🧬 biology

HIPNO: Symmetry-Aware Physics-Informed Neural Operators for Noninvasive Hemodynamic Inference

The paper introduces HIPNO, a symmetry-aware physics-informed neural operator that resolves scale symmetry in hemodynamic inference by parameterizing a network in a quotient space to accurately recover vascular decay and flow dynamics from non-invasive pressure signals, outperforming baseline methods in both accuracy and counterfactual robustness.

Original authors: Yunbei Pan, Jiahang Sha, Simon A. Lee, Maxime Cannesson, Wei Wang, Jeffrey N. Chiang

Published 2026-08-12

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Yunbei Pan, Jiahang Sha, Simon A. Lee, Maxime Cannesson, Wei Wang, Jeffrey N. Chiang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine trying to figure out how much water is flowing through a complex network of pipes just by listening to the sound of the water hitting the walls. In the world of medicine, doctors often need to know exactly how much blood is pumping through a patient's body and how tight or loose the blood vessels are. This is crucial for making life-or-death decisions during surgery or in the intensive care unit. Usually, the only way to get this information is to stick a thin tube (a catheter) directly into a major artery. It's like measuring the water pressure inside a fire hose by sticking a needle right into the stream—it works perfectly, but it's risky, painful, and can cause infections, so doctors only do it for the sickest patients.

For everyone else, doctors rely on non-invasive tools like a blood pressure cuff that squeezes your arm or a sensor that shines light through your finger. These tools are safe and easy, but they have a blind spot: they can tell you the pressure, but they can't easily tell you the flow or the "tightness" of the vessels. It's like knowing the water pressure in a pipe but not knowing if the pipe is wide open or clogged, or how hard the pump is working. Scientists have been trying to build computer programs that can guess these hidden numbers just from the safe, easy signals, but there's a tricky math problem in the way. Different combinations of flow and vessel tightness can create the exact same pressure reading, making it impossible for a computer to know which one is the "real" answer without getting confused.

This is where a new study called HIPNO comes in. The researchers built a smart computer system that acts like a detective who knows the rules of physics. Instead of trying to guess the hidden numbers directly, the system learns to speak a special "secret language" that strips away the confusion. It separates the "flow" (how hard the heart is pushing) from the "tone" (how tight the vessels are) by realizing that while these two things can change together to keep the pressure the same, they actually behave differently when you look at them through the lens of physics. By training on data from nearly 2,600 patients, HIPNO learned to predict these hidden states with much greater accuracy than previous methods. It can now estimate how fast the blood is decaying in the vessels—a key sign of vascular health—about 32% more accurately than a standard guess.

The paper doesn't just say "we guessed better"; it proves that the system understands the mechanism. When the researchers simulated what would happen if they gave a patient a drug to relax the vessels (vasodilation) or a drug to tighten them (vasoconstriction), HIPNO correctly predicted the direction of the change in over 90% of cases. It even figured out how to combine its predictions with a patient's heart rate and other basic info to estimate cardiac output (the total volume of blood pumped) with a reasonable degree of accuracy. However, the authors are careful to note that while the system is a powerful research tool, it still needs a little bit of help from a real-world measurement (like a single blood pressure reading or a flow measurement) to set the absolute scale. It's like a map that shows the perfect shape of a city but needs one known street sign to tell you exactly where you are standing. Until more testing is done, the tool is intended for research to help doctors understand what's happening inside the body, rather than for making immediate clinical decisions on its own.

The Story of HIPNO: Solving the "Pressure Puzzle"

The Problem: The Great Mix-Up
Imagine you are trying to guess how hard a person is pedaling a bike (flow) and how much friction the brakes are applying (resistance) just by looking at how fast the bike is going (pressure). The tricky part is that if you pedal harder but also loosen the brakes, the speed stays the same. If you pedal softer but tighten the brakes, the speed also stays the same. To a computer looking only at the speed, these two very different situations look identical. This is called "scale symmetry," and it's the reason why previous computer models got stuck. They could guess the pressure perfectly, but they couldn't tell you if the patient had a weak heart or tight vessels because the math allowed for multiple "correct" answers that were actually wrong.

The Solution: The "Secret Language" of Physics
The HIPNO team realized that instead of trying to guess the raw numbers (like exact flow or exact resistance), they should teach the computer to learn the ratio between them. They created a new set of coordinates, or a "secret language," that naturally separates these mixed-up variables.

Flow Drive (U): Think of this as the "pedaling power" normalized by the size of the pipes.
Vascular Tone (τWK): Think of this as the "decay time" or how quickly the pressure drops when the heart stops pumping for a split second. This tells you how "leaky" or "tight" the vessels are.

By forcing the computer to learn in this secret language, the mix-up disappears. The computer can now see that a change in pressure caused by a drug that relaxes vessels looks different from a change caused by a drug that makes the heart pump harder, even if the pressure reading looks similar at first glance.

The Experiment: Testing the Detective
The researchers tested HIPNO on a massive dataset from 2,562 patients, using over 945,000 tiny 10-second windows of data. They compared HIPNO to other smart models that just tried to guess the pressure waveform.

The Result: HIPNO didn't just guess the pressure; it guessed the "vascular decay time" (a proxy for how the vessels are behaving) with 32% lower error than a standard baseline.
The Mechanism Check: This is the coolest part. The researchers played a "what-if" game. They asked the computer: "What happens to the pressure if we make the vessels 25% more relaxed?" HIPNO correctly predicted that the pressure would drop. They asked, "What if we make the heart pump 25% harder?" HIPNO correctly predicted the pressure would rise. In 90% or more of the scenarios, the computer's reaction matched the expected direction of real-world medicine. This proves the computer isn't just memorizing patterns; it's actually learning the physics of the body.

The Catch: The Missing Ruler
There is one thing HIPNO can't do alone: it can't tell you the exact absolute number of liters of blood flowing without a little help. Because the math allows for a "scale" to be shifted, the system knows the shape of the flow and the ratio of the resistance, but it needs one external measurement (like a single blood pressure reading or a known flow value) to set the "ruler" and tell it exactly how big the numbers are. The paper shows that once you give it that one reference point, it can calibrate itself to estimate cardiac output quite well.

Why It Matters
This work is a big step toward "non-invasive hemodynamic monitoring." If this technology matures, doctors might one day be able to see exactly how a patient's blood vessels and heart are interacting just by looking at a simple finger sensor and an ECG, without needing to stick a needle in an artery. It could mean safer surgeries and better care for patients who are too risky to have invasive tubes. However, the authors are very clear: this is currently a research tool. It's a brilliant prototype that solves the math puzzle of symmetry, but it still needs more testing to be used as a standalone medical device in a hospital. The paper suggests that with the right calibration, we are getting closer to a future where we can see the invisible forces of the circulatory system without ever breaking the skin.

Technical Summary: HIPNO – Symmetry-Aware Physics-Informed Neural Operators for Noninvasive Hemodynamic Inference

Problem Statement

Continuous hemodynamic monitoring is critical for guiding treatment in surgery and intensive care, yet gold-standard signals (e.g., cardiac output, vascular tone) typically require invasive catheterization, restricting their use to critically unstable patients. While non-invasive signals like ECG and photoplethysmography (PPG) are routinely available, reconstructing continuous volumetric flow and vascular tone from them remains a challenge due to the scarcity of clinical ground truth.

A fundamental obstacle in physics-informed hemodynamic inference is structural non-identifiability. Specifically, the observation model (pressure) exhibits scale symmetry: proportional rescaling of cardiac flow ( $Q_{in}$ ), vascular resistance ( $R_1, R_2$ ), and arterial compliance ( $C$ ) can generate identical pressure waveforms. Consequently, distinct latent physiological states map to the same observable signal, making it impossible to uniquely recover absolute physical parameters (like absolute flow or resistance) from pressure data alone. Existing data-driven models often fail to address this ambiguity, while standard physics-informed approaches may leave the symmetry intact in the inverse problem.

Methodology

The authors introduce HIPNO (Hemodynamic Inference via Physics-informed Neural Operators), a framework designed to resolve scale ambiguity by operating directly in the invariant quotient space of the underlying symmetry group.

1. Symmetry-Aware Formulation

The authors formalize the scale symmetry of the 3-element Windkessel model (Proposition 1). They demonstrate that for any scaling factor $s > 0$ , the transformation $C' = sC$, $Q'_{in} = sQ_{in}$ , $R'_1 = R_1/s$ , and $R'_2 = R_2/s$ leaves the pressure waveform unchanged.

To eliminate this ambiguity, HIPNO parameterizes the network in quotient coordinates that are invariant under this transformation:

Compliance-normalized flow: $U = Q/C$
Decay time constant: $\tau_{WK} = R_2 C$
Characteristic-impedance coordinate: $\kappa = R_1 C$

In this space, the boundary equations depend only on $U$ , $\tau_{WK}$ , and $\kappa$ , effectively quotienting out the scale degree of freedom. The model predicts these invariant coordinates alongside the pressure waveform.

2. Neural Operator Architecture

HIPNO employs a hybrid architecture coupling signal encoders with a physics-conditioned continuous trunk:

Encoders: Separate convolutional Transformer encoders process ECG and PPG signals. A cross-attention decoder fuses these modalities to capture complementary timing information (e.g., QRS timing vs. pulse arrival).
Cardiac Phase Learning: The model learns a monotone cardiac phase $\phi(t)$ to align within-beat morphology, enabling evaluation at arbitrary times.
Trunk Network: A DeepONet-style trunk maps Fourier features of phase and space to continuous arterial fields (cross-sectional area $A$ , normalized flow $U$ , and transmural pressure $P_{tm}$ ). This branch is conditioned on the global window representation via Feature-wise Linear Modulation (FiLM).
Physics Branch: Enforces the Windkessel ordinary differential equations (ODEs) and arterial PDE residuals. It predicts $\tau_{WK}$ , $\kappa$ , and venous pressure $P_v$ .
Direct Branch: A separate pathway predicts the pressure waveform directly from the inputs.
Blending: The final output is a weighted blend of the direct and physics branches, $P(t) = \alpha P_{dir}(t) + (1-\alpha)P_{phy}(t)$ , optimized via a curriculum learning strategy.

3. Training and Data

The model was trained on 945,499 intraoperative windows from 2,562 patients derived from the VitalDB dataset. The training objective includes:

Reconstruction Loss: Masked MSE on pressure waveforms, standardized shape, and derivatives.
Physics Loss: Windkessel coupling, PDE residuals, and periodicity constraints.
Time Constant Supervision: A log-MSE loss against a pressure-derived proxy $\tau_{wave}$ (fit to diastolic decay) to guide the learning of $\tau_{WK}$ .
Scale Constraints: Priors on parameters and a fixed compliance normalization ( $C=1$ ) to define the quotient space.

Key Results

1. Prediction of Vascular Decay

HIPNO predicts the decay time constant $\tau_{WK}$ (a proxy for vascular tone) with significantly higher accuracy than a population baseline.

Performance: On held-out test windows, HIPNO achieved a 32% reduction in log-scale Mean Absolute Error (MAE) compared to a baseline predicting the training cohort median ($0.297$ vs. $0.434$).
Correlation: The model captures inter-patient variation with a Pearson correlation of $r=0.678$ on the log scale.

2. Mechanism Separation (Counterfactuals)

A critical validation of the quotient coordinate formulation is mechanism separation. The authors performed counterfactual perturbations on the learned coordinates:

Separation of Tone and Flow: Perturbing $\tau_{WK}$ (vascular tone) while holding $U$ (flow drive) fixed produced directional responses consistent with vasodilation/vasoconstriction in >90% of windows. Conversely, perturbing $U$ while holding $\tau_{WK}$ fixed produced flow-driven responses.
Clinical Relevance: This separation allows the model to distinguish between hypovolemia (low flow) and vasodilation (low tone), which present similarly in pressure-only models but require opposite treatments. In contrast, pressure-only baselines cannot disentangle these mechanisms.

3. Pressure Reconstruction

HIPNO maintains high fidelity in pressure reconstruction while generating the additional physiological coordinates.

Accuracy: The Mean Arterial Pressure (MAP) MAE was $10.88$ mmHg, comparable to state-of-the-art reversible waveform models (INN-PAR: $10.34$ mmHg) and foundation models (PaPaGei: $9.01$ mmHg).
Trade-off: The waveform correlation was slightly lower ($0.672$ vs. $0.817$ for INN-PAR), with errors concentrated in systolic peaks, but the model successfully recovers the invariant coordinates necessary for mechanistic insight.

4. Cardiac Output Calibration

The authors demonstrated that the invariant coordinates can be used to calibrate monitor Cardiac Output (CO).

Ablation: Adding HIPNO's flow coordinate ( $\log U$ ) to a calibration model using EHR covariates and device indicators yielded marginal improvements ( $<0.02$ L/min MAE change).
Conclusion: While the coordinates do not independently outperform clinical covariates, they can be integrated without degradation, supporting the feasibility of using non-invasive signals to refine CO estimates when an external scale reference is available.

Significance and Claims

The paper claims that HIPNO demonstrates how physics-informed neural operators can learn in the identifiable quotient space of a mechanistic observation model.

Resolving Ambiguity: By explicitly identifying and quotienting out the scale symmetry, HIPNO recovers interpretable latent parameters (vascular tone and flow drive) that are otherwise mathematically undetermined from pressure alone.
Clinical Utility: The model provides a continuous, non-invasive index of arterial compliance and vascular tone ( $\tau_{WK}$ ) and separates flow drive from tone, addressing a key limitation in current hemodynamic monitoring.
Explicit Scale Limitation: The authors modestly note that while the relative dynamics are recovered, the absolute physical scale (e.g., absolute cardiac output in L/min) remains unidentified without an external reference (e.g., a compliance reference or a flow measurement). The formulation explicitly identifies this missing measurement as the necessary next step for absolute quantification.

The work establishes a generalizable four-step procedure for symmetry-aware operator design: (1) identify the symmetry group, (2) form invariant coordinates, (3) train the operator in that space, and (4) identify the external measurement required to fix the scale. This approach is presented as applicable beyond hemodynamics to other inverse problems involving scale symmetries.