Severity-dependent alterations of functional network segregation and integration in tobacco use disorder

Previous research on brain network topology in Tobacco Use Disorder (TUD) has been inconsistent, likely due to overlooking the heterogeneity of addiction severity. Consequently, how these topological alterations manifest across different smoking severities is not fully understood.

Resting-state functional magnetic resonance imaging (fMRI) and clinical data were collected from 102 males (24 heavy smokers, 36 light smokers, 42 healthy controls). Based on the fMRI data, we computed global graph metrics and applied Network-Based Statistics (NBS) analysis to identify abnormal subnetworks, and performed correlation analyses between global graph metrics and clinical scales. A Support Vector Machine (SVM) classifier was constructed using functional connectivity (FC) strength, graph metrics, and multi-scale fusion features to discriminate the severity of smoking at the individual level within a rigorous repeated stratified nested cross-validation framework.

Graph-theory analysis revealed that patients with Tobacco Use Disorder (TUD) exhibited significant large-scale topological reorganization, characterized by increased global integration, indexed by higher global efficiency (E_glob), and reduced local segregation, indexed by a lower clustering coefficient (C_p). Network-Based Statistics (NBS) identified two distinct subnetworks showing reduced connectivity in the TUD group, primarily involving the default mode, limbic, and sensorimotor networks. Subgroup analyses further demonstrated a severity-dependent pattern of network alterations. Light smokers showed increased E_glob with relatively preserved C_p, whereas heavy smokers exhibited a significant reduction in C_p accompanied by more extensive disruption of connectivity across distributed higher-order networks. In contrast, no significant subnetworks were detected in the light smoker group. Importantly, correlation analysis revealed a significant negative association between C_p and Fagerström Test for Nicotine Dependence (FTND) scores. Finally, a multi-scale machine learning model integrating network features and evaluated using 10 × 5 repeated stratified nested cross-validation achieved optimal classification of smoking severity (ensemble AUC = 0.8380), outperforming models based on single-modality features.

TUD involves a severity-dependent pathophysiological gradient, which may reflect a hypothesized shift from compensatory strengthening of global integration in light smokers to advanced-severity weakening of local network organization in heavy smokers. The severity-dependent reduction in local segregation, together with widespread hypoconnectivity, suggests multi-scale network disruption. These multi-scale functional network features provide preliminary evidence supporting their potential utility for severity stratification in TUD.