Machine learning-based identification and brain network characterization of flight cadets compared with air traffic control (ATC) students using static and dynamic functional connectivity.

Introduction: This study used ATC students with a comparable civil aviation educational background and aviation-related cognitive demands, but no actual flight experience, as the control group. This study aimed to systematically compare the performance of static and dynamic functional connectivity in distinguishing flight cadets and to explore time-varying brain network characteristics associated with flight training. Methods: Resting-state functional magnetic resonance imaging (rs-fMRI) data were collected from 39 flight cadets and 37 ATC students, from which static functional connectivity (sFC) and dynamic functional connectivity (dFC) features were extracted. Following feature selection, five classifiers, including Support Vector Machine (SVM), eXtreme Gradient Boosting (XGBoost), Random Forest (RF), Gaussian Naive Bayes (GNB), and K-Nearest Neighbors (KNN), were systematically evaluated. Results: The primary analysis showed that dFC generally achieved better classification performance than sFC. The optimal dFC-SVM model reached an Area Under the ROC Curve (AUC) of 0.949, with an accuracy of 0.895, a sensitivity of 0.945, and a specificity of 0.848, whereas the best sFC model achieved an AUC of only 0.828. Under a more stringent supplementary validation procedure, the dFC-SVM and sFC-SVM models achieved AUCs of 0.65 and 0.59, respectively. The window-length sensitivity analysis showed that model performance remained within a similar range across the tested window lengths. High-contribution connections identified in the primary analysis involved brain regions such as the left insula and parahippocampal gyrus and were distributed across the Default Mode Network (DMN), Central Executive Network (CEN), and Sensorimotor Network (SMN). Discussion: Taken together, dFC may contain additional time-varying discriminative information beyond static connectivity. However, the magnitude of this advantage was limited in the supplementary validation, and the related findings should be regarded as preliminary evidence requiring further validation. By adopting a comparable control design within the aviation domain, this study further characterized neurofunctional representations associated with flight training experience and provided neuroimaging clues for subsequent longitudinal validation incorporating training stage, behavioral performance, and flight performance, as well as for exploring adaptive monitoring of flight training and multidimensional auxiliary assessments.