PaperlayerView exact revision on arXiv

Source-linked AI summary

SIFPBPNet: A Dual-Path Network for Wearable and Cuffless Blood Pressure Estimation via Individualized Steady-state Representation

Shuailong Tang, Xiaoyu Li, Donglin Xie, Wei Chen, Guangpu Zhu, Yelei Li, Yali Zheng

arXiv:2609.12690v1cs.LG

TL;DR

Population-based PPG models face individual heterogeneity, including similar waveforms associated with different BP levels. SIFPBPNet combines historical individual steady-state features with current-segment estimation, reducing SBP MAE by 8.5% versus SEM-ResNet and DBP MAE by 4.0% versus VGGNet.

  • Problem

    Population-based PPG models can regress toward the mean because similar waveforms may correspond to different BP levels across individuals.

    Source: §I · p. 1

  • Method

    SIFPBPNet pairs an SFP that represents individual steady-state information from historical PPG with an IFP that estimates BP from current PPG using that prior.

    Source: §I · p. 1

  • Results

    SIFPBPNet reduced SBP MAE by 8.5% versus SEM-ResNet and DBP MAE by 4.0% versus VGGNet, the strongest baselines for those metrics.

    Source: A. BP Estimation Performance · p. 5

  • Takeaways & Limitations

    The SFP module showed transferability across various deep-learning backbones, supporting its use as an individual-prior component.

    Source: §VI · p. 6

    Longitudinal validation of personalized baseline tracking remains future work.

    Source: §VI · p. 7

Abstract

from arXiv · show

Continuous and cuffless blood pressure (BP) monitoring using photoplethysmography (PPG) is of great interest for low-cost and personalized cardiovascular health management. However, significant population heterogeneity and the "one-to-many mapping" problem, where similar waveforms across individuals correspond to different BP levels, limit the accuracy of conventional population-based models. To address this challenge, we propose a dual-path architecture termed SIFPBPNet, which separately represents steady-state and instantaneous features, through a Steady-state Feature Path (SFP) and an Instantaneous Feature Path (IFP). The SFP employs a Graph Attention Network (GAT) to extract individual-specific and long-term characteristics from multi-day historical PPG trajectories. In parallel, the IFP captures short-term dynamics from current PPG segments and incorporates the steady-state prior via a cross-attention mechanism. Experiments on a large-scale wearable dataset demonstrate that SIFPBPNet achieves a Mean Absolute Error (MAE) of 8.57 and 5.97 mmHg for systolic and diastolic BP, respectively, outperforming state-of-the-art models. Furthermore, the SFP module consistently improves performance when integrated into various backbone architectures, yielding 2.8-13.1% relative MAE reductions for systolic BP. These results highlight the strong generalizability and plug-and-play transferability of the SFP module, underscoring its great potential for accurate cuffless BP monitoring.

I. INTRODUCTION

Cuff-based measurement is intermittent, while population-based PPG models struggle with individual variation and similar waveforms linked to different BP levels. SIFPBPNet responds with an individual-aware framework; earlier work spans physiological indicators and deep-learning representations.

I. INTRODUCTION

  • Hypertension affects more than one billion adults and accounts for approximately 19% of global mortality, motivating improved BP management.The passage attributes these figures to the WHO’s 2023 Global Report on Hypertension.Source: §I · p. 1
  • Intermittent cuff measurements may miss white-coat, masked, or nocturnal hypertension, while repeated inflation causes discomfort and limits adherence.These constraints motivate unobtrusive approaches for precise risk management.Source: §I · p. 1
  • Population-based models can regress toward the mean because similar PPG waveforms may correspond to different BP levels across individuals.The cited work reports this pattern in up to 33% of PPG segments.Source: §I · p. 1
  • SIFPBPNet pairs a steady-state path using historical PPG trajectories with an instantaneous path that incorporates the individual prior for BP regression.The proposed framework shifts from pure instantaneous mapping toward individual-prior estimation.Source: §I · p. 1

A. Physiological-driven Stage

  • Physiological-driven approaches used indicators such as pulse transit time and PPG morphology, with later methods combining morphological features and PTT.The passage describes machine-learning approaches including random forests and linear regression.Source: A. Physiological-driven Stage · p. 2

B. Deep Learning-driven Stage

  • Deep-learning methods learned signal representations, but estimation bias across demographics and the one-to-many mapping problem remained.Examples include VGGNet and PMFL with SEM-ResNet; similar waveforms can map to different BP levels in up to 33% of cases.Source: B. Deep Learning-driven Stage · p. 2

C. Personalized Modeling Stage

Prior personalized approaches use demographic attributes, BP-interval subspaces, or disease labels. SIFPBPNet adds a pathway that learns a dynamic individual prior from long-term PPG and combines it with current-segment features for personalized estimation.

C. Personalized Modeling Stage

  • Earlier personalized methods used demographic attention queries, BP-interval classification, or disease-label pre-stratification.The cited examples are Zheng et al., Ma et al., and Zhang et al..Source: C. Personalized Modeling Stage · p. 2
  • SIFPBPNet learns individualized steady-state representations from 10-day historical PPG trajectories, unlike approaches relying on static demographic attributes.The authors describe this as a dynamic personalized prior for heterogeneous populations.Source: C. Personalized Modeling Stage · p. 2

A. Model Overview

  • The dual-path architecture separates short-term instantaneous features from long-term steady-state features to support personalized BP estimation.The overview assigns current-record encoding to IFP and multi-day historical PPG aggregation to SFP.Source: A. Model Overview · p. 2
  • IFP combines current PPG features with morphological and demographic features, then incorporates SFP’s individual long-term information for BP regression.The overview describes gated fusion as providing personalized constraints for instantaneous regression.Source: A. Model Overview · p. 2
  • SFP uses a GNN and BP-related auxiliary tasks, is pre-trained on historical data, and is fixed before IFP training.The auxiliary tasks include age regression, individual-level hypertension detection, and high BP variability detection.Source: A. Model Overview · p. 2

B. PPG Encoder

The shared PPG encoder processes raw signals and their derivatives for both SFP and IFP. It combines multi-scale convolution, attention-based refinement, and a Transformer to produce segment-level representations.

B. PPG Encoder

  • The shared encoder takes a three-channel input comprising raw PPG and its first and second derivatives.It serves both SFP and IFP.Source: B. PPG Encoder · p. 2
  • The encoder combines an MRF-CNN for multi-scale local features with a Transformer for long-range dependencies.The resulting representations are used by both paths.Source: B. PPG Encoder · p. 2
  • Parallel convolutions use kernel sizes 3, 5, and 7; each output has C_k channels and length L′.The passage describes these convolutions as capturing features at different scales.Source: B. PPG Encoder · p. 2
  • Concatenated multi-scale features are refined by CBAM, whose channel and spatial attention highlight important features and suppress noise.Source: B. PPG Encoder · p. 2
  • Absolute PPG-segment timing is encoded with sine and cosine positional embeddings and added to enhanced features for Transformer input.Source: B. PPG Encoder · p. 2
  • Global average pooling of the transposed Transformer representations yields a feature vector for each PPG segment.Fig. 1 identifies the encoder as extracting latent features for both SFP and IFP.Source: B. PPG Encoder · p. 2, Figure 1 · p. 3

C. SFP

The SFP turns a user’s historical PPG records into an individualized steady-state representation. A GAT handles varying record counts, and task-guided pooling produces a prior for BP regression.

C. SFP

  • The GAT adaptively aggregates a variable number of historical-record nodes, letting the SFP handle different amounts of data while learning BP-related signatures.Its auxiliary tasks include age regression, individual-level hypertension detection, and high BP variability detection.Source: C. SFP · p. 3
  • Each graph node represents an encoded historical PPG segment, and the encoded features of an individual’s N records form the GAT input matrix.The matrix has N rows and Cfeat columns.Source: C. SFP · p. 3
  • GATv2 computes attention weights between neighboring nodes, then concatenates the attention-head features to update each node representation.The updated representation is defined for node i.Source: C. SFP · pp. 3–4
  • Task-guided pooling combines each node feature with a task embedding to weight nodes, then sums the weighted features and processes them into the steady-state prior.The resulting individualized feature is used in subsequent BP regression.Source: C. SFP · p. 4

D. IFP and Fusion Module

The IFP estimates BP from a current PPG segment while incorporating an individualized steady-state prior. It fuses signal, demographic, and morphological features before regression.

D. IFP and Fusion Module

  • The IFP encodes the current PPG segment and uses gated fusion to combine its instantaneous representation with demographic and morphological features.It then incorporates the SFP’s individualized steady-state prior through cross-attention.Source: D. IFP and Fusion Module · p. 4
  • Cross-attention uses the instantaneous representation as Query and the steady-state feature as Key and Value.For multi-task SFP pre-training, the steady-state feature stacks vectors from T auxiliary tasks.Source: D. IFP and Fusion Module · p. 4
  • Learnable projections map the features into a common dimension, and the resulting fused representation enters the regression head for BP estimation.The projections are WQ, WK, and WV; their common dimension is dk.Source: D. IFP and Fusion Module · p. 4

E. Training Setup

Training separates SFP pre-training from IFP training using different individuals. The SFP’s auxiliary tasks target age, hypertension, and BP variability as complementary steady-state cues.

E. Training Setup

  • SFP pre-training uses historical data from 40% of individuals and three tasks: age regression, individual-level hypertension detection, and high BPV detection.An individual is labeled high BPV when historical SBP or DBP standard deviation exceeds 10 or 7 mmHg, respectively.Source: E. Training Setup · p. 4
  • The auxiliary tasks are motivated as complementary priors: age reflects long-term vascular structural changes, while hypertension and BPV status characterize functional BP dynamics.The passage relates age regression to vascular aging patterns and PPG morphology to arterial stiffness and age-related vascular changes.Source: E. Training Setup · p. 4
  • The remaining 60% of individuals are used for IFP training and testing, with a 4:1:1 individual-level split that keeps each person within one set.This separates individuals between the SFP and IFP training stages as well.Source: E. Training Setup · p. 4

A. Dataset

The study uses the OPPO HBPM wearable BP dataset, with historical PPG records preceding prediction data. The prediction and historical subsets are described separately.

A. Dataset

  • 10 consecutive days of each person’s PPG data, with at least two daily periods, serve as historical input; later records are used for prediction.The source dataset records synchronized BP and PPG in morning, noon, and evening periods over approximately one to two months.Source: A. Dataset · p. 4
  • 790 individuals are represented, with mean age 38.53 ± 9.93 years; signals were recorded for 60 seconds at 250 Hz alongside cuff-based BP measurements.The sample comprises 46.33% males and 53.67% females; ground-truth BP was measured using an Omron oscillometric monitor.Source: A. Dataset · p. 4
  • After low-quality signals were excluded, the prediction subset contained 98,258 samples and the historical subset 20,342; Fig. 2 depicts BP distribution and historical-record counts.Table I presents statistics for the prediction and historical subsets.Source: A. Dataset · p. 5, Figure 2 · p. 5, Table I · p. 5

B. Preprocessing

The section describes filtering and quality screening of PPG signals, then reports SIFPBPNet’s OPPO HBPM performance and comparison with standalone deep-learning baselines.

B. Preprocessing

  • A fourth-order Butterworth filter (0.5–12 Hz) and template-matching quality detection selected 12-second, 3,000-point high-quality PPG segments.The filtering was applied to remove baseline drift and high-frequency noise.Source: B. Preprocessing · p. 5

A. BP Estimation Performance

  • SIFPBPNet achieved SBP and DBP MAEs of 8.57 and 5.98 mmHg, respectively, on the OPPO HBPM dataset.For SBP, ME ± SDE was 0.02 ± 10.76 mmHg; for DBP, it was −0.10 ± 7.70 mmHg.Source: A. BP Estimation Performance · p. 5, Table II · p. 5
  • SIFPBPNet reduced MAE by 8.5% for SBP and 4.0% for DBP versus the strongest standalone baselines.The strongest baselines were SEM-ResNet for SBP and VGGNet for DBP; SIFPBPNet also performed better across all reported evaluation metrics.Source: A. BP Estimation Performance · p. 5, Table III · p. 5

B. Ablation Study

The SFP improved performance across tested backbones, with ablations and pre-training comparisons examining its contribution and the auxiliary tasks used to train it.

B. Ablation Study

  • Adding SFP lowered SBP MAE from 9.07 to 8.57 mmHg; CBAM and Transformer integration produced incremental improvements beforehand.The authors associate the earlier gains with capturing local spatial features and long-range temporal dependencies.Source: B. Ablation Study · p. 5, Table IV · p. 6

C. Results of SPF Pre-training Task Configurations

  • Hypertension detection alone yielded SBP and DBP MAEs of 8.61 and 6.05 mmHg; adding age regression gave 8.57 and 5.98 mmHg, respectively.The combined tasks produced the best reported DBP result while maintaining competitive SBP performance.Source: C. Results of SPF Pre-training Task… · p. 6, Table V · p. 6

D. Transferability of SFP Module

VI. DISCUSSION & CONCLUSION

On a large-scale, real-world wearable dataset, incorporating steady-state features substantially enhanced BP estimation and markedly reduced SBP MAE versus the strongest baseline. The authors conclude that SIFPBPNet is an effective, scalable framework while identifying longitudinal validation, pretraining, and window-length evaluation as future work.

VI. DISCUSSION & CONCLUSION

  • SFP incorporation substantially enhances estimation performance and markedly reduces SBP MAE versus the strongest baseline on a large-scale, real-world wearable dataset.The passage reports no specific error values for this comparison.Source: §VI · p. 6
  • The SFP transfers across various deep-learning backbones, and the authors identify individual-level representation learning as critical to reliable, personalized BP estimation.Source: §VI · p. 6
  • SIFPBPNet is presented as an effective, scalable framework for accurate daily BP measurement in heterogeneous real-world wearable scenarios.Source: §VI · pp. 6–7
  • Future work will validate longitudinal baseline tracking, explore auxiliary and self-supervised or contrastive SFP pretraining to reduce labeled-data reliance, and vary PPG window lengths to balance context and deployment constraints.Source: §VI · p. 7
Loading 2609.12690v1…