A Novel Method for Identifying Bipolar Disorder Based on Diagnostic Texts

Hua Gao, Li Chen, Yi Zhou, Kaikai Chi, Sixian Chan

Abstract: Bipolar disorder is a complex condition characterized by episodes of mixed manic and depressive states, exhibiting complexity and diversity in symptoms, with high rates of misdiagnosis and mortality. To improve the accuracy of automated diagnosis for bipolar disorder, this study proposes a text-based identification method. Our approach focuses on two extreme emotional features, utilizing two temporal networks in the emotion feature module to extract depressive phase features and manic phase features from the text. Simultaneously, mixed-dilated convolutions are introduced in TextCNN to extract local features with a larger receptive field. By integrating feature information captured from different perspectives, we construct a multi-scale feature model that emphasizes both state features. We utilized a self-collected dataset comprising symptom descriptions of bipolar disorder patients from hospitals, achieving an accuracy of 92.5%. This work provides an accurate assessment of bipolar disorder, facilitating individuals to gain a rapid understanding of their condition and holds significant social implications.