Depression is one of the most common mental disorders that greatly affects a person's normal life. The wide range of symptoms of depression, as well as the similarity of its symptoms to other disorders, sometimes makes it difficult to diagnose. By diagnosing this disorder in its early stages, recovery is easier and adverse events such as self-harm and suicide are prevented. Previous studies have shown that brain signals are successful in diagnosing depression. Most previous studies used a large number of electrodes, did not conduct external validation, and used deep learning methods to achieve high accuracy. These methods have a high computational load and require powerful processors. They also act like a black box; meaning that they do not allow for interpretation of mechanisms. Reducing the computational load of diagnostic models based on artificial intelligence methods and making them interpretable helps these studies become applicable in clinical activities and accepted by the medical community. In this study, depression diagnosis was performed using 8 electrodes and machine learning methods for data recorded in the resting state with eyes closed in two datasets: MODMA and the local IUTDD dataset. The IUTDD dataset was collected in the Medical Engineering Laboratory of Isfahan University of Technology and the data recording method was similar to the MODMA dataset. The validation methods used in previous studies can be divided into two categories: department-based and individual-based. In this study, these methods were compared with each other. The accuracy of depression diagnosis in the MODMA and IUTDD datasets with the LDA classifier was 92.80% and 94.55% in the department-based method and 95.00% and 91.36% in the individual-based method, respectively. By adding a voting threshold in both methods, the accuracy of external assessment increases to 95.45%. Also, with the help of dynamic analysis of external eva‎luation results, the classification accuracy reaches 95.00% with the segment-based method and 94.55% with the person-based method. Therefore, choosing an appropriate combination of features in the input of machine learning models and using appropriate post-processing methods can increase the classification accuracy to a desirable level, while only 8 channels were used and the signal was recorded in the resting state. It also seems that the results obtained in studies that used a dataset for training and eva‎luating the model are more reliable with the person-based method. The diagnosis of depression severity is used in selecting appropriate treatment and eva‎luating the effects of treatment and has been less considered in previous studies. In this study, the diagnosis of depression severity was made with two multi-class classification methods and also the estimation of depression scores based on two valid psychological questionnaires. The level of depression was predicted using the multi-class SVM classifier in the MODMA and IUTDD datasets with an accuracy of 92.18% and 95.05%, respectively, and has increased significantly compared to previous studies. Also, with the KNN and RF regression methods, the depression score based on the two questionnaires PHQ-9 and BDI-II was estimated with an RMSE of 3.08±0.14 and 5.00±0.19, and has performed better compared to similar previous studies. The use of the intraclass regression method significantly reduces the estimation error. ...

فرزانه شايق بروجني

ماه گل توكلي

رسول امير فتاحي ورنوسفادراني , جلال ذهبي