Supervised Estimation of Granger-Based Causality between Time Series.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Brain effective connectivity analysis uses neural activity measurements (e.g., M/EEG) to identify causal interactions.
• Existing methods include model-based (e.g., Granger causality) and model-free approaches.
• The Granger criterion assumes autoregressive data models for causal inference.

Purpose of the Study:

• To compare a novel classification-based causality detection method with standard Granger causality analysis.
• To evaluate the performance and robustness of the supervised learning approach in detecting effective connectivity.

Main Methods:

• Formulated causality detection as a supervised learning classification task.
• Developed a feature space incorporating precedence and predictability measures for time series.
• Customized the approach for MAR (Vector Autoregression) models and compared two variations against standard Granger causality.

Main Results:

• The supervised method demonstrated superior performance compared to the standard Granger approach in simulations.
• The proposed method exhibited enhanced robustness to noise.
• Achieved 2nd place in the Biomag2014 causality detection competition and reduced false positives in rat neural recordings.

Conclusions:

• The supervised learning approach offers a more robust and effective method for brain effective connectivity analysis.
• This machine learning-based strategy advances causal inference in neuroscience signal processing.