Detecting Cannabis-Associated Cognitive Impairment Using Resting-State fNIRS

Abstract

Functional near infrared spectroscopy (fNIRS), an emerging, versatile, and non-invasive functional neuroimaging technique, promises to yield new neuroscientific insights, and tools for brain-computer-interface applications and diagnostics. In this work, we consider the novel problem of detecting cannabis intoxication based on resting-state fNIRS data. We examine several machine learning approaches and present an innovative data augmentation technique suitable for resting-state functional data. Our experiments suggest that a recurrent neural network model trained on dynamic functional connectivity matrices, computed on sliding windows, coupled with the proposed data augmentation strategy yields the best accuracy for our application. We achieve up to 90$\%$ area under the ROC on cross-validation for detecting cannabis associated intoxication at the individual-level. We also report an independent validation of the best performing model on data not used in cross-validation.