Post-traumatic stress disorder (PTSD) is a neurobiological disorder that may occur after exposure to a traumatic event. Electroencephalography (EEG) studies have shown PTSD patients were characterized by partial or whole-brain dysrhythmia. Besides, in animal models, theta (4Hz) activities were found highly correlated with fear expression. Nowadays, the application of machine learning (ML) algorithms allows us to perform detection and prediction with emotion-related biomarkers. Here, we hypothesized it is possible to use ML methods to detect and predict fear expression with several critical brain wave features in the PTSD animal model. Brain wave signals associated with fear expression were divided into early (10mins, 30mins, 2hrs, 4hrs ,6hrs) and late (day 1, 3, 7 and 14) time points after traumatic stress exposure, and signals were collected by in vivo local field potentials (LFPs) recording from the bilateral amygdala (AMY), medial prefrontal cortex (mPFC) and ventral hippocampus (vHipp), total of six recording sites. We analyzed signals with four ML methods, they are Fine Tree, Linear SVM, Fine Knn and Bagged Trees. Here we found Bagged Trees performed the best for detecting the presence of freezing behavior in all recording sites. Besides, we were able to distinguish freezing behavior in the PTSD model and immobility in the control group with ML methods, and we found different learning methods show a similar tendency with high accuracy in the same brain area. Last, we found signals from the AMY performed the best for predicting freezing behavior. And we conclude that the brain wave signals from the AMY with the Bagged Trees method could be the best combination for the detection and prediction of freezing behavior in the PTSD rat model.

Keywords: machine learning, post-traumatic stress disorder, fears, fear detection, fear prediction