Population‐based priors in cardiac model personalisation for consistent parameter estimation in heterogeneous databases

Roch Molléro, Xavier Pennec, Hervé Delingette, Nicholas Ayache, Maxime SermesantInternational Journal for Numerical Methods in Biomedical Engineering 2018 · DOI

TL;DR

This paper introduces the Iteratively Updated Priors (IUP) algorithm, a novel method for personalizing cardiac models. IUP leverages population-level data to create "priors" that guide parameter estimation, enabling more consistent and robust personalization, especially when clinical measurements are sparse or heterogeneous.

What Problem Does This Paper Address?

Creating accurate, patient-specific cardiac models is crucial for understanding and predicting heart behavior. However, these models often have many parameters, while clinical data for a single patient is usually limited. This makes it impossible to uniquely estimate all parameters, leading to an "ill-posed inverse problem." This ambiguity hinders the clinical application of personalized cardiac models. Existing methods often struggle with heterogeneous datasets where measurements vary significantly between patients or are missing altogether.

Methods

The core of the paper is the Iteratively Updated Priors (IUP) algorithm. It works by:

  1. Successive Personalization: The algorithm iteratively personalizes a population of cardiac models.
  2. MAP Estimation: For each patient, it uses Maximum A Posteriori (MAP) estimation to find parameter values that best fit the available measurements, guided by a prior probability distribution.
  3. Prior Update: Crucially, the prior probability distribution for the next iteration is derived from the distribution of the personalized parameters from the previous iteration. This creates a feedback loop where the population's collective knowledge informs individual estimations.
  4. Subspace Selection: As the algorithm converges, it naturally identifies a lower-dimensional linear subspace where the population's parameters tend to lie. This subspace represents the directions of variability that are observable from the data.
  5. IUP* Variant: A modified version, IUP*, is introduced to ensure a high goodness-of-fit for all cases, even with heterogeneous data, by adaptively reweighting the data-fit term.

The paper also links IUP to sparse regression techniques, showing it effectively minimizes a population-wide cost function with a penalty on the dimensionality of the parameter space.

Datasets

  • 137 complete cases: Used for initial validation and parameter subspace selection.
  • 811 heterogeneous cases: A larger, more realistic dataset with varying measurements (e.g., volume, pressure, ejection fraction) and missing data, used to demonstrate the robustness of the IUP* algorithm.

Key Insights and Ideas

  • Population-based priors as a regularization tool: Instead of relying on a priori knowledge, IUP learns the population's parameter distribution from the data itself, using it to regularize parameter estimation.
  • Dimensionality reduction as a consequence of data limitations: The algorithm naturally finds a lower-dimensional subspace where parameters are best constrained by the available data, effectively identifying "observable" parameter directions.
  • Handling heterogeneous data: The IUP* variant is crucial for real-world scenarios where measurements are incomplete or vary significantly, ensuring all patients achieve a good fit.
  • Bridging the gap between individual and population models: IUP creates a consistent framework that leverages population statistics to improve individual model personalization, especially for unobservable parameters.
  • External parameters can enhance priors: Incorporating readily available patient data (like height and weight) into the prior distribution can further improve the estimation of unobservable parameters.

Key Results

  • The IUP algorithm successfully reduces the parameter space to a minimal, sufficient linear subspace.
  • For the 137-case dataset, IUP identified a 4-dimensional subspace (out of 6 original parameters) that allowed for consistent personalization with high goodness-of-fit and minimal parameter variability.
  • For the 811-case heterogeneous database, the IUP* algorithm achieved a high goodness-of-fit for all cases, even those with missing measurements, by leveraging population-based priors.
  • When external parameters (height, weight) were integrated into the prior, the correlation between estimated parameters and these external factors improved, particularly for parameters that were otherwise unobservable.

Limitations and Future Work

  • Model non-linearity: The non-linear nature of cardiac models means that parameter estimation can be non-convex, potentially leading to multiple possible solutions or subspaces.
  • Parameter subspace relevance: While the algorithm identifies mathematically relevant subspaces, their physiological interpretability still requires expert supervision.
  • Computational cost for 3D models: Applying IUP to complex 3D cardiac models with computationally expensive simulations remains a significant challenge.
  • Future Work:
    • Investigating different prior distribution shapes beyond Gaussian.
  • Exploring multifidelity approaches to reduce computational burden for 3D models.
  • Developing methods to automatically assess the physiological relevance of selected parameter subspaces.
  • Further integration of external patient data to improve prior distributions.

Summary updated

Abstract

Personalised cardiac models are a virtual representation of the patient heart, with parameter values for which the simulation fits the available clinical measurements. Models usually have a large number of parameters while the available data for a given patient are typically limited to a small set of measurements; thus, the parameters cannot be estimated uniquely. This is a practical obstacle for clinical applications, where accurate parameter values can be important. Here, we explore an original approach based on an algorithm called Iteratively Updated Priors (IUP), in which we perform successive personalisations of a full database through maximum a posteriori (MAP) estimation, where the prior probability at an iteration is set from the distribution of personalised parameters in the database at the previous iteration. At the convergence of the algorithm, estimated parameters of the population lie on a linear subspace of reduced (and possibly sufficient) dimension in which for each case of the database, there is a (possibly unique) parameter value for which the simulation fits the measurements. We first show how this property can help the modeller select a relevant parameter subspace for personalisation. In addition, since the resulting priors in this subspace represent the population statistics in this subspace, they can be used to perform consistent parameter estimation for cases where measurements are possibly different or missing in the database, which we illustrate with the personalisation of a heterogeneous database of 811 cases.
BibTeX
@article{mollero2018populationbased,
  author = {Roch Molléro and Xavier Pennec and Hervé Delingette and Nicholas Ayache and Maxime Sermesant},
  title = {{Population‐based priors in cardiac model personalisation for consistent parameter estimation in heterogeneous databases}},
  journal = {International Journal for Numerical Methods in Biomedical Engineering},
  year = {2018},
  doi = {10.1002/cnm.3158},
}
Figure