Whole-brain functional activity and connectivity for the classification of subjective tinnitus: a machine learning study
TL;DR
The findings indicate that the ROI-based rs-fMRI machine-learning provides preliminary proof-of-concept evidence for the objective confirmation of subjective tinnitus.
Abstract
Background and purpose Clinical evaluation of subjective tinnitus mainly depends on patients’ self-reported auditory complaints, and standardized neuroimaging biomarkers for characterizing its central brain functional abnormalities remain lacking. The aim of this study is to utilize the resting-state functional magnetic resonance imaging (rs-fMRI) machine learning technique based on the region of interest (ROI), and to construct an exploratory classification framework for subjective tinnitus by analyzing functional activities and connectivity. Methods The rs-fMRI data of 63 patients with subjective tinnitus (38.79 ± 15.79) and 84 healthy controls (HCs) (42.01 ± 9.45) were collected from the Department of Otorhinolaryngology, the first affiliated Hospital of Nanchang University. Five analysis methods were used: regional homogeneity (ReHo), amplitude of low frequency fluctuation (ALFF), fraction amplitude of low frequency fluctuation (fALFF), resting state functional connectivity (RSFC) and degree centrality (DC). A total of 7,134 features are extracted after z conversion. Then, the predicted features were selected through Mann–Whitney U test, the variables with high pairwise correlation (the correlation coefficient is greater than 0.75) were removed, the least absolute shrinkage and selection operator method was used to screen the features. Finally, a machine learning model was constructed by combining logistic regression (LR), support vector machine (SVM) and random forest (RF), and the performance differences of the three models were compared. Results 21 features are retained, including 3 zRSFCs, 1 zALFFs, 6 zfALFFs, 3 zDCs, and 8 zReHos. Based on these 21 features, the model accuracy and area under the curve constructed by LR, SVM and RF were 75.51% and 0.80, 80.27% and 0.82, 73.47% and 0.79, respectively. Conclusion Our findings indicate that the ROI-based rs-fMRI machine-learning provides preliminary proof-of-concept evidence for the objective confirmation of subjective tinnitus. The imaging information based on rs-fMRI has the potential to become a neuroimaging biomarker for tinnitus.