DIA-PINN. A physics-informed machine learning method to estimate global intrinsic diastolic chamber properties of the left ventricle from pressure-volume data

This paper introduces DIA-PINN, a physics-informed neural network framework that accurately and robustly estimates global intrinsic diastolic properties of the left ventricle from instantaneous pressure-volume data, outperforming traditional optimization methods by providing reliable, initialization-insensitive estimates of stiffness, relaxation, and elastic recoil.

The Gist

Imagine your heart is a high-performance pump, specifically the Left Ventricle (LV). For this pump to work perfectly, it needs to do two things during its "rest" phase (diastole):

1. Relax: It needs to soften up quickly to let blood in (like a deflating balloon).
2. Recoil: It needs to have a bit of springiness to help push blood back out (like a rubber band snapping back).

Doctors have long used a "gold standard" test called a Pressure-Volume (PV) Loop to measure how well the heart is doing these two things. Think of a PV loop as a map that tracks the heart's pressure and size at every split second.

However, reading this map has always been tricky. Traditional methods are like trying to solve a complex puzzle by looking at only a few pieces at a time. They are sensitive to noise, require a human to guess where to start, and often get stuck in "local minima" (getting lost in a dead-end of the puzzle).

Enter DIA-PINN, the new hero of this story.

What is DIA-PINN?

DIA-PINN stands for Diastolic Physics-Informed Neural Network. That's a mouthful, so let's break it down with a simple analogy.

Imagine you are trying to teach a robot how to drive a car.

• Traditional Machine Learning: You show the robot 1,000 videos of cars driving and say, "Guess the rules." The robot might learn weird patterns (like "cars always turn left at red lights") that aren't actually true physics.
• Physics-Informed Learning (PINN): You give the robot the videos, BUT you also give it the actual laws of physics (Newton's laws, friction, gravity) and say, "You can learn from the videos, but you cannot break the laws of physics."

DIA-PINN is exactly this. It is an AI that looks at the heart's pressure and volume data, but it is forced to follow the known biological rules of how a heart muscle relaxes and stretches. It doesn't just "guess" the answer; it solves a math problem where the answer must make physical sense.

How Does It Work? (The "Two-Part" Heart)

The researchers programmed DIA-PINN to understand that the heart's pressure during rest is made of two distinct ingredients mixed together:

1. The "Active" Ingredient (Relaxation): This is the heart muscle actively letting go of tension. It's like a person slowly unclenching a fist.
2. The "Passive" Ingredient (Stiffness & Spring): This is the heart's natural rubberiness. It's like a spring that resists being stretched but wants to snap back.

DIA-PINN looks at the messy, real-world data from a patient's heart and tries to separate these two ingredients to tell the doctor: "Your heart relaxes at this speed, but it is this stiff."

Why Is This Better Than the Old Way?

The old method (called Global Optimization or GOM) is like trying to find a hidden treasure on a map by guessing coordinates. If you start guessing in the wrong place, you might get stuck in a small valley and think you found the treasure, when the real treasure is on a mountain nearby. It's very sensitive to where you start.

DIA-PINN is like having a GPS that knows the terrain.

• It doesn't care where you start: Even if you give it random starting guesses, it always finds the same, correct answer. It's not easily confused.
• It handles noise: Real heart data is messy (like static on a radio). Because DIA-PINN knows the "laws of physics," it can ignore the static and find the true signal.
• It needs more data to shine: The study showed that DIA-PINN works best when the heart is tested under changing conditions (like squeezing the veins to lower blood pressure). This is like testing a car on a hill, a flat road, and a curve to really understand its engine. When the heart is tested this way, DIA-PINN gives the most accurate map possible.

The Results

The researchers tested DIA-PINN on two things:

1. Fake Data: They created 4,000 perfect, computer-generated heart loops. DIA-PINN recovered the "true" numbers almost perfectly, beating the old method.
2. Real Patients: They tested it on 59 real people (some with heart failure, some healthy). DIA-PINN's results matched the old method's best results but were much more consistent and didn't get confused by bad starting guesses.

The Bottom Line

DIA-PINN is a smarter, more reliable way to measure how well a heart relaxes and stretches. By combining the flexibility of modern AI with the rigid rules of physics, it gives doctors a clearer, more accurate picture of heart health.

Think of it as upgrading from a hand-drawn sketch of a heart's mechanics to a 3D, physics-accurate simulation that never gets tired, never gets confused, and always follows the rules of nature. This could help doctors diagnose heart failure earlier and tailor treatments more precisely in the future.

Technical Summary

1. Problem Statement

Assessing the intrinsic global diastolic properties of the left ventricle (LV) is critical for understanding heart failure and other cardiac diseases. While invasive pressure-volume (PV) loop analysis is the gold standard, current methodologies face significant limitations:

• Traditional Fitting: Conventional methods fit specific segments of the PV loop (e.g., isovolumic relaxation or end-diastolic points) separately. This assumes relaxation and volume-dependent properties are uncoupled, ignores active relaxation during late diastole, and is highly sensitive to noise, landmark selection errors, and beat-to-beat variability.
• Global Optimization Method (GOM): A previous advancement, the GOM, fits complete diastolic periods using physiologically informed equations. However, it suffers from sensitivity to parameter initialization and bounds, often converging to spurious local minima, which compromises reproducibility and robustness, especially in noisy clinical data.

There is a need for a method that combines the mechanistic interpretability of constitutive models with the flexibility and robustness of machine learning to estimate diastolic parameters without relying on careful manual initialization.

2. Methodology: DIA-PINN Framework

The authors developed DIA-PINN (Diastolic Physics-Informed Neural Network), a framework that embeds a mechanistic model of LV diastole directly into a neural network training process.

A. Constitutive Model

The physical core of the model decomposes diastolic pressure ($P$) into two components:

1. Active Relaxation ($P_{act}$): Modeled as an exponential decay function dependent on the time constant of relaxation ($\tau$) and the pressure at the onset of diastole ($P_0$).
2. Passive Recoil/Stiffness ($P_{pas}$): Modeled as a piecewise logarithmic function dependent on volume ($V$), stiffness coefficient ($\beta$), elastic recoil coefficient ($\alpha$), equilibrium volume ($V_{eq}$), and dead volume asymptote ($V_0$).

The total pressure is defined by a set of six constitutive parameters: $\Theta = \{\tau, \beta, \alpha, V_{eq}, V_0, V_{max}\}$.

B. Neural Network Architecture

• Input: Time ($t$), measured LV Volume ($V$), and measured LV Pressure ($P$).
• Output: Predicted Pressure ($P_{pred}$) across the entire diastolic phase.
• Trainable Variables: The six constitutive parameters ($\Theta$) are treated as global trainable variables within the network, rather than fixed weights.
• Architecture: A feedforward neural network (FNN) with two hidden layers (5 neurons each), using Sigmoid activation functions.
• Optimizer: Adam optimizer with a learning rate of 0.001.

C. Loss Function

The training loss ($L$) is a multi-component function designed to enforce three objectives simultaneously:

1. Physics-Informed Term ($L_{phy}$): Ensures the predicted pressure satisfies the governing constitutive equations (the sum of active and passive components).
2. Data Fidelity Term ($L_{data}$): Minimizes the mean squared error between the predicted pressure and the measured clinical/synthetic pressure waveforms.
3. Physiological Plausibility Term ($L_{phys}$): Uses ReLU-like constraints to enforce biological bounds (e.g., positive stiffness, realistic volume ranges), preventing non-physiological solutions.

3. Validation Strategy

The study evaluated DIA-PINN through two distinct phases:

1. Synthetic Data (Monte Carlo Simulations):
   • Generated 4,000 synthetic PV loops with systematically varied physiological parameters.
   • Tested under two conditions: Single-beat data and Vena Cava Occlusion (VCO) sequences (simulating preload reduction).
   • Added random and Gaussian noise to test robustness.
   • Compared performance against the Global Optimization Method (GOM) under various initialization and bounding constraints.
2. Clinical Data:
   • Applied to invasive PV data from 59 patients (39 with Heart Failure with Preserved Ejection Fraction [HFpEF] and 20 healthy controls).
   • Compared DIA-PINN estimates directly against the previously validated GOM.

4. Key Results

A. Synthetic Data Performance

• Accuracy: DIA-PINN recovered all constitutive indices with Intraclass Correlation Coefficients (ICC) near unity.
• Robustness to Initialization: Unlike GOM, DIA-PINN was insensitive to parameter initialization. It consistently converged to the same parameter subspace regardless of random starting points, whereas GOM frequently failed or converged to local minima without tight bounds.
• Impact of Preload Manipulation:
  • Single-beat: DIA-PINN performed well but showed higher dispersion in estimating the elastic recoil coefficient ($\alpha$) and dead volume ($V_0$).
  • VCO (Multi-beat): Performance improved markedly. Varying preload enhanced parameter identifiability, reducing bias to near-zero for all parameters.
• Comparison to GOM: DIA-PINN outperformed GOM in recovering parameters, particularly when parameter bounds were not strictly constrained a priori.

B. Clinical Data Performance

• Correlation: DIA-PINN estimates strongly correlated with GOM estimates for relaxation time ($\tau$, $R=0.98$), equilibrium volume ($V_{eq}$, $R=0.99$), and stiffness ($\beta$, $R=0.90$).
• Elastic Recoil: The elastic recoil coefficient ($\alpha$) showed moderate correlation ($R=0.77$), consistent with synthetic findings where this parameter is harder to identify without low-volume data.
• Stability: DIA-PINN provided stable estimates across both HFpEF patients and controls, demonstrating generalizability.

5. Key Contributions

1. Novel Framework: Introduced DIA-PINN, the first PINN framework specifically designed to infer global intrinsic diastolic properties directly from raw invasive PV loop data.
2. Initialization Independence: Solved the critical limitation of traditional optimization methods by demonstrating that PINNs can converge to physiologically correct solutions without the need for precise initial guesses or narrow parameter bounds.
3. Unified Modeling: Successfully integrated active relaxation and passive stiffness/recoil mechanisms into a single, differentiable model, allowing for the simultaneous estimation of coupled time- and volume-dependent forces.
4. Open Source: The authors released the full Python code, making the methodology reproducible and adaptable.

6. Significance and Limitations

Significance:
DIA-PINN bridges the gap between mechanistic modeling and machine learning in cardiology. It offers a robust, reproducible, and interpretable tool for quantifying diastolic dysfunction. By removing the dependency on manual initialization, it lowers the barrier for clinical translation and improves the reliability of diastolic phenotyping in heart failure research.

Limitations:

• Parameter Identifiability: The study acknowledges that the dead volume asymptote ($V_0$) and the elastic recoil coefficient ($\alpha$) remain difficult to identify accurately if the PV data does not extend into low-volume regimes (below the equilibrium volume). This is a fundamental mathematical limitation of the constitutive model when data is sparse in that region.
• Data Requirements: While single-beat analysis is possible, the method performs significantly better with preload reduction maneuvers (VCO), which may not always be feasible in all clinical settings.

Conclusion

DIA-PINN represents a significant advancement in cardiac mechanics modeling. By embedding physiological laws into neural network training, it provides a superior alternative to traditional fitting and optimization methods, offering high accuracy, initialization insensitivity, and the ability to handle noisy clinical data effectively.