TL;DR
This paper introduces a novel sparse Bayesian non-linear regression method to accurately pinpoint multiple electrical activation onsets in the heart using non-invasive body surface potential mapping (BSPM). This advancement is crucial for personalizing cardiac electrophysiology models, leading to better diagnosis and treatment of heart conditions.
What Problem Does This Paper Address?
Understanding and predicting heart conditions, as well as tailoring treatments like cardiac resynchronization therapy (CRT), often requires creating patient-specific models of cardiac electrophysiology (EP). A key challenge in building these models is accurately estimating parameters from non-invasive data. While previous work has focused on simpler cases with a single electrical activation onset, real-world scenarios often involve multiple onsets (e.g., during biventricular pacing) or complex cardiac conditions like scar tissue. Existing methods struggle with these complexities, and crucially, often don't provide a measure of confidence in their estimations. This paper aims to bridge this gap by developing a method that can handle multiple onsets and provide uncertainty estimates, making cardiac EP personalization more robust and reliable.
Methods
The core of this paper is the application of a Sparse Bayesian Non-linear Regression technique, specifically a Relevance Vector Machine (RVM), for estimating cardiac activation maps. Here's a breakdown of their approach:
- Non-invasive Personalization Framework: They leverage a fast forward EP model to generate a large database of simulated cardiac electrical activity and corresponding body surface potentials (BSPM). This simulated data is then used to train their model.
- Feature Extraction from BSPM: To capture essential information from the BSPM signals, they extract specific descriptors related to timings, areas, and extremum signs of the QRS complex.
- Dimensionality Reduction of Myocardial Shape: The 3D myocardial geometry can be complex. To handle this, they employ a hierarchical decomposition using eigenvectors of the stiffness matrix (related to Laplacian operator). This reduces the dimensionality of the myocardial shape representation, focusing on large spatial variations and making the regression problem more tractable. They found that using the first 400 modes provided a good balance between accuracy and dimensionality.
- Relevance Vector Machine (RVM) Regression: The RVM is a sparse kernel-based regression method that uses a sparsity-inducing prior within a Bayesian framework. Unlike other methods like Elastic-Net, RVM automatically estimates the noise level and doesn't require manual tuning of regularization parameters. It also provides a posterior probability for estimated quantities, offering a measure of confidence. They applied the RVM to regress the reduced-space myocardial activation times from the extracted BSPM features.
- Onset Localization: After estimating the activation map, they further extract the locations of the activation onsets by identifying the points of minimal activation time.
Datasets
- Simulated Patient-Specific Database: The primary dataset used for training was generated by simulating cardiac EP models. This database consisted of 1000 simulations with fixed conduction velocity and two randomly selected onset locations on the myocardial surface (endocardium and epicardium).
- Clinical Data: For evaluation, they used data from a single non-ischaemic patient who underwent simultaneous biventricular pacing. This included:
- Body Surface Potential Mapping (BSPM) signals from a CardioInsight jacket (205 torso sensors, 1kHz sampling rate).
- Ventricular myocardial geometry (derived from CT scans).
- Torso sensor locations.
- Pacing lead locations.
- Cardiac fiber orientations estimated from CT images.
Key Insights and Ideas
- RVM for Sparse and Confident Estimation: The RVM's ability to perform sparse regression and provide uncertainty estimates is a significant advantage. This means it can automatically select relevant features and give a measure of how reliable the estimated activation times are.
- Dimensionality Reduction is Key: Reducing the complexity of the myocardial geometry using spectral shape analysis (eigenmodes of the stiffness matrix) is crucial for making the non-linear regression problem computationally feasible and effective.
- Handling Multiple Onsets: The proposed method successfully addresses the challenge of localizing multiple activation onsets, which is a significant step beyond previous work limited to single onsets.
- Bridging Simulated and Real Data: The framework effectively bridges the gap between simulated data (used for training) and real patient data (for evaluation), demonstrating its practical applicability.
Key Results
- Accurate Onset Localization: The RVM regression on the reduced shape space achieved an error distance of less than 23mm for both pacing leads, with a mean error of 18.4mm. This is a notable improvement over other methods tested.
- Improved Activation Map Reconstruction: The RVM method, particularly when applied to the reduced space, produced smoother and more accurate activation maps compared to kernel ridge regression or RVM applied independently to each vertex.
- Confidence Estimation: The RVM provided standard deviation (std) estimates for the activation times, indicating areas of higher and lower confidence. For the left ventricle, the estimated std was below 17ms.
- Computational Efficiency: The RVM regression on the reduced space was efficient, requiring only 400 regressions for 14K activation times.
Limitations and Future Work
- Multivariate Regression Limitation: The current RVM implementation primarily performs multivariate regression of a single output scalar, not a vector. While extensions exist, this is a noted limitation.
- Scar Tissue Handling: The method showed promise in capturing scar boundaries but may require further refinement for accurate estimation in the presence of scar tissue that blocks activation waves.
- Myocardial Shape Precision: The use of a generic mesh due to limitations in precise myocardial shape availability from CT scans could be a source of error. Future work will involve testing with patient-specific myocardial meshes.
- Error Source Apprehension: While RVM provides regression error estimates, other error sources (e.g., lead localization, BSPM data quality) need to be fully accounted for to improve confidence in the results.
- Validation on More Patients: The method was tested on only one patient, necessitating more experiments for robust validation.
- Estimating Additional Parameters: Future work aims to extend the approach to robustly estimate other cardiac parameters, such as local conduction velocity.
Summary updated