A natural language processing pipeline to advance the use of Twitter data for digital epidemiology of adverse pregnancy outcomes

Background: In the United States, 17% of pregnancies end in fetal loss: miscarriage or stillbirth. Preterm birth affects 10% of live births in the United States and is the leading cause of neonatal death globally. Preterm births with low birthweight are the second leading cause of infant mortality in the United States. Despite their prevalence, the causes of miscarriage, stillbirth, and preterm birth are largely unknown. Objective: The primary objectives of this study are to (1) assess whether women report miscarriage, stillbirth, and preterm birth, among others, on Twitter, and (2) develop natural language processing (NLP) methods to automatically identify users from which to select cases for large-scale observational studies. Methods: We handcrafted regular expressions to retrieve tweets that mention an adverse pregnancy outcome, from a database containing more than 400 million publicly available tweets posted by more than 100,000 users who have announced their pregnancy on Twitter. Two annotators independently annotated 8109 (one random tweet per user) of the 22,912 retrieved tweets, distinguishing those reporting that the user has personally experienced the outcome (“outcome” tweets) from those that merely mention the outcome (“non-outcome” tweets). Inter-annotator agreement was κ = 0.90 (Cohen's kappa). We used the annotated tweets to train and evaluate feature-engineered and deep learning-based classifiers. We further annotated 7512 (of the 8109) tweets to develop a generalizable, rule-based module designed to filter out reported speech—that is, posts containing what was said by others—prior to automatic classification. We performed an extrinsic evaluation assessing whether the reported speech filter could improve the detection of women reporting adverse pregnancy outcomes on Twitter. Results: The tweets annotated as “outcome” include 1632 women reporting miscarriage, 119 stillbirth, 749 preterm birth or premature labor, 217 low birthweight, 558 NICU admission, and 458 fetal/infant loss in general. A deep neural network, BERT-based classifier achieved the highest overall F1-score (0.88) for automatically detecting “outcome” tweets (precision = 0.87, recall = 0.89), with an F1-score of at least 0.82 and a precision of at least 0.84 for each of the adverse pregnancy outcomes. Our reported speech filter significantly (P < 0.05) improved the accuracy of Logistic Regression (from 78.0% to 80.8%) and majority voting-based ensemble (from 81.1% to 82.9%) classifiers. Although the filter did not improve the F1-score of the BERT-based classifier, it did improve precision—a trade-off of recall that may be acceptable for automated case selection of more prevalent outcomes. Without the filter, reported speech is one of the main sources of errors for the BERT-based classifier. Conclusion: This study demonstrates that (1) women do report their adverse pregnancy outcomes on Twitter, (2) our NLP pipeline can automatically identify users from which to select cases for large-scale observational studies, and (3) our reported speech filter would reduce the cost of annotating health-related social media data and can significantly improve the overall performance of feature-based classifiers. © 2020 The Authors