Data on potential risk factors in pregnancy are limited. Meanwhile, in the United States, 17% of pregnancies end in fetal loss [1], and birth defects and preterm births are the leading causes of infant mortality [2]. In previous work [3], we developed an automated natural language processing pipeline that identifies users who announced their pregnancy on Twitter and collects all of their tweets on an ongoing basis. We have also demonstrated that their tweets can be used for observational studies [4-6]. However, selecting users for such studies involves additional processing to address a limitation of our pipeline—namely, that many of the users refer to a pregnancy either that occurred prior to the availability of their tweets or for which we could not determine the prenatal period. To streamline the use of Twitter as a source of data, the objective of this study was to advance a downstream system developed in our previous work [7] and evaluate its upstream use for identifying tweets that indicate the availability of Twitter data during pregnancy and can be used to extract dates marking the beginning and end of the 40-week prenatal period.

We deployed Pregex on the Twitter timelines of more than 550,000 users—mostly users identified by our original pipeline [3]—and detected approximately 235,000 tweets that were posted by more than 100,000 users. For validation, 3 annotators labeled a random sample of 4017 matching tweets—1 tweet per user and up to 100 tweets per regular expression—to identify whether they self-report an ongoing pregnancy, and the correct beginning and end dates were extracted (Multimedia Appendix 1). Among these 4017 tweets, 3716 (90%) were dual annotated and 400 (10%) were annotated by all 3 annotators. For 381 (95%) of these 400 tweets, the 3 annotators agreed on whether the tweet self-reports an ongoing pregnancy, agreeing that 378 (99%) of these 381 tweets do. For 376 (99%) of these 378 tweets, the 3 annotators agreed on whether the correct beginning and end dates were extracted. After resolving disagreements among all 4017 tweets, we established that Pregex had a precision of 0.96 for identifying ongoing pregnancies, where precision = true positives / (true positives + false positives). Among the 3875 true positives, Pregex had a precision of 0.99 for extracting dates marking the beginning and end of the 40-week prenatal period.

Because pregnancy is a common event, our rule-based approach can identify tweets during pregnancy with high precision and on a large scale, facilitating the use of Twitter as a complementary source of data for observational studies. In real time, Pregex is detecting approximately 50 new users daily, taking as input approximately 15,000 tweets returned from the Twitter streaming application programming interface that matches pregnancy-related keywords derived from the regular expressions (Multimedia Appendix 2). Among the 142 false-positive tweets in our evaluation that did not self-report an ongoing pregnancy, 42 (29%) mention a due date that refers to a deadline (eg, payments), which we will address in future work.