This analysis is part of a larger ongoing study, the Swiss CASCADE cohort, which follows adult (aged ≥18 years) men and women from families that harbor pathogenic variants associated with HBOC or LS. The cohort includes individuals who had genetic testing, confirming either the presence or the absence of the familial pathogenic variant, and their untested relatives with unknown mutation status. Eligible participants may have had a cancer diagnosis, or they could be cancer-free at the time of enrolment in the study. Recruitment takes place at 8 different oncology and genetic testing centers in the German-, French-, and Italian-speaking regions of Switzerland. The study collects survey data designed to elicit factors that enhance cascade genetic testing and cancer surveillance for HBOC and LS. A subsample of participants has consented to provide narrative data regarding family communication of test results. For the purposes of this paper, we focused only on individuals who have had genetic testing for HBOC-associated pathogenic variants and accepted to provide narrative data.

The study protocol has been approved by the Ethics Committee of Northwest Switzerland (BASEC 2016-02052) and is publicly available (ClinicalTrials.gov NCT03124212) [30]. We also used available data from participants in the K-CASCADE study (ClinicalTrials.gov NCT04214210) in South Korea, which focuses on HBOC. K-CASCADE and the collaboration of the 2 studies has been approved by local ethics committees (Severance Hospital Institutional Review Board: 4-2020-0520). K-CASCADE is identical to the Swiss CASCADE in respect to scope, research design, participant eligibility criteria (except for age ≥19 years), and data collection methodology. Participants to K-CASCADE are recruited from 5 hospitals in South Korea [31].

Narrative data included in this paper were collected from 44 individuals living in Switzerland and 9 in South Korea. The in-depth interviews were conducted between April 2019 and June 2021 either face to face or online (after April 2020 due to the COVID-19 pandemic) by trained research staff in German, French, Italian, English, and Korean using the same interview guide. Interview questions were designed to explore general communication patterns within family networks and specific experiences and barriers of family communication regarding genetic risk including discussions with health care providers. Examples of questions included in the interview guide are “What are some issues (barriers) that people might experience, related to sharing genetic risk information with family members?” and “Think of your own experience of (not) sharing genetic risk information with family members. What did you do and how did you decide about it?” Interviews were recorded, and all narrative data were transcribed verbatim in the original language in Microsoft Word and translated into English for this paper.

Survey data were collected on an ongoing basis, starting in fall 2017 and occurring approximately 18-24 months apart. Self-administered surveys assessed demographic and clinical characteristics [30]. The surveys also included investigator-developed items that have been associated with family communication and intention to inform relatives about genetic cancer risk. These items assess informational support among family members, preference for patient-mediated communication of genetic testing results, and perceived utility of genetic testing for relatives (Textbox 1). These items are scored on 7-point Likert-type scales ranging from 1 “Strongly Disagree” to 7 “Strongly Agree.” Respondents also completed the Informing Relatives Inventory (IRI), a 37-item scale assessing knowledge, motivation, and self-efficacy to disclose genetic cancer risk to relatives [32]. IRI items are also scored on a 7-point Likert-type scale, with higher overall score indicating greater intention to inform relatives about genetic cancer risk.

First, we examined narratives to assess “openness of communicating” genetic test results and cancer risk with relatives and with health care providers. Second, we categorized the text of each narrative as describing either a positive or negative sentiment toward experiences with genetic testing and health care services. Third, we examined whether demographic and clinical characteristics and sentiment, as expressed in the narrative, can predict “openness of communicating” genetic risk from tested individuals to relatives.

The ability of NLP to identify and predict different levels of “openness of communication” was evaluated following a multistep framework (Figure 1), which was divided into three phases: (1) preprocessing, (2) training, and (3) performance evaluation. All computations were performed in R software (version 3.6.3; R Foundation for Statistical Computing) [33]. We have made our analysis publicly available through the Zenodo open data repository [34].

To start data processing, we broke down each text into individual tokens. We then applied functions to remove stop words and special characters. All texts were converted to lower case. We also applied part-of-speech tagging to extract phrases from the text corpus, used a latent Dirichlet allocation model to generate the most appropriate topics, and computed the term frequency–inverse document frequency to indicate the significance of a word in the text corpus [35,36].

For developing the model, the overall data set was split randomly, with 70% of data used in the training phase by using the “openness of communication” score as the dependent variable. To examine whether the demographic and clinical characteristics and sentiment features of each narrative predicted “openness of communication” scores, we used a linear regression model based on the following steps. Initially we performed a univariate analysis to identify those independent variables exhibiting more than 60% absolute correlation with one another. These variables were excluded to avoid multicollinearity. Then, we continued with a multivariate analysis using a stepwise linear regression to identify possible predictors of the dependent variable and remove nonsignificant independent variables. As an alternative model, we attempted to use an artificial neural network. We built a fully connected network with 1 hidden layer, 1 input and 1 output layer, and 5 neurons. Optimization was done through the Broyden-Fletcher-Goldfarb-Shanno method. Early stopping was utilized to avoid overfitting. However, we ended up discarding the artificial neural network from the analysis because it showed no improvement compared to the linear regression. Finally, the performance of the models was evaluated using the area under the curve (AUC), sensitivity, specificity, and root mean square error (RMSE).

In this phase, we tested the model using the remaining 30% of the database (validation cohort). The performance of the models was evaluated using the same metrics as in the training phase, ie, AUC, sensitivity, specificity, and RMSE.

Narrative and survey data from 53 individuals are included in this paper. Participants were aged 32-76 years. Most were female (47/53, 89%), married (41/53, 77%), and carriers of the familial pathogenic variant (51/53, 96%). Approximately 2 in 3 (32/53, 60%) had a prior diagnosis of cancer (Table 1). The Swiss and the Korean samples were not statistically different in respect to age (P=.71), prior cancer diagnosis (P=.38), educational level (P=.17), and employment status (P=.14).

The average “openness of communication” score was 29.8 (SD 19.5; range –9 to 76), indicating an overall trend toward open communication. Narratives from these 53 individuals included 5837 unique unigrams, 4183 bigrams, and 654 trigrams. The most frequently appearing nontrivial words are shown in Figure 2. Based on the NRC Emotion Lexicon, the 10 most common positive words were “time,” “true,” “children,” “talk,” “finally,” “information,” “positive,” “doctor,” “understand,” and “daughter”. The 10 most common negative words were “cancer,” “sick,” “feel,” “risk,” “negative,” “died,” “difficult,” “fear,” “disease,” and “bad.”

The correlation coefficient between the “openness of communication” score and IRI in the Swiss data was r=0.46, indicating a moderate positive correlation.

Attitude toward genetic testing and health care services varied among participants, but it was overall positive. “Trust” appeared as the strongest positive emotion, whereas “fear” and “sadness” appeared as the strongest negative emotions in the text corpus based on the NRC Emotion Lexicon. The least perceived emotions were “surprise” and “anger.” Figure 3 describes the frequencies of words identified in the corpus for each emotion.

The R² for the overall model was 0.87 (adjusted R²=0.85; P<.001). A stepwise linear regression identified 5 significant predictors of “openness of communication” score, ie, the overall net sentiment of the narrative and fear, which were obtained based on the NRC Emotion Lexicon; informational support among family members; educational level; and being single (Table 2). Specifically, findings showed that both the higher overall net sentiment score of the narrative (P=.007) and also greater fear (P=1.97 × 10^–5) were strongly associated with higher “openness of communication” scores. There was a positive correlation between “openness of communication” score and the statement “In our family when I have a health problem there is great willingness to share information with each other” (P=.005). Participants with nonacademic education were also more likely to communicate genetic risk with their relatives (P=.02). Lastly, there was a positive correlation between being single and “openness of communication” scores (P=.047).

The predictive accuracy of the model using a stepwise linear regression for the training and testing data sets reached 0.85 (AUC=0.92, specificity=0.86, and sensitivity=0.82) and 0.72 (AUC=0.69, specificity=0.62, and sensitivity=0.83), respectively. Figure 4 presents the receiver operating characteristic curves that visualize the accuracy improvement between the training and testing data sets applying linear regression.

The predicted values are plotted against the target values and are shown on a scatter plot for the linear regression model (Figure 5). The linear regression model achieved RMSEs of 11.76 (training data set) and 16.04 (testing data set). In this case, our model performs more accurately when it yields lower values of RSME.

We analyzed 53 narratives regarding intrafamilial communication of genetic cancer risk associated with HBOC. NLP enabled the analysis of unstructured narratives from different languages and identified the most frequently used words or combination of words describing openness in family communication of genetic cancer risk. This was the first study in which we applied NLP and sentiment analysis to better understand factors driving open intrafamilial communication regarding genetic cancer risk. Our findings showed that sentiment plays a crucial role and that emotions are a pervasive feature that predict intrafamilial communication in this particular population. Sentiment analysis performed on all interviews provided scores demonstrating positive or negative emotional valence, which were highly predictive of the direction of intrafamilial communication in this context. The higher overall net sentiment scores predicted greater openness in intrafamilial communication, whereas the lower overall net sentiment scores predicted closed or absent communication. This finding provides insights consistent with social penetration theory related to self-disclosure of carrying a cancer-causing genetic variant [41,42]. The depth of self-disclosure, ie, the degree to which the individual reveals personal and private information involving unusual traits and painful memories, reflects the degree of intimacy of a relationship. In the context of HBOC intrafamilial communication, self-disclosure of personal genetic information may be opposed by the desire to retain privacy and to avoid creating uncertainty and unpredictability in interpersonal relationships. Anticipating future negative emotions, such as regret or conflict, categorizes genetic risk information as a considerable emotional threat [43]. This finding was captured in our analysis as the overall net sentiment of each narrative, and its predictive value was confirmed based on the performance of our models. Taken together, findings indicate that sentiment can be used to frame genetic cancer risk as an opportunity for proactive risk reduction and for enhancing technology-mediated HBOC intrafamilial communication.

Our linear regression model explained more than 80% of the variance in openness of communication and achieved good performance in both the training and the testing samples. Our findings show that NLP was highly accurate in analyzing unstructured narratives from individuals of different cultural and linguistic backgrounds (Swiss German, French, Italian, English, and Korean) and in quantifying openness of communication in intrafamilial discussions about genetic cancer risk. The “openness of communication” scores were also validated against IRI. IRI was developed on the premise that increased genetic knowledge, positive motivation, and increased self-efficacy are prerequisites of increased intention to inform relatives about genetic risk. Although “intention to inform relatives” is closely related to “openness of communicating genetic risk,” the 2 concepts are not identical, which was also confirmed in our data with a moderate positive correlation between the 2 scores. An individual may have high intention to inform relatives about their genetic risk despite difficulties in communication within their family.

Creating a new lexicon for openness in communication enriched with terms from different sources contributes to the innovation of our approach and the generalizability and applicability of our findings. Our lexicon can be further used and expanded in future projects, providing a solid foundation for the use of NLP in the growing field of research in interpersonal communication, focusing on family communication and health care and technology-mediated communication [44]. Sentiment analysis can be further utilized in the era of precision medicine and precision public health for message tailoring and message framing. Extracting sentiment polarities can be highly informative in improving consumer experiences when using digital health platforms in promoting precision public health campaigns. For example, trust in the health care system has been associated with use of cancer surveillance, whereas conflicting messages from providers create a sense of disorientation and mistrust [45-48].

Our findings also indicated a greater likelihood of open intrafamilial communication in those who were single, had a nonacademic education, and higher informational support within their family network. These findings should be interpreted with caution and should be replicated with analyses of narratives from larger, and possibly more diverse, samples.

Studies in different domains have also considered sentiment for analyzing textual communication in social media such as Twitter or Facebook [5,9,11]. However, one significant strength of our approach was that narrative data were combined with the demographic and clinical characteristics of participants, which can increase the applicability of findings. Another important strength was the use of several sentiment lexicons to select the most suitable for this context. Sentiment scores originating from the NRC Emotion Lexicon were the most appropriate to predict “openness of communication,” whereas the other 2 sentiment lexicons (AFINN and Bing Liu) were highly correlated, resulting in a predictive algorithm of lesser importance. Studies have shown that the selection of inappropriate lexicons may impact prediction performance [39,40]. Finally, NLP can automate parts of text analysis and can be used as an assisting tool to help researchers navigate through large volumes of text data.

One limitation of our study was the small sample size and the size of the available corpus, which did not allow us to include possible significant covariates and to fully explore the potential of the NLP methodology, including sentence structure and length of words. Despite this limitation, the results of our study can be used as indicators of various narrative features, such as overall sentiment and fear, which can be important predictors of interpersonal communication and self-disclosure in this specific population. Important features of NLP analysis, such as sentence structure and length of words, can be investigated with a larger number of narratives and larger number of corpora. The analytical approach we describe in this paper can be further improved by using larger samples. Further development of a robust model will advance a more precise assessment and reach higher accuracy.

We demonstrate how various features from narratives can be used to predict “openness of communication” in individuals carrying a pathogenic variant connected to HBOC. Although our methodology requires further exploration and our findings require replication with larger samples, this is an important first step to understand how individuals and the public may react in discourses involving communication of genetic cancer risk. Overall, this experimental analysis provides evidence that our approach is promising and can be further used in the field of technology-mediated communication and precision public health.