Abnormal white matter microstructure in tobacco use disorder: A machine learning study based on whole-brain skeletonized DTI data.

BACKGROUND

There is growing evidence that cigarette smoking is associated with abnormal white matter (WM) microstructure. However, traditional research using group-level mass-univariate statistical analysis, despite strict multiple comparison corrections, may still yield false positives or false negatives. In this study, a multivariate machine learning method was performed to enhance the identification of smoking-related WM regions.

METHODS

The whole-brain skeletonized maps of diffusion metrics derived from diffusion tensor imaging of 60 tobacco use disorder (TUD) participants and 66 non-TUD subjects were utilized as classification features. A linear support vector machine (SVM) classifier was trained to differentiate between TUD and non-TUD individuals, and smoking-related WM regions were determined using discriminative score. Correlation analysis was conducted to evaluate the relationship between classification scores and smoking-related variables among TUD participants.

RESULTS

The SVM classifier achieved an accuracy over 0.80 and an area under the curve exceeding 0.91, with maximal discriminative weights localized to the anterior corona radiata, posterior thalamic radiation, and genu of the corpus callosum (i.e., forceps minor). Moreover, classification scores were positively correlated with the onset age of cigarette use.

CONCLUSIONS

Our multivariate approach using whole-brain skeletonized maps of diffusion metrics outperformed those using regional features, demonstrating superior classification performance. The WM discriminative regions identified by our approach may offer a more effective way to distinguish individuals with nicotine addiction, potentially advancing our understanding of smoking-related brain alterations.