Personalized and Explainable Blood Pressure Estimation from PPG via Hybrid CNN--Morphological Features

Myung-Kyu Yi; Jongshill Lee; Jeyeon Lee; In Young Kim

cs.CVAug 11, 2026

Personalized and Explainable Blood Pressure Estimation from PPG via Hybrid CNN--Morphological Features

Authors: Myung-Kyu Yi, Jongshill Lee, Jeyeon Lee, In Young Kim

Organizations: Department of Biomedical Engineering, Hanyang University,, Seoul 04763, South Korea

Abstract

Continuous cuffless blood pressure (BP) monitoring using photoplethysmography (PPG) offers a promising solution for personalized healthcare. However, existing methods have two major limitations. Handcrafted feature-based approaches rely on precise fiducial point detection and are limited to short-term analysis, while deep learning models, despite their accuracy, often operate as black boxes with limited physiological interpretability. To address these challenges, we propose a physiology-guided hybrid framework for personalized BP estimation that couples a convolutional neural network (CNN) branch capturing global and local waveform dynamics with a morphology-prior branch that explicitly encodes person-specific vascular characteristics. By embedding a morphology-based feature set that explicitly encodes individual vascular characteristics, the proposed framework enhances personalization and reduces dependence on large-scale training datasets. Evaluated on a subset of the MIMIC-III database under a subject-specific (personalized) testing protocol, the proposed personalized physiology-guided hybrid approach achieved mean absolute errors (MAEs) of 3.77 mmHg for systolic BP and 2.36 mmHg for diastolic BP, corresponding to relative improvements of 43.7% and 32.4% over a subject-specific (personalized) CNN-only baseline. SHAP-based analysis confirmed that the introduced morphology-prior features align with individual vascular characteristics, reinforcing per-subject interpretability. These findings highlight the potential of personalized, physiology-guided hybrid learning with novel morphological descriptors for accurate and explainable BP monitoring in real-world settings.

Figures & tables

Comparison Criteria	Conventional Morphological Features	Proposed Hemodynamic Features
Domain	Geometric Signal Description	Vascular and Hemodynamic Properties
Abstraction Level	Descriptive: Quantifies the visible shape of the PPG waveform (width, amplitude, slope).	Physiology-oriented: Encodes vascular properties such as stiffness, compliance, and damping through energy- and decay-related indices.
Mathematical Basis	Point-wise: Heavily relies on fiducial points (e.g., peak, valley), making it sensitive to noise and localization errors.	Area/Rate-based: Uses region-level integrals and decay rates over systolic/diastolic phases, improving robustness to fiducial errors and high-frequency artifacts.
Link to BP	Empirical Association: Uses waveform shape changes that correlate with BP but are only weakly tied to specific regulatory mechanisms.	Physiology-guided Association: Features are constructed to reflect determinants of BP—such as arterial stiffness, vascular compliance, and peripheral resistance—rather than geometry alone.

Table 1: Comparison between conventional morphological features and the proposed hemodynamic features. The proposed features are designed to complement geometric descriptors by incorporating simple hemodynamic considerations.

Figure 1: Schematic illustration of the proposed morphological features extracted from a single PPG cycle

Feature	Mathematical Meaning	Physiological Interpretation	Design Principle
SUA	Accumulated first derivative from onset to systolic peak	Proxy for stroke volume and arterial compliance	Hemodynamic Relevance
PNRS	Mean first derivative after dicrotic notch	Reflective wave slope, indicator of arterial stiffness	Hemodynamic Relevance
NPVA	Area difference before and after notch, normalized by total area	Balance of systolic and diastolic energy	Global Energy Representation
SDE	Sum of squared second-order derivatives	Signal roughness; reflects vascular tension or resistance	Differential Shape Encoding
TE	Shannon entropy of post-notch segment	Morphological irregularity in diastolic tail	Global Energy Representation
HET	Time to reach 50% of total energy	Symmetry of energy distribution; waveform temporal centroid	Global Energy Representation

Table 2: Proposed Features with Mathematical and Physiological Interpretations and Design Principles

Figure 2: Schematic of the proposed hybrid CNN–Prior model. The architecture comprises two parallel branches: (1) a CNN module that processes raw PPG waveforms to extract local morphological patterns, and (2) a prior branch that encodes handcrafted physiological features. These representations are fused and passed through a fully connected regression head for personalized SBP/DBP prediction.

Aspect	Fiducial-Based Handcrafted Features	End-to-End Deep Models	Proposed Hybrid CNN + Physiological Priors
Input Representation	Discrete waveform landmarks	Raw waveform samples	Raw PPG + priors
Physiological Interpretability	Moderate	Low	High
Robustness to Noise / Motion	Low	Moderate	High
Data Requirements	Low	High	Moderate
Computational Complexity	Low	High	Moderate

Table 3: Comparison of the proposed approach with two common alternatives for PPG-based BP estimation.

Parameter	Value
Optimizer	Adam
Learning rate	$1\times 10^{-3}$
Loss	MSE (joint SBP/DBP regression)
Batch size	32
Epochs	100
Dropout	0.5

Table 4: Training configuration for the CNN branch model.

Configuration	MAE SBP	MAE DBP	Params (K)	FLOPs (M)
CNN Only (Baseline)	6.70 $\pm$ 1.02	3.49 $\pm$ 0.65	60.0	3.12
CNN + Conventional Features	3.99 $\pm$ 0.48	2.78 $\pm$ 0.61	101.9	3.16
CNN + Proposed Features (Ours)	3.77 $\pm$ 0.50	2.36 $\pm$ 0.40	101.9	3.16

Table 6: Ablation study on the effect of incorporating conventional and proposed features into the CNN model for personalized BP estimation. The MAE is reported as mean

\pm

standard deviation over 5 folds, and the improvement over the personalized CNN-only baseline is statistically significant (

p<0.05

, paired t-test).

Configuration	MAE SBP	MAE DBP
CNN Only (Baseline)	9.14 $\pm$ 1.14	5.01 $\pm$ 1.03
CNN + Conventional Features	9.03 $\pm$ 1.02	4.97 $\pm$ 1.10
CNN + Proposed Features (Ours)	8.73 $\pm$ 1.15	4.52 $\pm$ 1.20

Table 7: Ablation study on the effect of incorporating conventional and proposed features into the CNN model for cross-subject BP estimation. The MAE is reported as mean

\pm

standard deviation over 5 folds, and the improvement over the cross-subject CNN-only baseline is statistically significant (

p<0.05

, paired t-test).

Figure 3: Bland–Altman analysis for SBP estimation showing a mean difference of $+0.06\,\mathrm{mmHg}$ (red solid line) and the 95% limits of agreement (LoA) of $-12$ to $+12\,\mathrm{mmHg}$ (red dotted lines), indicating acceptable agreement with minimal systematic bias. Note that the plot is generated on a per-segment basis; points at higher SBP levels may correspond to transiently elevated segments from otherwise normotensive participants.

Figure 4: Bland–Altman analysis for DBP estimation showing a mean difference of $+0.22\,\mathrm{mmHg}$ (red solid line) and the 95% limits of agreement (LoA) of $-9$ to $+9\,\mathrm{mmHg}$ (red dotted lines), indicating acceptable agreement with minimal systematic bias. Because the analysis is also segment-based, points at higher DBP levels may reflect transiently elevated segments within normotensive subjects.

Year	Study	Dataset	Personalization Approach	MAE (SBP / DBP)
2022	Leitner et al. Baseline [ 27 ]	MIMIC-III Subset	None	4.59 / 2.72
2022	Leitner et al. Transfer [ 27 ]	MIMIC-III Subset	Pretraining + Fine-tuning	3.52 / 2.20
2023	Fan et al. [ 28 ]	MIMIC-II + Real-world	Pretraining + Adapter	6.68 / 3.91
2022	Zhao et al. [ 33 ]	MIMIC-III Subset	Pretraining + Fine-tuning	8.44 / 8.26
2024	Wang et al. [ 34 ]	Real-world (IoT)	Personalized incremental learning	3.11 / 2.47
2024	Guo et al. [ 35 ]	Multi-domain	Meta-learning + Fine-tuning	7.46 / 4.51

Table 8: Comparison with representative personalized BP estimation methods.

Figure 5: Global SHAP summary plot. PNRS and SUA exhibit the largest contributions, followed by TE, while PSI and SDE have moderate effects. WCC, EAR, SRR, NPVA, and HET exhibited relatively low SHAP values, suggesting that their marginal contribution may stem from redundancy with dominant features (e.g., PNRS and SUA) or reduced variance across the study cohort. Despite their lower rankings, these features may still capture secondary physiological cues in broader populations or under stress conditions.

Figure 6: Mean absolute SHAP value ranking. PNRS and SUA dominate the model’s predictions, TE is moderately influential, and PSI and SDE provide secondary contributions.

Figure 7: Radar chart of normalized SHAP values. PNRS and SUA form the largest wedges, with TE, PSI, and SDE contributing moderately. Remaining features are near-zero.

Figure 8: SHAP waterfall for segment #1. High PNRS and TE drive positive shifts, while SUA slightly decreases prediction.

Figure 9: SHAP waterfall for segment #2. Negative TE dominates downward influence; PNRS provides a positive offset.

Figure 10: SHAP waterfall for segment #3. PNRS and TE again lead positive contributions, confirming global ranking trends.

Explore similar work

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
DMT: Demographic Conditioning, Morphology-Enhanced Transformer for Cuffless Blood Pressure Estimation from PPG Signals
Jun 9, 2026Yidan Shen, Neville Mathew, Maham Rahimi +4Photoplethysmography Wearable Health Monitoring
Dept. of Electrical and Computer Engineering University of Houston Houston, USA · Dept. of Engineering Technology University of Houston Houston, USA · Dept. of Cardiovascular Surgery Houston Methodist Hospital Houston, USA +1
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

Jul 29, 2026J/K move · Enter open · S save

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

Kindeep K. Dhatt, Tengyue Wu, Hanbang Hua, Yayun Du

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

Wearable Health Monitoring Cuffless Blood Pressure Estimation

Open paper AskPDF arXiv

Continuous cuffless blood pressure (BP) monitoring remains challenging due to motion artifacts, physiological variability, and the limited robustness of conventional pulse transit time (PTT) models under dynamic conditions. Many prior approaches rely on multi-second windows to stabilize estimation, an assumption that is frequently violated during real-world monitoring with intermittent signal corruption. Here, we show that discriminative BP-related information is preserved at the single-beat level and present a lightweight multi-modal wearable framework for continuous BP estimation. The system integrates synchronized chest electrocardiography (ECG) and ear-clip reflectance photoplethysmography, each co-located with a 6-axis inertial measurement unit to provide motion context. We introduce a hybrid learning architecture in which a one-dimensional convolutional neural network extracts a 64-dimensional embedding from individual PPG beats and fuses it with 30 physiology-grounded features, including PTT statistics and heart rate variability, followed by LightGBM regression. The method was evaluated using a multi-phase stress protocol (

n=10

) and the PulseDB public dataset with subject-disjoint validation. Across 30 independent runs, the model achieved mean absolute errors of

4.02 \pm 0.21

~mmHg for systolic BP and

1.79 \pm 0.05

~mmHg for diastolic BP, corresponding to a 28.2% reduction in combined MAE relative to baseline models. By enabling beat-wise estimation without long temporal context, this framework supports computationally efficient cuffless BP monitoring suitable for wearable deployment under practical resource constraints. The source code for this work is available at https://github.com/SYMBIOX-Lab/BP-wireless.

Abstract

Explore similar work

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

DMT: Demographic Conditioning, Morphology-Enhanced Transformer for Cuffless Blood Pressure Estimation from PPG Signals

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