Accelerating Computation of DCM for ERP in MATLAB by External Function Calls to the GPU

Area of Science:

• Neuroscience
• Computational Neuroscience
• Biophysics

Background:

• Dynamic Causal Modelling for Event Related Potentials (DCM for ERP) is crucial for studying neuronal effective connectivity.
• The Expectation Maximization (EM) algorithm used in DCM is computationally intensive, posing challenges for analysis speed and model selection.
• Existing MATLAB implementations face limitations due to computational demands and potential combinatorial model explosion.

Purpose of the Study:

• To enhance the performance of DCM for ERP analysis in MATLAB.
• To implement a parallel computing scheme utilizing graphics processing units (GPUs) for faster DCM for ERP estimation.
• To evaluate the efficiency and accuracy of the proposed GPU-based approach.

Main Methods:

• Developed a parallel computing scheme for DCM for ERP using external function calls to a GPU.
• Dynamically partitioned and distributed DCM computations to threads for parallel processing based on model complexity and hardware.
• Evaluated performance using synthetic data and validated accuracy with experimental data.

Main Results:

• Achieved a significant acceleration of DCM for ERP computation, with speedup factors up to 155 for parallel parts in simulations.
• Experimental data showed an average speedup of approximately 7 per model, varying with complexity.
• The GPU-based implementation yielded qualitatively similar group-level results to the original MATLAB implementation.

Conclusions:

• The proposed GPU-based implementation offers a substantial speed improvement for DCM for ERP analysis.
• This tool can serve as an efficient screening method for selecting the most likely DCM models.
• The approach provides valuable guidance for potential future clinical applications, such as online diagnosis.