Towards Extracting Medical Family History from Natural Language Interactions: A New Dataset and Baselines

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 1255–1260, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 1255 Towards Extracting Medical Family History from Natural Language Interactions: A New Dataset and Baselines Mahmoud Azab1, Stephane Dadian1, Vivi Nastase2, Larry An1, Rada Mihalcea1 1University of Michigan, Ann Arbour, Michigan, mazab@umich.edu, sdadian@umich.edu lcan@umich.edu, mihalcea@umich.edu 2University of Heidelberg, Heidelberg, Germany nastase@cl.uni-heidelberg.de Abstract We introduce a new dataset consisting of natu- ral language interactions annotated with med- ical family histories, obtained during interac- tions with a genetic counselor and through crowdsourcing, following a questionnaire cre- ated by experts in the domain. We describe the data collection process and the annotations performed by medical professionals, including illness and personal attributes (name, age, gen- der, family relationships) for the patient and their family members. An initial system that performs argument identification and relation extraction shows promising results – average F-score of 0.87 on complex sentences on the targeted relations. 1 Introduction Individual family history is an important medical tool for diagnosis, disease prevention and treat- ment. Macro analyses of family histories provide useful and important information for determining disease susceptibility by researching genetic influ- ences between family histories and diseases. This can be used to determine levels of risk for certain illnesses for a given family member, and poten- tially develop a protocol of preventative screen- ing. In-person appointments with genetic coun- selors are time-consuming and suffer from limi- tations such as long wait times and small num- ber of genetic counseling centers (Sutphen et al., 2010; Nathan et al., 2016). Moreover, collecting a patient’s family history away from pressured con- sultation environment results in more accurate and complete information (Nathan et al., 2016). An interesting solution would be the deploy- ment of a dialog agent to collect such data. The interaction between the agent and the patient can vary between two extremes: (a) pose very spe- cific questions about each member of the family and their history – difficult because of the need to identify very clearly the family relations in real time during the interaction, but easier because the answers will be short and focused; (b) a sys- tem that asks more generic questions, and elicits more narrative answers from the patients – eas- ier because the interaction with the patient is more generic and does not have to be adjusted in real time, but harder because the elicited answers are longer and more free-form. To support research in this area, we choose to explore the second end of the spectrum, which favours a simpler interac- tion and more intense processing for relation ex- traction. We introduce a dataset consisting of free- form natural language responses to a medical can- cer family history questionnaire created by experts in the domain. The dataset consists of 20,774 an- notated relations (illness, name, age, gender, fam- ily relationships) in 228 family histories answer- ing 159 questions each, collected through …

Content truncated to 500 words