TL;DR
This paper tackles a critical, often overlooked source of uncertainty in computational fluid dynamics models of blood circulation: "model mismatch." By developing and applying a Bayesian framework that explicitly accounts for discrepancies between the model and reality, the authors demonstrate that this approach corrects biased parameter estimates and provides more reliable predictions compared to traditional methods. Their findings highlight the superiority of a nonlinear vessel wall model with radius-dependent stiffness for accurately representing pulmonary hemodynamics.
What Problem Does This Paper Address?
Computational models are increasingly vital for understanding cardiovascular diseases and predicting treatment outcomes. However, before these models can be trusted for clinical decision-making, they must be rigorously tested and their uncertainties quantified. A significant challenge in this process is "model mismatch" – the inherent discrepancy between a simplified mathematical model and the complex biological reality it aims to represent, as well as potential errors in how measurement noise is modeled. Traditional methods often minimize the mean squared error (MSE) between model predictions and data, implicitly assuming a perfect model and correctly specified noise. This paper argues that ignoring model mismatch leads to biased parameter estimates and overly confident, unreliable predictions, hindering the clinical utility of these models.
Methods
The core of this research lies in a Bayesian inference framework designed to explicitly account for model mismatch. Instead of solely focusing on parameter uncertainty, the authors incorporate uncertainty in the model form itself. They achieve this by representing model mismatch using Gaussian Processes (GPs), a powerful tool for modeling functions with uncertainty. Their Bayesian approach jointly samples parameters for both the physiological model and the model mismatch function from their posterior distribution using Markov Chain Monte Carlo (MCMC). This allows them to simultaneously quantify uncertainty in parameters, model form, and measurements. For model selection, they employ the Watanabe-Akaike Information Criterion (WAIC), a standard metric for comparing probabilistic models.
Datasets
The study utilizes two types of data:
- Physiological Data: Measured pulmonary artery blood pressure data from a control mouse lung. This data was obtained invasively via catheterization and averaged over 20 cardiac cycles.
- Synthetic Data: Generated from forward simulations of the mathematical model, with added non-stationary, additive Gaussian correlated errors. This synthetic data was used to validate the proposed method and assess its performance under controlled conditions.
Key Insights and Ideas
- Model Mismatch is Crucial: The paper strongly emphasizes that ignoring model mismatch leads to biased parameter estimates and underestimation of uncertainty, making predictions unreliable for clinical applications.
- Gaussian Processes for Mismatch: Using Gaussian Processes is an effective way to model and quantify the uncertainty arising from model discrepancy and incorrect noise assumptions.
- Bayesian Inference for Robustness: A Bayesian approach that jointly infers model parameters and model mismatch parameters provides a more comprehensive and reliable quantification of uncertainty.
- MSE is Sub-optimal with Mismatch: Minimizing Mean Squared Error (MSE) is a suboptimal inference strategy when model mismatch is present, as it doesn't account for error correlations and penalizes models for poor uncertainty quantification.
- Nonlinear Models Offer Better Fit: A nonlinear pressure-area relationship for vessel walls, especially when combined with radius-dependent stiffness, provides a better fit to the pulmonary hemodynamics data than simpler linear models.
Key Results
- Bias Correction: The proposed method, which accounts for model mismatch, successfully corrects the bias observed in parameter estimates and haemodynamic predictions when compared to traditional MSE-based methods.
- Improved Uncertainty Quantification: The Bayesian approach with model mismatch correction leads to wider, more realistic uncertainty bounds in both parameter and output spaces.
- Model Selection Favors Nonlinearity: Formal model selection using WAIC consistently favored nonlinear vessel wall models with radius-dependent stiffness over linear models, indicating a better representation of pulmonary hemodynamics.
- Synthetic Data Validation: Synthetic data experiments confirmed that incorporating model mismatch leads to more accurate parameter recovery and better uncertainty estimates.
- Best Model Identified: The nonlinear pressure-area relationship with stiffness dependent on the unstressed radius (model I) was found to be the best predictor of pulmonary hemodynamics in the control mouse.
Limitations and Future Work
- Single Mouse Data: The physiological conclusions are based on data from a single mouse, limiting generalizability. Future work should involve comparing models across multiple subjects.
- WAIC vs. Marginal Likelihood: The study used WAIC, an asymptotically based criterion. Future work could explore alternative model selection approaches, such as marginal likelihood, which may be more robust but computationally intensive.
- Distinguishing Mismatch Sources: It was not possible to separately distinguish between model discrepancy and measurement error due to the retrospective nature of the data and lack of prior knowledge. Future work could explore methods to disentangle these contributions.
- Boundary Conditions and Network Structure: The study focused on one type of boundary condition and assumed a fixed network geometry. Future research could investigate more complex boundary conditions and model uncertainty in network structure.
- Real-time Applications: While the study demonstrates the potential for efficient surrogate models (emulators) for faster predictions, achieving real-time personalized treatment planning remains a significant computational challenge, especially for complex models with many parameters.
Summary updated