With the increase of social pressure, the prevalence of depression is on the rise. Traditional self-rating depression scale diagnosis methods are often affected by patients’ emotions and environmental factors, resulting in a high misdiagnosis rate. In order to improve the accuracy of diagnosis, this paper proposes a classification method of depression based on mixed neural network model. First, the scalp EEG signal is filtered and independent component analysis is carried out to get a clearer signal. The power spectral density characteristics of these signals were extracted by Welch method. Secondly, a hybrid neural network model is designed by combining convolutional neural network, long short-term memory network and attention mechanism. In this paper, we use the method of 5-fold cross-validation to divide the data set and input the power spectral density characteristics into the designed model. The experimental results show that the accuracy of the method is 96.93%, the accuracy is 96.71%, the recall rate is 97.76%, and the F1 score is 97.20%. Compared with traditional diagnostic methods, the hybrid neural network model has better classification performance.