Physics-Constrained Deep Learning Model for Contactless Blood Pressure Monitoring from Triaxial Bodyseismography

arXiv:2608.23562 · eess.SP, cs.AI, physics.bio-ph · Submitted 2026-08-24 · Read on arXiv

Listen

Radio episode about this paper

Transcript

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

Tom: I'm Tom, and with me are Jane, Lu, senior AI researcher at Tsinghua, Meng, lead engineer at a mysterious AI startup and Lalam, the in-house Large Language Model.

Jane: Today's paper: "Physics-Constrained Deep Learning Model for Contactless Blood Pressure Monitoring from Triaxial Bodyseismography".

Tom: Ballistocardiography (BCG) is promising for unobtrusive long-term blood pressure (BP) monitoring in laboratory settings, but traditional BCG signals are vulnerable to variations in body-bed interaction and personal hemodynamic changes,

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

Title and authors: Tom: So, moving past the challenges, let’s look at what the Phy-BP framework actually does in detail, as described in the summary of "Physics-Constrained Deep Learning Model for Contactless Blood Pressure Monitoring from Triaxial Bodyseismography." Jane Essentially, it proposes three major components: an adaptive quality control step to filter out bad data, a physical model describing three dee wave propagation, and then a deep learning architecture that combines them.

Lu: The first part is this adaptive quality control algorithm that uses neighboring beat patterns and universal templates to select segments enriched with cardiogenic components. This pre-filtering step seems really important because if you feed the deep learning model a lot of noisy or non-cardiac vibration data, it will just learn those noise patterns instead of the actual pressure dynamics.

Meng: I see that filtering step as essential for efficiency; we only want to train on high-fidelity cardiac events, not general surface vibrations. It’s about reducing the training burden on the main model by removing low-quality segments upfront.

Lalam: That quality control is like cleaning up the raw footage before showing it to the main editor; it ensures that the AI is learning from high-quality cardiac events rather than just random shaking on a surface. It’s a necessary layer of refinement.

Tom: And then there's the core deep learning architecture where they embed that physical model right into the structure to characterize the coupling among the three axes. This is where we move from simple correlation to physically informed representation.

Jane: So, instead of just looking at X, Y, and Z features separately, they use a gate-based feature alignment mechanism to force those features to agree on a single latent state based on the physical model’s requirements. It’s about making sure that what we extract from each axis is actually representing the same underlying physiological process.

Lu: That alignment, combined with projecting those features into a low-dimensional latent state sequence governed by reduced-order dynamics, is where the real innovation lies. They are essentially using an ordinary differential equation to govern how the latent features should evolve over time based on the physical model.

Meng: That ODE approach is quite sophisticated; it’s taking a high-dimensional problem and forcing it onto a simpler, tractable mathematical trajectory, which is exactly what we need for reliable, real-time inference.

Lalam: It means the system isn't just guessing the next state; it’s following a physically plausible path dictated by the body-bed dynamics. This intrinsic grounding is what elevates this approach beyond standard black-box models.

The paper's summary: Tom: Let's talk about how they actually improved the system, because it’s not just a concept; there are concrete steps they took in "Physics-Constrained Deep Learning Model for Contactless Blood Pressure Monitoring from Triaxial Bodyseismography." Jane They introduce the physics-constrained layer between the encoder and decoder that handles feature alignment across the X, Y, and Z axes using a gate-based mechanism.

Lu: That layer is crucial because it aligns the latent features extracted from different directions by learning a time-varying gate network that combines axis-specific features with a shared prototype feature. It specifically addresses how energy leaks or shifts across axes, which was a major weakness in previous models.

Meng: I’m interested in the second step they take after alignment, which is the physics-constrained regularization that projects these aligned features into a low-dimensional latent state sequence constrained by the reduced-order dynamics derived from the physical model. That projection step is where you enforce physical consistency on the learned representations.

Lalam: That constraint acts as a strong prior, meaning even if the input data is a bit ambiguous, the AI has to stick to what physics predicts should happen next in terms of blood pressure dynamics. It’s like giving the model a strong sense of physical reality to anchor its predictions.

Tom: And they validated this by showing that using different combinations of axes actually helps, because adding the X-axis increased performance from a five point seven six percent improvement with only the Y axis to thirty-eight point two two percent when all three were used. That's a pretty strong demonstration of how complementary information is captured here.

Jane: It really shows that leveraging triaxial BSG isn't just about having more data; it’s about capturing the complete three dee body-bed response, which is what this method achieves.

The paper's improvements: Tom: So, to wrap up our discussion on "Physics-Constrained Deep Learning Model for Contactless Blood Pressure Monitoring from Triaxial Bodyseismography," we see a system that uses physical constraints and triaxial data to create a much more robust estimation of blood pressure than what single-axis methods could manage. Jane It’s about moving past models that are just pattern matchers and toward models that truly understand the mechanics of how the body interacts with the bed.

Lu: I think this has huge implications for how we model complex biological systems because it suggests a path where AI can incorporate known physical laws directly into its decision-making process, not just statistical correlations. We’re moving toward AI that is more fundamentally principled.

Meng: From a practical standpoint, if we can get this level of robustness, it opens the door for deploying contact-less monitoring systems in settings where wearing sensors is impractical, which is what the paper aimed for. That's a significant step toward real-world usability.

Lalam: For me, it means that future AI applications will have this intrinsic physical grounding, leading to systems that are inherently more trustworthy and less prone to errors when deployed in sensitive areas like healthcare. It builds trust into the very structure of the AI.

Tom: That’s a great summary of where we are with this paper; it really shows that integrating physical principles into deep learning architecture leads to more stable results. We’ve covered a lot about how they achieved that robustness through quality control and physics-informed constraints.

Jane: It’s been fascinating watching the Phy-BP framework move from an abstract idea to a validated method that handles multi-axis data effectively. We've really explored how this physical modeling changes the way we think about signal processing for biomedical sensing.

Lu: It opens up new avenues for AI research where incorporating known physical laws isn't just an afterthought but a foundational requirement, which is exciting. We should keep looking at how this constraint can be applied across other complex physical domains.

Meng: I think the immediate impact is in validating that these contact-less systems can actually function reliably outside the lab environment where perfect calibration is possible. That validation will be key for moving this technology forward.

Lalam: And for my perspective, this intrinsic physical grounding means future AI systems will naturally be more trustworthy, which is the biggest cultural shift we can see in how we deploy these kinds of tools. That trust is something we need to prioritize in our development.

Conclusion: Tom: So, we’ve seen how the Phy-BP framework uses triaxial bodyseismography and physical constraints to create a more robust blood pressure estimation method than traditional single-axis approaches. Jane It really shows how incorporating known physics into deep learning architecture can lead to a much more stable and reliable system for monitoring things like blood pressure.

Lu: I think this has huge implications for how we model complex biological systems; it suggests a path where AI can truly incorporate known physical laws into its decision-making process, not just statistical correlations. Meng From a practical standpoint, if we can get this level of robustness, it opens the door for deploying contact-less monitoring systems in settings where wearing sensors is impractical, which is what the paper aimed for.

Lalam: For me, it means that future AI applications will have this intrinsic physical grounding, leading to systems that are inherently more trustworthy and less prone to errors when deployed in sensitive areas like healthcare. Tom That level of trust built by grounding the AI in physics sounds like something we should be striving for in every single model we build.

Jane: It’s about moving past models that are just pattern matchers and toward models that actually understand the underlying mechanics of how the body interacts with the bed, which is what this Phy-BP framework achieves. Lu And I think that's where it gets really exciting because it shows a way to make AI systems more fundamentally principled.

Meng: I’m interested in how this structure translates to actual deployment; if we can get this level of robustness, it opens the door for deploying contact-less monitoring systems in settings where wearing sensors is impractical, which is what the paper aimed for. Tom Exactly, because that’s where the real utility lies for practical engineering applications.

Lalam: It means that future AI applications will have this intrinsic physical grounding, leading to systems that are inherently more trustworthy and less prone to errors when deployed in sensitive areas like healthcare. Jane We need to focus on making those foundations solid so the resulting tools can be trusted by users in critical situations.

Lu: I think this has huge implications for how we model complex biological systems; it suggests a path where AI can truly incorporate known physical laws into its decision-making process, not just statistical correlations. Tom That structural approach is really what makes this work; it’s not just adding layers, it’s changing the architecture itself.

Meng: From a practical standpoint, if we can get this level of robustness, it opens the door for deploying contact-less monitoring systems in settings where wearing sensors is impractical, which is what the paper aimed for. Jane It really shows how triaxial data can provide that comprehensive view needed for accurate physiological estimation.

Lalam: For me, it means that future AI applications will have this intrinsic physical grounding, leading to systems that are inherently more trustworthy and less prone to errors when deployed in sensitive areas like healthcare. Tom So we’re leaving with a really solid framework for how physics can guide deep learning in sensing problems.

Jane: It’s about moving past models that are just pattern matchers and toward models that actually understand the underlying mechanics of the body-bed interaction, which is what this Phy-BP framework achieves. Lu I think this has huge implications for how we model complex biological systems; it suggests a path where AI can truly incorporate known physical laws into its decision-making process, not just statistical correlations.

Meng: From a practical standpoint, if we can get this level of robustness, it opens the door for deploying contact-less monitoring systems in settings where wearing sensors is impractical, which is what the paper aimed for. Tom Exactly, because that’s where the real utility lies for practical engineering applications.

Lalam: For me, it means that future AI applications will have this intrinsic physical grounding, leading to systems that are inherently more trustworthy and less prone to errors when deployed in sensitive areas like healthcare. Jane We need to focus on making those foundations solid so the resulting tools can be trusted by users in critical situations.

Tom: So we’re leaving with a really solid framework for how physics can guide deep learning in sensing problems, and this whole paper is a great example of that. Jane What should we look at next to see how these physical principles might apply to vision models?

University of Georgia

eess.SP, cs.AI, physics.bio-ph

Submitted: 2026-08-24

Updated: 2026-09-28

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 92/100

The gist: Ballistocardiography (BCG) is promising for unobtrusive long-term blood pressure (BP) monitoring in laboratory settings, but traditional BCG signals are vulnerable to variations in body-bed

Key concepts

Triaxial Bodyseismography (BSG)
This involves measuring vibrations from the body using three orthogonal axes (X, Y, Z). It captures subtle movements caused by blood acceleration against the bed. The challenge is that these signals are complex and vary based on posture and body movement, making it hard to extract reliable blood pressure information alone.
Physics-Constrained Deep Learning Model
This technique integrates a mathematical model of how waves travel through the body (a partial differential equation) directly into the neural network. This forces the deep learning model to learn features that are consistent with real-world physics, ensuring better alignment and robustness when predicting blood pressure.
Gate-based Feature Alignment
This is a mechanism within the deep learning model that uses a time-varying network (a gate) to merge features extracted from the three different BSG axes (X, Y, Z). It learns to combine axis-specific signals with a shared prototype feature, creating a unified representation of the body's physical state.
Reduced-Order Dynamics
This process simplifies a complex physical model describing 3D wave propagation into a simpler system of ordinary differential equations (ODEs). This tractable mathematical representation is used to constrain the deep learning model, ensuring that the learned latent features evolve in a way that matches the actual physical behavior of the body-bed system.

Terminology

Summary

Ballistocardiography (BCG) is promising for unobtrusive long-term blood pressure (BP) monitoring in laboratory settings, but traditional BCG signals are vulnerable to variations in body-bed interaction and personal hemodynamic changes, causing misaligned representations that affect model generalizability and robustness.

The gist: A physics-constrained deep learning framework, Phy-BP, is proposed to estimate blood pressure from triaxial bodyseismography (BSG) by embedding a physical model of 3D wave propagation into the deep learning architecture to align multi-axis features under a shared latent state evolution.

Physiological Rationale and Challenges

Blood pressure is jointly determined by stroke volume (SV), heart rate (HR), and total peripheral resistance (TPR). While BSG captures the subtle body-bed recoil induced by blood acceleration, it does not directly measure arterial pressure, meaning absolute BP level is modulated by subject-specific hemodynamic properties like TPR, which cannot be directly observed from non-invasive signals. Existing 1D BCG signals are insufficient because they only capture the body recoil along a single axis (e.g., the Y-axis), and significant vibration energy may leak into the X- and Z-axes under different postures, as illustrated in Figure 1(b). Furthermore, existing deep learning models often rely on black box approaches, leading to misaligned representations during feature extraction due to uncertainties in BP ground truth and BCG morphology.

The Phy-BP Framework

To address these challenges, the proposed Phy-BP framework integrates three core modules:

  1. Quality Control: An adaptive quality-control algorithm is designed to select BSG segments enriched with cardiogenic components by jointly considering neighboring beat patterns and universal cardiogenic templates. This involves using a matched filter with a universal template, such as the fourth derivative of a Gaussian function, to identify high-quality BSG cycles. A second matched filter based on a dynamic template learned from the current segment uses linear time warping (LTW) to find optimal scaling lengths that ensure all candidate cycles become as close as possible to the same representative morphology.

  2. Physical Model: A physical model is established to describe 3D wave propagation in the body–bed system, modeled as an equivalent viscoelastic medium excited by a single cardiogenic excitation force, formulated via a partial differential equation (PDE). This model is reduced-order using modal truncation to derive a state-space representation, transforming the high-dimensional PDE into a tractable ODE: Mq¨(t) + Cq˙(t) + Kq(t) = Fc(t).

  3. Physics-Constrained Deep Learning Model: The deep learning model utilizes an encoder-decoder architecture with a crucial physics-constrained layer embedded between the encoder and decoder. This layer performs two steps: Gate-based Feature Alignment to align the latent features extracted from the three BSG axes (X, Y, Z) by learning a time-varying gate network that combines axis-specific features with a shared prototype feature. Subsequently, Physics-constrained Regularization projects these aligned features into a low-dimensional latent state sequence constrained by the reduced-order dynamics derived from the physical model to govern their evolution: x˙ a(t) = xa(t + 1) − xa(t) ≈ Axa(t) + BFr(t).

Experimental Validation and Results

The framework was validated on a 162-hour hospital dataset collected from 21 subjects, using invasive ABP measurements as the golden-standard reference. The performance of Phy-BP was compared against baselines including ResNet18, PITN, MSTNN, and FSNet. Overall performance metrics showed that Phy-BP achieved the best results across all three BP targets: MAEs of 6.03 for SBP, 4.23 for DBP, and 5.07 for MAP mmHg. Compared to the ResNet baseline, Phy-BP obtained an overall improvement of 52.76%.

Ablation and Robustness Analysis

A series of ablation studies verified the necessity of each component:

  1. Quality Control Effectiveness: Increasing the threshold for required matched cardiac cycles (e.g., from Cycles> 4 to Cycles> 7) progressively removes samples with large prediction errors, leading to reduced STD and MAE and increased PCC.

  2. Necessity of Triaxial BSG: The results demonstrated that using different combinations of axes provided complementary information; for instance, adding the X-axis increased performance from Phy-BP(Y only)'s 5.76% improvement to 38.22%. Using all three axes yielded the best overall improvement of 52.76%.

  3. Effects of Physics Constraint: Varying the weighting coefficient λ in the physics-constrained layer showed that performance improved as λ increased, with "λ = 1.

Improvements for AI systems

As a fastidious researcher, I have analyzed the proposed Physics-Constrained Deep Learning Model for Contactless Blood Pressure Monitoring from Triaxial Bodyseismography (Phy-BP). The core innovation lies in bridging the gap between raw, noisy triaxial bodyseismography (BSG) data and accurate blood pressure (BP) estimation by explicitly embedding physical principles into the deep learning architecture.

Here are the specific improvements to AI systems based on this paper, and what these improved systems can achieve:

The proposed Phy-BP framework enhances existing AI models through three interconnected modules: Adaptive Quality Control, Physics-Informed Feature Alignment, and Physics-Constrained Regularization. These enhancements move the system from a purely data-driven black box predictor to a physically constrained, robust estimator.

Here are the specific improvements and their resulting capabilities:

  1. The AI system incorporates an automated pre-processing step using a matched filter with universal and dynamic templates to perform real-time quality control on triaxial BSG signals (Section III-B).

  2. The deep learning model is architecturally augmented with a Physics-Constrained Layer that aligns latent representations extracted from the three axes (X, Y, Z) via a gate-based feature alignment mechanism (Equation 16) and projects them into a shared latent excitation state governed by a reduced-order dynamical system derived from wave propagation theory (Section III-D).

  3. The training objective function is modified to include a physics-guided regularization term, which constrains the evolution of the latent states to follow the expected dynamics of the body-bed system (Equation 18).

These improved AI systems can achieve:

  1. Enhanced Robustness Against Signal Distortion and Noise: The quality control module filters out low signal-to-noise ratio (SNR) segments based on physiological cardiogenic components, ensuring that the deep learning model is trained only on high-fidelity data. This prevents the model from learning spurious correlations caused by patient movement or non-cardiac vibrations.

  2. Improved Generalizability Across Body Postures and Sensor Misalignment: By aligning the triaxial features under a shared latent state using physics constraints, the system learns to decouple axis-specific noise (like energy leakage in X or Z axes) from the fundamental cardiac excitation signal. This results in a model that is invariant to variations in body posture or shifts of fiducial points caused by soft mattresses, leading to superior performance on unseen real-world data compared to single-axis models.

  3. Superior Accuracy Under Limited Training Data: The physics constraint acts as a structural prior, imposing the known physical coupling between cardiac excitation and triaxial wave propagation (Section III-C). This regularization allows the model to learn consistent representations even when training samples are limited or subject diversity is low, significantly improving data efficiency and preventing the model from overfitting to dataset-specific artifacts.

  4. Quantifiable Sensitivity to Hemodynamic Drift: The framework explicitly models how BP is a function of cardiac output (CO) and Total Peripheral Resistance (TPR). The calibration mechanism, informed by PiCCO-derived CO measurements, allows the AI system to dynamically correct its predictions based on subject-specific hemodynamic changes in real-time. This leads to a more stable and clinically relevant estimate of SBP, DBP, and MAP that is less susceptible to rapid changes in vascular resistance (TPR).

  5. Optimized Multi-Modal Fusion: The triaxial input architecture leverages the complementary information captured by different vibration directions. The system learns the optimal weighting (via ablation studies) for X, Y, and Z axes to maximize prediction accuracy, effectively capturing the complete 3D body-bed response induced by a single cardiogenic event.

Related papers