Deep Depression Detection Based on Feature Fusion and Result Fusion

Hua Gao, Yi Zhou, Li Chen, Kaikai Chi

Abstract: Depression, as a severe mental disorder, has significant impacts on individuals, families, and society. Accurate depression detection is of great significance. To this end, we propose a deep depression detection based on feature fusion and result fusion. We have introduced a dataset consisting of answers that correspond to four questions. Using the answers to the first three questions provided by the patients, we propose a text feature extraction method based on a pre-trained BERT model and an improved TextCNN model to extract features and introduce an attention mechanism for feature fusion. The vector obtained by fusing the features of the first three questions is inputted into a classifier to obtain classification results. These results, along with the classification results obtained from the fourth question, are fused using result fusion techniques to obtain the final outcome. To evaluate our method, we conducted experiments on the dataset collected from the specialized hospital for mental illness. We compare our depression detection algorithm with mainstream ones. Our method achieves an accuracy of 94.2% on our dataset. The results show that our method achieves an accuracy rate of 7.7% higher than the highest accuracy rate among the compared mainstream models.