Depression and bipolar disorder (BD) are mental disorders that are often misdiagnosed as each other because their symptoms often overlap. Depression is characterized by prolonged negative emotion, which is persistent feelings of sadness. In contrast, BD is characterized by repeated extreme changes of emotions, consisting of manic (very happy) and depression (very sad) episodes. Because of overlapping symptoms, changes of emotions indicating depression or BD are difficult to recognize from individual posts. We propose two deep learning models to detect depression and BD from user posts, considering individual posts and multiple posts. A novel method to extract emotion features is designed by fine-tuning a transformer model to extract the intensity of emotions of the posts. Text features are extracted using word embedding models that have been fine tuned on mental health data, and we further consider topic features by using a topic modeling method. Multiple posts are considered by grouping them based on a time interval between the posts. The features of each group are concatenated, input into a convolution neural network (CNN), and go through a long short-term memory (LSTM). The attention mechanism is used to pay attention to important groups. The most important groups can be observed further to explain why the model classified depression and BD users. Our proposed model outperforms other models, achieving an F1-Score of 0.9589. Based on our experiment, the changes of emotions within four days are necessary for distinguishing depression and BD, which aligns with considerations in diagnosing symptoms of depression and BD.

Doctors prescribe drugs for the patient with the objective of curing the patient. Some drugs cannot be consumed together since doing so may cause negative effects. This can be avoided by knowing the effects caused by consuming combinations of drugs. However, for complex cases of a patient, it can be difficult to decide the best combination of drugs. Therefore, automatic drug recommendation method was used to recommend drugs with minimal negative effects. It is performed by using a deep learning model which is trained on drug data. A graph called drug-drug interaction (DDI) is used to represent the drugs and effects of consuming one drug with other drugs. Additionally, information about the combination of drugs prescribed in the past for a patient is also important for drug recommendation. It can also be represented as a graph called drug concurrence relation (DCR). The DDI and DCR graphs can be input to the deep learning model through an encoding process. In this paper, we propose a graph encoding-enhanced transformer (GEET) to recommend drugs. The DDI and DCR graphs are encoded by using Graph Attention Network (GAT). The graph encoding model has multi-head attention, which makes the GEET model aware of the most important DDI and DCR from the graphs. Additionally, the encoding outputs are combined, and activation function and normalization methods are used to improve the performance. The model has been evaluated on the publicly available MIMIC-III dataset and has the best results on F1, Jaccard and PRAUC scores compared to the models proposed by the existing related research papers.