High-Dimensional Causal Mediation Model for Detecting Neuroimaging Biomarkers

Discovering neuroimaging biomarkers that link pharmacological interventions to neurobehavioral outcomes is imperative for advancing mental disorder treatments. Functional connectivity, a pivotal biomarker, reveals neural interactions and helps decipher complex neurological and psychiatric conditions. We propose a high-dimensional mediation model to map neural pathways and pinpoint sparse connectivity mediators. However, small sample sizes in neuroimaging studies often lead to unstable parameter estimates. To address this, we introduce "Pathway Network," an innovative penalization technique for biomarker discovery and hub detection. This convex penalty combines the Pathway LASSO and Network-constrained penalties, aiming to stabilize mediation effect estimates and integrate brain network information through a Laplacian matrix grounded in graph theory. To solve the optimization problem, we developed an Alternating Direction Method of Multipliers algorithm and created the "HDMAADMM" R package. Extensive simulations validate the effectiveness of this approach. Application to neuroimaging data from individuals with internalizing psychopathology further demonstrates its clinical utility. This method represents a significant step forward in biomarker detection within high-dimensional mediation analysis, with potential applications in genetics, precision medicine, and beyond.