CMTN: A Convolutional Multi-Level Transformer to Identify Suicidal Behaviors Using Clinical Notes

Suicide has become a significant cause of concern worldwide over recent years. The early identification and providing treatment of individuals having suicidal tendencies are necessary for preventing suicides. Past suicidal behavior information of an individual is recorded in the electronic health records (EHR) reports which can help to understand a patient’s current mental health condition. In this paper, to identify the people who are ideating and are anticipating attempting suicide, we propose a novel model named CMTN, which utilizes the textual EHR data for the prediction of suicidal behaviors. The proposed framework employs convolutional and transformer layers to capture local and global relationships in the text and the attention mechanism to assess the significance of various input text components. Overall, the suggested model has achieved the highest precision for the SA class with a score of 0.97 and the highest recall and f1-score of 0.56 and 0.52, respectively, for the SI class, compared with all other state-of-the-art and baseline models. We have also employed different embeddings such as BERT, BioBERT, and PubMedBERT to our state-of-the-art model and illustrated the model’s improved performance. In addition, we have also shared the data alignment and annotation extraction algorithms in this paper, allowing other researchers to generate the dataset, thereby expediting development in the prevention of suicides.