Predicting Openness of Communication in Families With Hereditary Breast and Ovarian Cancer Syndrome: Natural Language Processing Analysis

Abstract

Background In health care research, patient-reported opinions are a critical element of personalized medicine and contribute to optimal health care delivery. The importance of integrating natural language processing (NLP) methods to extract patient-reported opinions has been gradually acknowledged over the past years. One form of NLP is sentiment analysis, which extracts and analyses information by detecting feelings (thoughts, emotions, attitudes, etc) behind words. Sentiment analysis has become particularly popular following the rise of digital interactions. However, NLP and sentiment analysis in the context of intrafamilial communication for genetic cancer risk is still unexplored. Due to privacy laws, intrafamilial communication is the main avenue to inform at-risk relatives about the pathogenic variant and the possibility of increased cancer risk. Objective The study examined the role of sentiment in predicting openness of intrafamilial communication about genetic cancer risk associated with hereditary breast and ovarian cancer (HBOC) syndrome. Methods We used narratives derived from 53 in-depth interviews with individuals from families that harbor pathogenic variants associated with HBOC: first, to quantify openness of communication about cancer risk, and second, to examine the role of sentiment in predicting openness of communication. The interviews were conducted between 2019 and 2021 in Switzerland and South Korea using the same interview guide. We used NLP to extract and quantify textual features to construct a handcrafted lexicon about interpersonal communication of genetic testing results and cancer risk associated with HBOC. Moreover, we examined the role of sentiment in predicting openness of communication using a stepwise linear regression model. To test model accuracy, we used a split-validation set. We measured the performance of the training and testing model using area under the curve, sensitivity, specificity, and root mean square error. Results Higher “openness of communication” scores were associated with higher overall net sentiment score of the narrative, higher fear, being single, having nonacademic education, and higher informational support within the family. Our results demonstrate that NLP was highly effective in analyzing unstructured texts from individuals of different cultural and linguistic backgrounds and could also reliably predict a measure of “openness of communication” (area under the curve=0.72) in the context of genetic cancer risk associated with HBOC. Conclusions Our study showed that NLP can facilitate assessment of openness of communication in individuals carrying a pathogenic variant associated with HBOC. Findings provided promising evidence that various features from narratives such as sentiment and fear are important predictors of interpersonal communication and self-disclosure in this context. Our approach is promising and can be expanded in the field of personalized medicine and technology-mediated communication.