Classifying Vulnerability to Sleep Deprivation Using Resting-State Functional MRI Graph Theory Metrics

• Published in: Frontiers in neuroscience Vol. 15; p. 660365
• Main Authors: Xu, Yongqiang, Yu, Ping, Zheng, Jianmin, Wang, Chen, Hu, Tian, Yang, Qi, Xu, Ziliang, Guo, Fan, Tang, Xing, Ren, Fang, Zhu, Yuanqiang
• Format: Journal Article
• Language: English
• Published: Switzerland Frontiers Research Foundation 07.06.2021; Frontiers Media S.A
• Subjects: Algorithms; Artificial intelligence; Brain architecture; Brain mapping; Classification; Cognitive ability; Functional magnetic resonance imaging; Homeostasis; Laboratories; Learning algorithms; Machine learning; Magnetic resonance imaging; Medical imaging; Neuroimaging; Neuroscience; psychomotor vigilance task; Sensorimotor system; Sleep and wakefulness; Sleep deprivation; Software; Support vector machines; Thalamus; Vigilance; vulnerability; sleep deprivation; vulnerability; psychomotor vigilance task; machine learning; functional magnetic resonance imaging
• ISSN: 1662-453X; 1662-4548; 1662-453X
• DOI: 10.3389/fnins.2021.660365

Sleep deprivation (SD) has become very common in contemporary society, where people work around the clock. SD-induced cognitive deficits show large inter-individual differences and are trait-like with known neural correlates. However, few studies have used neuroimaging to predict vulnerability to SD. Here, resting state functional magnetic resonance imaging (fMRI) data and psychomotor vigilance task (PVT) data were collected from 60 healthy subjects after resting wakefulness and after one night of SD. The number of PVT lapses was then used to classify participants on the basis of whether they were vulnerable or resilient to SD. We explored the viability of graph-theory-based degree centrality to accurately classify vulnerability to SD. Compared with during resting wakefulness, widespread changes in degree centrality (DC) were found after SD, indicating significant reorganization of sleep homeostasis with respect to activity in resting state brain network architecture. Support vector machine (SVM) analysis using leave-one-out cross-validation achieved a correct classification rate of 84.75% [sensitivity 82.76%, specificity 86.67%, and area under the receiver operating characteristic curve (AUC) 0.94] for differentiating vulnerable subjects from resilient subjects. Brain areas that contributed most to the classification model were mainly located within the sensorimotor network, default mode network, and thalamus. Furthermore, we found a significantly negative correlation between changes in PVT lapses and DC in the thalamus after SD. These findings suggest that resting-state network measures combined with a machine learning algorithm could have broad potential applications in screening vulnerability to SD.