Research On Feature Extraction And Classification Of Obsessive-Compulsive Disorder Patients Based On Resting State EEG

Obsessive-compulsive disorder(OCD)is a clinically heterogeneous psychiatric disorder.According to the fourth edition of the American Diagnostic and Statistical Manual of Mental Disorders,according to the level of insight,OCD can be divided into two categories: good insight and poor insight.Different OCD patients have distinct genetic backgrounds and neurobiological bases,making the clinical diagnosis and determination of insight levels challenging.With the development of modern medical technology,researchers have been using electroencephalogram(EEG)or magnetic resonance imaging(MRI)for OCD prediction and identification,making it a highly focused research area.However,most pattern recognition studies for OCD currently utilize features from single modalities,which limits the discriminative ability of features.In this study,we developed a multimodal feature fusion method based on EEG data,aiming to achieve reliable identification of OCD patients and healthy controls,as well as patients with poor and good insight.(1)We analyzed single-channel EEG features,extracting power spectra,approximate entropy,and sample entropy features for classification of OCD patients and healthy controls,as well as patients with poor and good insight.The results showed that the discriminative ability based on approximate entropy features was the best,with classification accuracies of 79.07% and 82.26% for OCD patients vs.healthy controls and poor insight vs.good insight patients,respectively.(2)We then turned our attention to brain networks,using coherence networks,transfer entropy networks,and independent effective coherence networks to analyze EEG signals.Among these three networks,the network spatial features of independent effective coherence networks demonstrated the best classification performance,with classification accuracies of 74.49% and 87.10% for OCD patients vs.healthy controls and poor insight vs.good insight patients,respectively.(3)Through correlation-based feature selection,we combined approximate entropy and network spatial features of independent coherence networks from all frequency bands for classification.This method can effectively differentiate between OCD patients and healthy controls,as well as patients with poor insight and good insight,achieving classification accuracies of 85.71% and 96.77%,and F1 scores of 88.33% and 90.91%,respectively.This result indicates that the fused features of approximate entropy and independent effective coherence network spatial patterns can serve as reliable discriminators for differentiating OCD patients from healthy controls,as well as patients with poor and good insight.In this study,approximate entropy provided information on the complexity of individual channels in brain regions,while the network spatial features of independent effective coherence networks offered direct causal information on whole-brain interactions.Therefore,we fused these two features to achieve more accurate recognition results,providing a new idea for pattern recognition of OCD patients and different insight level OCD patients.At the same time,by extracting differential EEG features,we also explored the pathological mechanisms of OCD and insight level variations of OCD.