Cross-View Attention Fusion Net: A Prior-Guided Dual-View Representation Learning for Cardiac Output Estimation from Short-Term PPG Signals

arXiv:2605.19666 · physics.med-ph, cs.LG · Submitted 2026-05-19 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Cross-View Attention Fusion Net".

Jane: The gist: CVAF-Net is a prior-guided, dual-view deep learning architecture designed for cardiac output estimation from short,

Tom: First, who's behind it and why it matters.

Title and authors: Tom: So we're looking at this paper now, it’s called "Cross-View Attention Fusion Net: A Prior-Guided Dual-View Representation Learning for Cardiac Output Estimation from Short-Term PPG Signals". It sounds pretty technical, but the goal is estimating cardiac output from short segments of photoplethysmography signals.

Jane: Right, it sounds complicated because it’s using two different ways to look at the data at the same time. The authors are Yaowen Zhanga, Bo Cuia, Libera Fresiellob, Peter H. Veltinka, Dirk W. Donkerb and Ying Wanga from the University of Twente and Utrecht hospitals.

Lu: I’m really interested in how they bridge feature-engineered machine learning with end-to-end deep learning methodologies. It’s about moving past just using raw PPG descriptors alone to something more comprehensive.

Meng: And the core idea seems to be that they are tackling the problem of accurate cardiac output estimation from short, fixed-length PPG segments without needing perfect pulse detection, which is a big deal for wearable tech because motion artifacts can mess up timing.

Lalam: It’s about taking raw PPG as one view and a feature sequence map or FSM as another view, and then using attention to let them talk to each other.

The paper's summary: Tom: So the paper explains that they are using a dual-view deep learning architecture called CVAF-Net. This model processes the raw PPG signal as a temporal view and also uses a feature sequence map, or FSM, as a structured view that has some prior guidance built into it.

Jane: That’s right, so it’s not just one way of looking at the data; they are capturing the dynamic sequence of the PPG and the structural features extracted from it simultaneously. This is meant to help them capture more complex interactions than either view could alone.

Lu: The temporal encoder uses a sliding window long short term memory with a multi-head self-attention mechanism to get those sequential dependencies, while the spatial encoder looks at an FSM derived from eight key features like first derivative and FFT amplitude to grab structural information.

Meng: So they’re building these two encoders separately, one for time and one for structure, before they bring them together in a fusion module. That sounds like a lot of data handling on the backend.

Lalam: The real innovation there is the cross-view attention fusion module, where the FSM representation acts as a query to select relevant temporal information from the raw PPG branch based on that structural context.

The paper's improvements: Tom: The authors highlight a few key improvements they made. First, they introduced this cross-view attention fusion mechanism which lets the FSM-derived structural representations query both the FSM and temporal representations under a given prior-guided feature context.

Jane: That means the system can learn to select specific dynamic features from the raw temporal data based on what it sees in that structural context, which helps when physiological states are complex. It’s about making the model more relevant dynamically.

Lu: They also have this dual-branch architecture that learns complementary representations from raw 1D PPG segments and that eight-channel FSM view for non-invasive cardiac output estimation <ref:2605.19666#pg3>. This lets them capture both the sequential dynamics and the multi-scale structural information at once.

Meng: From an engineering standpoint, it’s interesting because they claim this approach achieves a twelvefold reduction in FLOPs compared to leading Transformer-based models while still maintaining accuracy. That's a significant computational efficiency gain for wearable hardware.

Lalam: They also showed that the cross-view attention fusion proved better than simple concatenation methods when modeling the complex interdependencies between those two different views, which is a key architectural choice there.

Conclusion: Tom: So to wrap this up, CVAF-Net seems to offer a high-precision and computationally efficient solution for continuous wearable-based cardiac output monitoring. It does achieve a mean absolute error of zero point one nine L/min on simulated data and maintains high accuracy in real-world scenarios with a minimum MAE of one point two zero L/min.

Jane: Plus, the plausibility analysis showed that the estimated cardiac output aligns with physiological laws, correlating expectedly with age, heart rate and systemic vascular resistance. This suggests it’s capturing actual physiology rather than just fitting noise patterns.

Lu: The ablation study confirmed that removing the spatial encoder caused the most significant performance degradation across all tested datasets. It proved that the structural view is really important for this estimation task.

Meng: And from a practical standpoint, achieving that twelvefold reduction in FLOPs means it could actually be deployable on resource-constrained wearable hardware right now.

Lalam: For me, the use of fixed-length PPG segments is relevant because it avoids the need for precise pulse cycle detection, which is important for continuous monitoring.

Tom: So that’s CVAF-Net, a prior-guided dual-view deep learning architecture designed to estimate cardiac output from short-term PPG signals. It’s promising for wearable monitoring, though future work will focus on noise and uncertainty awareness.

Yaowen Zhanga, Bo Cuia, Libera Fresiellob, Peter H. Veltinka, Dirk W. Donker, Ying Wanga

Department of Biomedical Signals and Systems, University of Twente Department of Cardiovascular and Respiratory Physiology, University Medical Center Utrecht

physics.med-ph, cs.LG

Submitted: 2026-05-19

Updated: 2026-10-09

License: http://creativecommons.org/licenses/by-nc-nd/4.0/

Importance score: 65/100

The gist: The gist: CVAF-Net is a prior-guided, dual-view deep learning architecture designed for cardiac output estimation from short, fixed-length PPG segments by simultaneously processing raw PPG as a

Key concepts

Dual-View Architecture: The model uses two separate processing branches: one for the raw PPG time series (temporal view) and another for a structured feature map derived from key PPG features (spatial view). This allows the system to analyze the signal from both a sequential and a structural perspective simultaneously.
Feature Sequence Map (FSM): The FSM is a structured representation created by extracting eight specific features from the PPG signal, such as derivatives and FFT amplitudes. This map provides the model with rich, high-level structural information about the underlying physiological signal that goes beyond just raw data.
Cross-View Attention Fusion: This core mechanism integrates the temporal and spatial views. It uses the structured FSM view as a query to selectively extract relevant information from the raw temporal PPG branch, effectively learning how structural context influences time-series dynamics for better cardiac output estimation.

Terminology

Benchmark — 2× in this paper · explained in 2 episodes
Transformer — 2× in this paper · explained in 2 episodes
Accuracy — 1× in this paper · explained in 1 episode
Representation Learning — 1× in this paper · explained in 3 episodes

Summary

The gist: CVAF-Net is a prior-guided, dual-view deep learning architecture designed for cardiac output estimation from short, fixed-length PPG segments by simultaneously processing raw PPG as a temporal view and a feature sequence map (FSM) as a structured, prior-guided view.

Motivation and Architecture

The paper addresses the challenge of accurately estimating cardiac output (CO) from photoplethysmography (PPG) signals by bridging the gap between feature-engineered machine learning approaches and end-to-end deep learning methodologies The Cross-View Attention Fusion Network (CVAF-Net) is a prior-guided, dual-view deep learning architecture designed for CO estimation from short, fixed-length PPG segments CVAF-Net simultaneously processes raw PPG as a temporal view and a feature sequence map (FSM) as a structured, prior-guided view The model was designed to capture complex interactions between these two views by learning how the FSM-derived representation queries both FSM and temporal representations.

Dual-View Representation Learning

CVAF-Net employs a dual-branch architecture consisting of a temporal encoder and a spatial encoder The temporal encoder processes the PPG time-series data using a sliding window long short term memory (LSTM) augmented with a multi-head self-attention mechanism to capture dynamic and sequential dependencies The spatial encoder operates on a feature map derived from the PPG signal, utilizing a one-dimensional ResNet-18 followed by a multi-head self-attention layer to extract high-level spatial and structural features from the FSM. The FSM is constructed from eight key features extracted from the PPG signal, including first derivative, second derivative, Fast Fourier transform (FFT) full amplitude, Hilbert transform, discrete stable wavelet transform (DSWT) approximation coefficients (a), and detail coefficients (d1-d3).

Cross-View Attention Fusion

The core of the model is the cross-view attention fusion module, which integrates the temporal features and FSM-based features to model their complex interdependencies The fusion mechanism uses the FSM-based representation as a query to select relevant temporal information from the raw PPG branch under a given structural context. This is achieved by projecting features into a joint space and generating two attention maps, alpha intra for spatial self-relations and alpha cross for cross-modal dependencies. The attended values are aggregated using learnable scalars c1 and c2 to adaptively balance the influence of the structural view and the temporal view during training. The final fused representation is compressed into a segment-level feature vector Vf lat via global average pooling across all windows before being passed through a multi-layer perceptron (MLP) to compute the CO value.

Experimental Validation and Performance

CVAF-Net was validated independently using 5-second, 15-second, and 30-second segments across three datasets: simulated pulse waves (SPW), vasoconstriction provocation (VP), and resting/cycling activities (RC) The model consistently outperformed most of the benchmark machine learning and deep learning methods while achieving a mean absolute error (MAE) of 0.19 L/min on the simulated data. In real-world scenarios, it maintained high accuracy (minimum MAE: 1.20 L/min) Plausibility analysis confirmed that estimated CO aligns with physiological laws, correlating expectedly with age (rho = −0.274), heart rate (rho = 0.894), and systemic vascular resistance (rho = −0.740).

Conclusion and Implications

CVAF-Net offers a high-precision, computationally efficient, and generalizable solution for continuous wearable-based CO monitoring It achieves a twelvefold reduction in FLOPs compared to leading Transformer-based model The study suggests that CVAF-Net provides a more favorable accuracy-computation trade-off than the Transformer baseline. However, clinical relevance will depend on more than average MAE, and future work should focus on noise-aware training and uncertainty-aware prediction. The use of fixed-length PPG segments is relevant for wearable monitoring as it avoids explicit dependence on precise pulse cycle detection.

Ablation and Plausibility

The ablation study verified the contribution of each module, showing that removing the spatial encoder led to the most significant performance degradation across all datasets. The cross-view attention fusion proved superior to simple concatenation methods in modeling complex interdependencies between views Plausibility analysis showed that predicted CO exhibited a robust positive correlation with HR and negative associations with SVR on the RC dataset, consistent with exercise physiology. This suggests that CVAF-Net captured activity-related hemodynamic structure rather than merely fitting arbitrary waveform patterns. The study concludes that CVAF-Net is a promising direction for wearable CO monitoring while larger clinical and free-living validation remains necessary before practical deployment.

References

[1] World Health Organization. Cardiovascular diseases (cvds), 7 2025

[2] G. A. Roth, G. A. Mensah, C. O. Johnson, et al. Global burden of cardiovascular diseases and risk factors, 1990-2019 Journal of the American College of Cardiology, 76(25):2982–3021, 12 2020

[3] D. Rusinaru, Y. Bohbot, F. Djelaili, et al. Normative reference values of cardiac output by pulsed-wave doppler echocardiography in adults The American Journal of Cardiology, 140:128–133, 2 2021

[4] C. Vahdatpour, D. Collins, S. Goldberg. Cardiogenic shock Journal of the American Heart Association, 8(8):e011991, 4 2019

[5] B. F. Geerts, L. P. Aarts, J. R. Jansen Methods in pharmacology: measurement of cardiac output British Journal of Clinical Pharmacology, 71(3):316–330, 3 2011

[6] J. Grensemann Cardiac output monitoring by pulse contour analysis the technical basics of less-invasive techniques Frontiers in Medicine, 5:64, 3 2018

[7] B. Saugel, J. L. Vincent Cardiac output monitoring by pulse contour analysis Current Opinion in Critical Care, 24(3):165–172, 6 2018

[8] P. H. Charlton, P. A. Kyriacou, J. Mant, V. Marozas, P. Chowienczyk J Alastruey Wearable photoplethysmography for cardiovascular monitoring Proceedings of the IEEE, 110(3):355–381, 3 2022

[9] Q. Y. Lee, S. J. Redmond, G. S. Chan, et al Estimation of cardiac output and systemic vascular resistance using a multivariate regression model with features selected from the finger photoplethysmogram and routine cardiovascular measurements BioMedical Engineering OnLine, 12(1):19, 3 2013

[10] L. Wang, E. Pickwell-MacPherson, Y. P. Liang, Y. T. Zhang Noninvasive cardiac output estimation using a novel photoplethysmogram index 2009 Annual International Conference of the IEEE Engineering in Medicine and Biology Society, pages 1746–1749, 9 2009

[11] E. Ipar, L. J. Cymberknop, R. L.

Improvements for AI systems

A cross-view attention fusion mechanism that allows FSM-derived structural representations to query both FSM and temporal representations, enabling the model to learn which raw temporal patterns are relevant under a given prior-guided feature context. This enables the system to select specific dynamic features based on learned structural context, improving relevance in complex physiological states.
A dual-branch architecture that learn[s] complementary representations from raw 1D PPG segments and an 8-channel FSM view for non-invasive CO estimation. This allows the model to simultaneously capture sequential dynamics (temporal view) and multi-scale structural information (spatial/FSM view).
A computational efficiency improvement by achieving a twelvefold reduction in FLOPs compared to leading Transformer-based model while maintaining accuracy, making it feasible for deployment on resource-constrained wearable hardware.

Abstract

Accurate cardiac output (CO) estimation from photoplethysmography (PPG) is promising for unobtrusive hemodynamic monitoring, but remains difficult since CO is jointly determined by cardiac function and vascular tone. Conventional feature-based models use physiologically meaningful PPG descriptors, yet depend on accurate pulse detection and may miss latent temporal relationships. In contrast, fully end-to-end deep learning models learn directly from raw PPG but often underuse established PPG-derived prior information. Here, we introduce the Cross-View Attention Fusion Network (CVAF-Net), a prior-guided dual-view deep learning model for CO estimation from short, fixed-length PPG segments. CVAF-Net processes raw PPG as a temporal view and a feature sequence map (FSM) as a structured prior-guided view, and fuses the two representations through cross-view attention. The model was independently evaluated using 5-, 15-, and 30-s segments from three datasets: simulated pulse waves (3323 subjects), vasoconstriction provocation (79 subjects), and resting/cycling activities (10 subjects), and was compared with multiple machine learning and deep learning benchmarks. CVAF-Net outperformed most benchmark methods and achieved performance comparable to a state-of-the-art Transformer-based model, with a mean absolute error (MAE) of 0.19 L/min (MAPE: 3.95%) on simulated data and high accuracy in real-world settings (minimum MAE: 1.20 L/min). Importantly, CVAF-Net reduced FLOPs by twelvefold compared with the leading Transformer-based model. Plausibility analysis showed physiologically consistent CO estimates, with expected correlations with age (ρ= -0.274), heart rate (ρ= 0.894), and systemic vascular resistance (ρ= -0.740). These findings indicate that CVAF-Net provides an accurate, computationally efficient, and generalizable approach for continuous wearable-based CO monitoring.