PsyBERTpt: A Clinical Entity Recognition Model for Psychiatric Narratives

Abstract

Mental disorders are among the most complex disorders to treat due to the scarcity of biomarkers that identify and quantify the severity of the disease, as is commonly available in other areas of medicine. The practice of psychiatry uses semi-structured and unstructured data to record the mental and behavioral states of patients, which are impressions of the physician about the patient, and therefore important information for prognosis. Most of this data lacks standardization, making it difficult to use for quantitative analysis through computational tools since clinical decision models are based on structured data. In this work, a team of psychiatrists and computer scientists developed a methodology based on Natural Language Processing to extract relevant information from admission clinical notes of a psychiatric emergency service. With the use of BERT, we developed psyBERTpt, a prediction model capable of extracting multiple types of information considered relevant to psychiatric practice.