only for Pulse-PPG pretraining experiments; other settings retain 125 Hz.The filtering suppresses baseline drift and high-frequency noise.
  • Evaluation setup: The evaluation includes four calibration-free baselines, two calibration-based methods, and multiple BP change point definitions with periodic calibration fixed at 120 minutes.Calibration-free models are fine-tuned using the same protocol, while PPG2BP-Net and TTC represent input-level and parameter-level calibration, respectively.

6. Results on Real Data

On real data, BP estimation errors increase during unstable intervals, while interval-balanced training improves robustness. Change point–aware re-calibration, including online CPD triggers, improves calibration efficiency and accuracy over periodic or random strategies.

  • Fluctuation-Aware Evaluation: Unstable intervals increase BP estimation error by approximately 2–4 mmHg MAE for SBP and 1–2 mmHg for DBP across models and training strategies.Whole-recording evaluation therefore underestimates errors during clinically important fluctuation periods.
  • Fluctuation-Aware Evaluation: Interval-balanced training reduces unstable-region SBP error while preserving comparable whole-interval performance, including Inception1D from 17.04 to 16.40 mmHg and Pulse-PPG from 17.84 to 16.41 mmHg.The results support balancing training samples across stable and unstable intervals.
  • Offline Re-Calibration: PELT-based re-calibration achieves the lowest SBP and DBP MAE across TTC and PPG2BP-Net under a fixed 120-minute base calibration interval.Targeted calibration points substantially improve performance compared with periodic calibration alone.
  • Offline Re-Calibration: PELT uses 9.82 calibration points per case, outperforming the ΔBP strategy despite its 17.36 points per case and random re-calibration despite comparable calibration-point counts.These findings indicate that calibration timing, rather than frequency alone, drives the improvement.
  • Online Change Point Detection and Re-Calibration: Online CPD supports targeted re-calibration: Inception1D reaches 0.6457 weighted accuracy and 0.6149 IoU, while Online CPD-based re-calibration outperforms periodic and Random strategies.The online method achieves comparable or better accuracy using a similar number of calibration points per case.

7. Discussion

The discussion shows that fluctuation-aware evaluation exposes BP estimation failures hidden by aggregate metrics, while change point-triggered re-calibration improves robustness. It also identifies physiological and deployment limitations that require further validation.

  • Summary of findings: Fluctuation-aware evaluation reveals calibration-free model degradation averaging 2–4 mmHg higher MAE for SBP and 1–2 mmHg higher MAE for DBP.These failures would otherwise remain hidden under conventional whole-recording metrics.
  • Summary of findings: Fluctuation-stratified reporting should complement whole-interval evaluation, especially for patients with highly dynamic BP profiles.The discussion presents this reporting strategy as important for assessing model suitability in dynamic BP conditions.
  • Why change point timing helps: PELT detects joint SBP–DBP distributional shifts, whereas the ΔBP rule reacts to transient excursions and noise and produces redundant change points.The compared triggers therefore identify BP fluctuations using different assumptions and signals.
  • Linking detection quality to re-calibration gain: Raising AUROC from 0.5332 to 0.7275 lowers SBP MAE from 8.31 to 7.27 mmHg under the same PPG2BP-Neural setting.The discussion links detection quality directly to downstream re-calibration accuracy rather than merely to additional update triggers.
  • What transitions the detector captures: 85.4% of unstable segments occur under vasopressor administration, indicating that detected change points concentrate on pharmacologically driven hemodynamic shifts.These events are also described as periods when stale calibration references are most likely to limit performance.
  • Limitations: The framework’s generalizability, unsupervised deployment, penalty tuning, and fully cuffless adaptation remain limited by its data, training, fixed hyperparameter, and cuff-measurement requirements.PulseDB comes from ICU and operating-room settings; the online module uses ground-truth change points, PELT uses a fixed penalty, and detected change points still require occasional cuff measurements.

Appendix A. Further Implementation Details … A.3. Pseudo-Code for Online Change Point Detection via Sequential Clustering

Appendix A details the implementation of ground-truth BP change point detection, PPG preprocessing and encoder training, and online sequential-clustering detection. Ground-truth change points use PELT, while online detection uses persistent centroid deviations under duration-dependent thresholds.

  • A.1. Blood Pressure Change Point Detection with PELT: PELT detects ground-truth BP change points from reference SBP and DBP sequences.The implementation uses the PELT version provided by the ruptures library.
  • A.1. Blood Pressure Change Point Detection with PELT: The PELT implementation used in practice is provided by the ruptures library.
  • A.1. Blood Pressure Change Point Detection with PELT: Table 6 summarizes the hyperparameter settings for ground-truth BP change point detection using PELT.
  • A.2. Calibration-Free PPG to BP Estimation and Online Change Point Detection: Table 7 summarizes hyperparameters for PPG preprocessing, encoder training, and sequential clustering.
  • A.2. Calibration-Free PPG to BP Estimation and Online Change Point Detection: The encoder-training configuration is shared by calibration-free PPG-to-BP estimation and supervised contrastive learning for CPD.
  • A.2. Calibration-Free PPG to BP Estimation and Online Change Point Detection: Encoder backbones include Inception1D, PaPaGei-S, Pulse-PPG, and AnyPPG, using their original architecture implementations.
  • A.3. Pseudo-Code for Online Change Point Detection via Sequential Clustering: Online change points are triggered by persistent deviations from the current cluster centroid under a duration-dependent similarity threshold.

Appendix B. Effect of Periodic Calibration Frequency

The ablation varies periodic calibration intervals from 120 to 10 minutes with PELT fixed, showing that shorter intervals improve accuracy while change point triggering adds complementary gains. These results indicate that effective re-calibration depends on timing as well as frequency.

  • Calibration frequency: The ablation fixes PELT and varies periodic calibration intervals of 120, 60, 30, and 10 minutes, reporting SBP and DBP mean absolute error.The resulting errors are reported in Figure 4.
  • Calibration frequency: Shorter periodic calibration intervals consistently improve estimation accuracy, suggesting that frequent re-calibration mitigates physiological-drift degradation.
  • Change point triggering: PELT-based change point triggering further improves performance across all periodic calibration intervals, including 30- and 10-minute intervals.The gains remain evident even when periodic calibration is already frequent.
  • Calibration timing: Re-calibration effectiveness depends on both how often and when it is applied, reinforcing fluctuation-aware timing strategies.

Appendix C. Qualitative Results

Qualitative examples show that periodic calibration develops persistent bias when patient hemodynamics shift, whereas PELT-based re-calibration detects transitions and maintains closer agreement with ground-truth BP.

  • Qualitative Results: Periodic calibration exhibits persistent bias during sustained hemodynamic shifts, failing to adapt to changing BP over the session.Figures 5–7 compare the Static baseline and PELT-based re-calibration against ground-truth BP.
  • Qualitative Results: PELT detects BP transitions through change point detection and re-calibrates accordingly, maintaining closer agreement with ground-truth BP throughout each session.The comparison is illustrated in representative examples in Figures 5–7.

Appendix D. Sensitivity Analysis of PELT Hyperparameters · Appendix E. Hyperparameter Sensitivity of the Online CPD Module · Appendix F. Further Evaluations

The appendices analyze sensitivity in PELT-based segmentation and the online CPD module, emphasizing meaningful fluctuation detection and the trade-off between accuracy and interval localization. No passage describing Appendix F. Further Evaluations was supplied.

  • Appendix D. Sensitivity Analysis of PELT Hyperparameters: PELT hyperparameters were selected by greedy search using Mean|∆Seg|, the average blood pressure change within detected segments.The criterion favors segmentations that capture meaningful BP fluctuations rather than noise-induced oversegmentation.
  • Appendix D. Sensitivity Analysis of PELT Hyperparameters: Selected default PELT hyperparameter values were marked with an asterisk in the sensitivity table.
  • Appendix D. Sensitivity Analysis of PELT Hyperparameters: The selected default PELT configuration provided the best overall trade-off across sensitivity analyses.
  • Appendix D. Sensitivity Analysis of PELT Hyperparameters: The cost model had the largest effect on PELT segmentation behavior.l1 and l2 produced more detected points per case but smaller segment-level BP changes, indicating sensitivity to noise rather than meaningful BP transitions.
  • Appendix E. Hyperparameter Sensitivity of the Online CPD Module: Online CPD hyperparameters were selected using intersection-overunion (IoU), reflecting overlap between predicted unstable intervals and reference intervals.
  • Appendix E. Hyperparameter Sensitivity of the Online CPD Module: Online CPD sensitivity results showed a trade-off between classification accuracy and interval-level localization quality.For ηbase, decreasing the value increased accuracy but reduced IoU.

F.1. BP Estimation Performance on Unstable BP Intervals with Targeted Re-Calibration … Appendix G. Clinical Interpretation of Detected Change Points

Targeted re-calibration, particularly PELT- or Online-CPD-triggered, reduces PPG2BP-Net’s error during unstable BP intervals and limits degradation relative to periodic calibration. Online CPD operates with low latency, while detected unstable segments largely coincide with vasopressor administration and clinically critical BP boundaries.

  • F.1. BP Estimation Performance on Unstable BP Intervals with Targeted Re-Calibration: PELT achieves 6.73 mmHg SBP MAE in unstable intervals, outperforming periodic calibration, ΔBP, and Random strategies.Periodic calibration yields 10.93 mmHg, while ΔBP and Random yield 9.46 and 9.10 mmHg, respectively.
  • F.1. BP Estimation Performance on Unstable BP Intervals with Targeted Re-Calibration: PELT has the smallest whole-to-unstable-interval SBP performance gap at 0.89 mmHg, compared with 1.41–1.66 mmHg for the remaining strategies.This indicates that PELT-triggered re-calibration better retains its benefit in difficult unstable segments.
  • F.2. Classification Baseline for CPD: A naive binary classifier is trained on 10-second PPG segments to predict PELT-derived BP change-point labels because no directly comparable PPG-only online CPD baseline exists.The classifier uses the described encoder configuration and training protocol.
  • F.3. Online CPD-Triggered Re-Calibration on Unstable BP Intervals: Online CPD achieves lower unstable-interval error than periodic, Classification, and Random triggers, while all triggered strategies degrade by only 0.42–0.49 mmHg versus 1.60 mmHg for periodic calibration.Periodic calibration yields 10.93 mmHg SBP MAE; Classification and Random yield 8.73 and 8.06 mmHg, respectively.
  • F.3. Online CPD-Triggered Re-Calibration on Unstable BP Intervals: The Classification trigger issues 12.31 points per case yet performs worse than the other triggered strategies, indicating that inaccurate localization is not offset by higher frequency.Both Classification and Online CPD use Inception1D as the CPD backbone.
  • F.4. Computational Cost: Online CPD requires one encoder pass per incoming 10-second PPG segment, with 9.2–31.0 ms forward-pass latency and sequential clustering below 0.05 ms.Cosine comparisons against stored centroids require O(K|C|d) operations.
  • Appendix G. Clinical Interpretation of Detected Change Points: 6,094 unstable segments (85.4%) occur during vasopressor administration, compared with 1,038 segments (14.6%) without vasopressors.The passage uses this distribution to characterize the physiological events underlying PELT-detected change points.
  • Appendix G. Clinical Interpretation of Detected Change Points: Unstable segments lie near SBP > 180 mmHg and MAP < 65 mmHg boundaries, where calibration-based models retain approximately 11–14 mmHg SBP error.The passage links this residual error to potentially delayed recognition of hypertensive emergencies.
Loading 2608.18639v1…