Using Deep Neural Networks with Intra- and Inter-Sentence Context to Classify Suicidal Behaviour

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 1303–1310 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 1303 Using Deep Neural Networks with Intra- and Inter-Sentence Context to Classify Suicidal Behaviour Xingyi Song1, Johnny Downs2, Sumithra Velupillai2, Rachel Holden2, Maxim Kikoler2, Kalina Bontcheva1, Rina Dutta2, Angus Roberts2 1Department of Computer Science, The University of Sheffield Sheffield, UK {x.song, k.bontcheva}@sheffield.ac.uk 2Maudsley Biomedical Research Centre, Institute of Psychiatry, Psychology and Neuroscience, King’s College London London, UK {johnny.downs, sumithra.velupillai, rina.dutta, angus.roberts}@kcl.ac.uk Abstract Identifying statements related to suicidal behaviour in psychiatric electronic health records (EHRs) is an important step when modeling that behaviour, and when assessing suicide risk. We apply a deep neural network based classification model with a lightweight context encoder, to classify sentence level suicidal behaviour in EHRs. We show that incorporating information from sentences to left and right of the target sentence significantly improves classification accuracy. Our approach achieved the best performance when classifying suicidal behaviour in Autism Spectrum Disorder patient records. The results could have implications for suicidality research and clinical surveillance. Keywords: Suicide Detection, Neural Network, Contextual Classification, Document Encoding 1. Introduction Suicide is the second leading cause of death among 15 to 29 year-olds, causing about 800,000 deaths per year world- wide (WHO, 2018). Mental health problems are a major risk factor for suicide attempts. For example, among the people who attempted (or completed) suicide in the US in 2010, approximately 44% were diagnosed with a men- tal health problem, and 31% were receiving mental health care (Parks et al., 2014). The ubiquity of Electronic Health Records (EHRs) in many countries, and the potential for the reuse of information contained within them, has led to attempts to model and predict patient suicide risk from the content of those records. In building such risk models, it is necessary to extract information about suicidal behaviour from the EHR. In common with much of the information in the mental health record, this is generally recorded in the free text, unstructured portion of the records. When applying natural language processing to EHRs, the EHR can be analysed at different levels of granularity (Velupillai et al., 2018). Considering these levels of gran- ularity for detection of suicidal behaviour, previous ap- proaches can be categorised into patient level - a set of EHRs that belong to a certain patient who has exhibited or exhibits suicidal behaviour (Barak-Corren et al., 2017; Tran et al., 2015; Walsh et al., 2017; Fernandes et al., 2018; Choi et al., 2018; Cook et al., 2016); mention level - word or sub-sentences that mention suicidal thoughts or behaviour (Haerian et al., 2012; Anderson et al., 2015; Downs et al., 2017; Gkotsis et al., 2016b); and combinations of these into document and patient level - aggregating mention level annotations to derive document- and patient-level la- bels (Velupillai et al., 2019). To the best of our knowledge, none of these previous ap- proaches have addressed this problem on a sentence level. There …

Content truncated to 500 words