Abstract
Bipolar disorder (BD) is a mental disorder with a major characteristic of extreme instability in emotion caused by repeated depression and manic episodes. Long observation time is needed before a psychiatrist diagnoses BD to avoid misdiagnosing it as depression. Recently, deep learning models have been used for detecting BD and depression. However, because of the limitations of accessing healthcare data, such as the privacy of the data, it is difficult to collect large amounts of data to train the models. Alternatively, studies on detecting BD and depression were conducted using social media data because people often express their emotions in social media posts. The posts can be represented as text features using word embedding models, such as bidirectional encoder representations from transformers (BERT). Recently, transformer-based word embedding models were fine-tuned using biomedical, clinical, and mental health domain data, improving the performance of detecting depression. However, the potential of those fine-tuned word embeddings specific for detecting BD has yet to be explored. In this study, we compared some fine-tuned word embeddings, analyzed the impact of using different domain knowledge in the fine-tuning process for detecting BD, and identified the limitation of using word embedding in detecting BD. Based on our experiments, the models achieve excellent performance with an F1-Score above or equal to 0.90. However, the performance of these models is not significantly different unless more computationally expensive models are used in the fine-tuning process of the word embedding models. Our experiment results also show that different types of domain knowledge in the fine-tuning process have no significant impact on detecting BD. Additionally, the instability in emotion cannot be identified and explained from the posts using word embeddings, decreasing the model's reliability and limiting the use of word embeddings for BD detection.