Research On Depression Recognition Based On Eye Movement Information

Depression is a common mental illness that is characterized by a prolonged period of low mood,from initial depression to final anguish.Unlike ordinary mood swings that only affect people briefly,depression can seriously affect patients’ daily lives and can even lead to suicide.In recent years,the prevalence of depression has been increasing year by year,and early detection of the disease can be of great help in the subsequent treatment of patients,but the current common diagnostic methods suffer from the difficulty of early identification.Eye movement data have been widely used in the field of depression detection because they are easy to record and can visually reflect psychological processes.In this thesis,we propose Feature sequence Convolution Attention(Fe CA)and Encoder Spatial Attention(En SA)models based on eye movement features for analyzing the abnormalities of depressed patients to assist doctors in diagnosis.The main work of the thesis is as follows:(1)Data collection and dataset construction.This thesis was divided into two experiments,in each experiment there were 158 subjects and 74 subjects respectively,where the number of depression group and control group were equal.All subjects completed the eye-movement smooth tracking experiment and the forward and backward saccade experiment according to the experimental requirements,and the data were collected by the researcher using an eye-tracking device.After the data were exported from the device,the data were parsed by IDF Converter software,and the original eye movement data were selected to build the original eye movement feature data set.(2)Research on depression recognition based on eye movement smoothing tracking experiment.In this thesis,the eye movement data of subjects tracking the target trajectory were recorded by an eye-tracking device,and based on this eye-movement sequence data,the Fe CA model was proposed for depression recognition,introducing the self-attention mechanism into the field of eye-movement depression recognition.By feature engineering the eye movement time series data,the time series predictive classification task is transformed into a window-based predictive classification problem,the attention weights of the feature series are obtained,and depression recognition is performed based on self-attention.(3)Research on depression recognition based on prosaccade and antisaccade experiments.The subjects’ saccadic responses were detected by setting forward and backward target stimuli,and the subjects’ eye-movement data were recorded.Based on this experiment,a model combining self-attention and spatial attention,En SA,was proposed.We obtained multi-headed attention information by performing attention calculation on the eye-movement data,and continued to calculate spatial attention weights accordingly,and combined the two for depression recognition.Finally,this thesis uses the idea of stacking method to fuse the prosaccade and antisaccade model as the base model and XGBoost as the meta-model into the eye-movement depression recognition model.The experimental results show that in the experiment based on eye movement smoothing tracking,the recognition rate of the model on this dataset reached 90.03%.The En SA model performed the best in the prosaccade and antisaccade experiments,with a recognition accuracy of 93.5% in the prosaccade experiment and 95.5% in the antisaccade experiment,and the accuracy of the antisaccade experiments was generally higher than that of the prosaccade.Meanwhile,the model achieved a recognition rate of98.25% for the dataset.