DLGC: Dictionary Learning Based Granger Causal Discovery for Cortico-Muscular Coupling

Investigating causal pathways between the brain and muscles is crucial for identifying biomarkers associated with movement disorders such as Parkinson's disease, multiple sclerosis, and dystonia. The transmission of information from the brain to muscles is complicated by the background activities and various forms of noise and interference. Thus, extracting meaningful causal patterns from recorded sensor data presents a significant challenge. This paper presents a novel approach to disentangling causal information from physiological signals while addressing measurement noise and other forms of perturbations. We propose a dictionary learning based autoregressive model capable of extracting meaningful features from physiological signals while simultaneously eliminating noise. To efficiently solve the non-convex, non-smooth optimization problem inherent in our approach, we employ a recent second-order proximal algorithm that leverages a local surrogate function of the objective function to converge to a potentially better local minimum. Our experimental results, conducted on real physiological signals, demonstrate the effectiveness of our proposed method in disentangling causal information and mitigating noise, thereby advancing our understanding of brain-muscle interactions in movement control.