Research On Depression Identification Method Based On Decision-level Fusion

Major depressive disorder(MDD)is a common mental illness that affects over 300 million people worldwide.Diagnosis of MDD is often subjective and uncertain.Studies have shown that MDD patients have characteristic differences in brain electrical signals compared to normal individuals,and that brain electrical signals have advantages such as non-invasiveness,low cost,and ease of operation.Researchers have therefore begun to apply brain electrical signals to the identification of MDD.Currently,extracting brain electrical features from multiple perspectives,constructing classifiers,and fusing decision information from multiple classifiers at the decision level is an effective method for improving classification performance.However,how to better fuse decision information from multiple classifiers at the decision level is a difficult problem that urgently needs to be addressed.Therefore,this article proposes two decision-level fusion methods from the perspective of decision-level fusion.The main contents are as follows:1)A decision-level fusion method based on SVM classification score(Sc-score)is proposed.Firstly,we extract brain electrical features from multiple perspectives.Secondly,we train SVM classifiers based on the extracted features,obtain the classification scores for each feature,construct the feature vector by the classification scores at the decision level,and finally input the feature vector into the SVM classifier again to obtain the final classification result.Our method can fully utilize the discriminant results of different feature classifiers because it can convert the decision information of each classifier into the probability of classification,thereby overcoming the problem of the voting method’s inflexibility in assigning weights to different classifiers.2)A decision-level fusion method is proposed by constructing credibility.Inspired by the k-nearest neighbor algorithm,we use the distance between the test sample and the nearest neighbor to construct a new credibility.The advantage of this approach is that it can solve the problem of some extreme test samples being too far from the SVM partition hyperplane during the classification process,resulting in less than ideal classification results.Specifically,we train a classifier for each extracted feature set,calculate a new credibility using the Sc-score and the distance between the nearest neighbor sample,construct a new credibility feature vector at the decision level,and use it for the final classification.Compared with the Sc-score-based decision-level fusion method,this method further improves classification performance.In conclusion,we propose two decision-level fusion methods by processing and analyzing brain electrical signals,which fully utilize the differences in brain electrical signal features and the complementarity between classifiers.In classification experiments on MDD patients and healthy populations,both decision-level fusion methods outperform traditional fusion methods.Based on the experimental results,the effectiveness of the proposed methods is demonstrated,providing new ideas for future assisted diagnosis of MDD.