Published on in Vol 26 (2024)

Preprints (earlier versions) of this paper are available at, first published .
Using Longitudinal Twitter Data for Digital Epidemiology of Childhood Health Outcomes: An Annotated Data Set and Deep Neural Network Classifiers

Using Longitudinal Twitter Data for Digital Epidemiology of Childhood Health Outcomes: An Annotated Data Set and Deep Neural Network Classifiers

Using Longitudinal Twitter Data for Digital Epidemiology of Childhood Health Outcomes: An Annotated Data Set and Deep Neural Network Classifiers

Research Letter

1Department of Biostatistics, Epidemiology, and Informatics, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA, United States

2Department of Health Sciences, University of Monterrey, San Pedro Garza García, Nuevo León, Mexico

3Department of Obstetrics and Gynecology, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA, United States

4Department of Computational Biomedicine, Cedars-Sinai Medical Center, West Hollywood, CA, United States

Corresponding Author:

Graciela Gonzalez-Hernandez, PhD

Department of Computational Biomedicine

Cedars-Sinai Medical Center

Pacific Design Center, Ste G549F

700 N San Vicente Blvd

West Hollywood, CA, 90069

United States

Phone: 1 310 423 3521


We manually annotated 9734 tweets that were posted by users who reported their pregnancy on Twitter, and used them to train, evaluate, and deploy deep neural network classifiers (F1-score=0.93) to detect tweets that report having a child with attention-deficit/hyperactivity disorder (678 users), autism spectrum disorders (1744 users), delayed speech (902 users), or asthma (1255 users), demonstrating the potential of Twitter as a complementary resource for assessing associations between pregnancy exposures and childhood health outcomes on a large scale.

J Med Internet Res 2024;26:e50652



Many children are diagnosed with disorders that can impact their daily lives and can last throughout their lifetime. For example, in the United States, 17% of children are diagnosed with a developmental disability [1] and 8% of them with asthma [2]. Meanwhile, data sources for assessing the association of these outcomes with pregnancy exposures are limited, as pregnancy registries typically follow infants for up to 1 year after birth. While our previous work [3,4] demonstrated the utility of Twitter as a source of data regarding pregnancy outcomes, the ability to continue collecting users’ tweets on an ongoing basis after birth may present opportunities to detect outcomes in childhood. Twitter data have been used to identify self-reports of attention-deficit/hyperactivity disorder (ADHD) [5], autism spectrum disorders (ASD) [6], and asthma [7], but not to identify reports of these disorders in users’ children. This study aimed to assess whether there are users who report having a child with ADHD, ASD, delayed speech, or asthma, and develop and evaluate an automated method for identifying these reports.

Ethical Considerations

The study data were collected and analyzed in accordance with the Twitter Terms of Service. The institutional review boards of the University of Pennsylvania and Cedars-Sinai Medical Center deemed this study exempt.

Data Collection

We searched for mentions of ADHD, ASD, delayed speech, and asthma among all the tweets posted by more than 100,000 users who reported their pregnancy on Twitter [8]. We then searched these matching tweets for references to a child and the user, and excluded tweets that matched specific patterns indicating the user’s own disorder. The query (Multimedia Appendix 1) returned 36,094 tweets (excluding retweets) posted by 11,712 users.


We used 400 matching tweets—100 per outcome—to develop annotation guidelines (Multimedia Appendix 2) for distinguishing those that report having a child with a disorder from those that do not. An additional 9334 tweets—1 random tweet per user—were then independently annotated: 8334 by 2 annotators and 1000 by all 3. Interannotator agreement (Fleiss kappa) was 0.88. After resolving disagreements among all 9734 tweets, we determined that 3019 (31%) reported having a child with a disorder and 6715 (69%) did not.

Automatic Classification

We split the 9734 tweets into 80% (n=7787) training (Multimedia Appendix 3) and 20% (n=1947) test data, and performed machine learning experiments using deep neural network classifiers based on bidirectional encoder representations from transformers (BERT) [9]: the BERT-Base-Uncased, RoBERTa-Large, and BERTweet-Large pretrained models in the Huggingface library. Our preprocessing included normalizing URLs and usernames, and lowercasing the tweets. For training, we used Adam optimization, 5 epochs, a batch size of 8, and a learning rate of 0.00001, based on evaluating after each epoch using a 5% split of the training set. We fine-tuned all layers of the models with our annotated tweets.

Table 1 presents the performance of the classifiers. The RoBERTa-Large [10] classifier achieved the highest overall F1-score (0.93). Table 1 also presents the performance of the RoBERTa-Large classifier for tweets that mention specific outcomes. We deployed the RoBERTa-Large classifier on the additional 26,360 unlabeled tweets that matched our query (Multimedia Appendix 1). Between the 9734 manually annotated tweets and the 26,360 automatically classified tweets, we identified 3806 total users who reported having a child with ADHD (n=678), ASD (n=1744), delayed speech (n=902), or asthma (n=1255).

Table 2 presents examples of tweets in the test set that were misclassified by the RoBERTa-Large classifier. While 28 (58%) of the 48 false positives do refer to the user’s child, 11 (39%) indicate that someone other than the user’s child has a disorder (tweet 1), and 9 (32%) indicate that a disorder is merely suspected or exhibited (tweet 2). Among the other 20 (42%) of the 48 false positives, 10 (50%) are reported speech, such as quotations (tweet 3). Among the 42 false negatives, 22 (52%) do not explicitly mention the user’s child (tweet 4)—for example, using a pronoun or name—and 14 (33%) do not explicitly indicate that the child has a disorder (tweet 5).

Table 1. Precision, recall, and F1-score of classifiers for the class of tweets that report having a child with attention-deficit/hyperactivity disorder (ADHD), autism spectrum disorder (ASD), delayed speech, or asthma, including the outcome-specific precision, recall, and F1-score for the RoBERTa-Large classifier.



Delayed speech0.940.960.95

Table 2. Sample false positives and false negatives of a RoBERTa-Large classifier for detecting tweets that report having a child with attention-deficit/hyperactivity disorder (ADHD), autism spectrum disorder (ASD), delayed speech, or asthma (with the text that matched the data collection query in italics).
Tweet numberTweetActualPredicted
1So Maxine Waters can be maskless on a plane but I can’t fly with my 2 year old cause she won’t wear a mask? Kids with autism are being banned from flying because they won’t wear a mask?+
2they treat my baby with asthma meds all the time but didn’t diagnose her with it im pretty sure she has it tho+
3Any tips for this mum: “My daughter is 10. My parents would like to gift her either a phone or a smart watch which is easy to use and won’t be easily damaged by a very active ADHD kid... I need help choosi… [URL]+
4Flying tomorrow...during a pandemic with a nonverbal 3 year old. We could use some prayers, please.+
5I wouldn’t change my child for anything in the world. I’m just curious to know where autism came from because me and his dad don’t have any family members that are autistic. It’s just a question out of curiosity+

Our ability to identify Twitter data during pregnancy for users who reported having a child with ADHD, ASD, delayed speech, or asthma suggests that Twitter could be a complementary resource for assessing associations between pregnancy exposures and childhood health outcomes, with potential clinical implications for informing prenatal care. The overall and outcome-specific performance for automatically identifying these outcomes demonstrates the feasibility of using Twitter data for observational studies on a large scale.


This work was supported by the National Library of Medicine (R01LM011176). The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. The authors thank Ivan Flores for contributing to software applications and Karen O’Connor for contributing to annotating the Twitter data. Generative artificial intelligence was not used for any portion of the study or manuscript writing.

Data Availability

The manually annotated training data are included with this article in Multimedia Appendix 3. In accordance with the Twitter Terms of Service, these tweets are made available as tweet IDs, which can be rehydrated as tweet objects if they remain public at the time they are requested through the Twitter API.

Authors' Contributions

AZK and JAGG contributed to the data collection, annotation, machine learning experiments, error analysis, and drafting of the manuscript. LDL provided guidance on pregnancy outcomes and edited the manuscript. GGH designed and guided the study and edited the manuscript.

Conflicts of Interest

None declared.

Multimedia Appendix 1

Data collection query.

TXT File , 11 KB

Multimedia Appendix 2

Annotation guidelines.

DOCX File , 28 KB

Multimedia Appendix 3

Training data.

TXT File , 165 KB

  1. Zablotsky B, Black LI, Maenner MJ, Schieve LA, Danielson ML, Bitsko RH, et al. Prevalence and trends of developmental disabilities among children in the United States: 2009-2017. Pediatrics. Oct 2019;144(4):e20190811. [FREE Full text] [CrossRef] [Medline]
  2. Zahran HS, Bailey CM, Damon SA, Garbe PL, Breysse PN. Vital signs: asthma in children - United States, 2001-2016. MMWR Morb Mortal Wkly Rep. Feb 09, 2018;67(5):149-155. [CrossRef] [Medline]
  3. Golder S, Chiuve S, Weissenbacher D, Klein A, O'Connor K, Bland M, et al. Pharmacoepidemiologic evaluation of birth defects from health-related postings in social media during pregnancy. Drug Saf. Mar 3, 2019;42(3):389-400. [CrossRef] [Medline]
  4. Klein AZ, O'Connor K, Levine LD, Gonzalez-Hernandez G. Using Twitter data for cohort studies of drug safety in pregnancy: proof-of-concept with β-blockers. JMIR Form Res. Jun 30, 2022;6(6):e36771. [FREE Full text] [CrossRef] [Medline]
  5. Guntuku SC, Ramsay JR, Merchant RM, Ungar LH. Language of ADHD in adults on social media. J Atten Disord. Oct 08, 2019;23(12):1475-1485. [CrossRef] [Medline]
  6. Hswen Y, Gopaluni A, Brownstein JS, Hawkins JB. Using Twitter to detect psychological characteristics of self-identified persons with autism spectrum disorder: a feasibility study. JMIR Mhealth Uhealth. Feb 12, 2019;7(2):e12264. [FREE Full text] [CrossRef] [Medline]
  7. Edo-Osagie O, Smith G, Lake I, Edeghere O, De La Iglesia B. Twitter mining using semi-supervised classification for relevance filtering in syndromic surveillance. PLoS One. 2019;14(7):e0210689. [FREE Full text] [CrossRef] [Medline]
  8. Klein AZ, Kunatharaju S, O'Connor K, Gonzalez-Hernandez G. Pregex: rule-based detection and extraction of Twitter data in pregnancy. J Med Internet Res. Feb 09, 2023;25:e40569. [FREE Full text] [CrossRef] [Medline]
  9. Devlin J, Cheng M, Lee K, Toutanova K. BERT: pre-training of deep bidirectional transformers for language understanding. 2019. Presented at: 17th Annual Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies; June 2-7, 2019;4171-4186; Minneapolis, MN.
  10. Liu Y, Ott M, Goyal N, Du J, Joshi M, Chen D, et al. RoBERTa: a robustly optimized BERT pretraining approach. arXiv. URL: [accessed 2023-07-07]

ADHD: attention-deficit/hyperactivity disorder
ASD: autism spectrum disorder
BERT: bidirectional encoder representations from transformers

Edited by T de Azevedo Cardoso; submitted 07.07.23; peer-reviewed by C Ni, E Guo; comments to author 22.08.23; revised version received 05.09.23; accepted 19.09.23; published 25.03.24.


©Ari Z Klein, José Agustín Gutiérrez Gómez, Lisa D Levine, Graciela Gonzalez-Hernandez. Originally published in the Journal of Medical Internet Research (, 25.03.2024.

This is an open-access article distributed under the terms of the Creative Commons Attribution License (, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in the Journal of Medical Internet Research, is properly cited. The complete bibliographic information, a link to the original publication on, as well as this copyright and license information must be included.