Background/Purpose:

Multiple studies revealed adverse side effects caused by different rheumatoid arthritis (RA) medications. However, little is known about gender differences in RA treatment and medications. Increasingly patients are sharing and searching for RA information on health-related social media. This study developed deep learning (DL) text analytical tools on top of health-related social media to identify gender-specific disparities in side effects reported by people taking 2 common medications for RA: Methotrexate/Trexall and Leflunomide/Arava.

Methods: Data from three well-known health-related social media, WebMD (https://www.webmd.com/drugs), DailyStrength (https://www.dailystrength.org) and and Drugs.com (https://www.drugs.com), were extracted from June 2013 to May 2019, including gender information and the text of posts. Since Drugs.com does not provide gender of the users, a top 3 name-to-gender inference service, namely NameAPI (https://www.nameapi.org) has been used to identify gender of the users based on their names. We segmented each post into a set of sentences as the units for DL, which used word2vec feature representation. An ensemble of deep neural networks was developed to identify sentences relating to drug adverse side effects. The training data consisted of a random subset of sentences annotated by three domain experts, where each sentence was classified as either describing a side effect (SE) or no side effect (No-SE). A total of 12,702 sentences (6,453 SE and 6,249 No-SE) were annotated, with high inter-annotator agreement of kappa=0.81. A baseline classifier for comparison was trained using Naïve Bayes with bag-of-word features.

Results:

On 4-fold cross-validation, the DL classifier achieved an AUC of 0.897. The accuracy of the ensemble DL classifier outperformed the Naïve Bayes classifier significantly (p=0.04). With the best-tuned configuration, the DL classifier was then applied on 39,191 unseen new sentences for large-scale trend analysis, in which the overall disparity and disparity in five common SEs are shown in Figure 1. To focus on only RA-related sentences, we constrained messages to those that explicitly included at least one of: “RA”, “R.A”, “R.A.”, “bone”, and “joint”. Figure 1 shows the visualization results. 

Conclusion: The study demonstrated feasibility of developing highly accurate DL classifiers for identifying RA medication side effects in social media, providing gender disparity information using this wealth of data. Although social media may reflect biased prevalence of conditions due to subjective user behaviors, we were able to faithfully describe the data. In addition to improving our methods, future work will emphasize verifying/comparing such disparity by applying similar analysis on EHRs along with generic social media (e.g.,  Twitter, Google+, and reddit). The present work elucidates construction of a holistic patient-centered decision support framework that can be merged with a diverse range of clinical data, such as EHRs and clinical notes to help clinicians and patients assess benefits and risks of RA treatments, conducting the development of personalized medicine within the RA context.

« Back to 2019 ACR/ARP Annual Meeting

ACR Meeting Abstracts - https://acrabstracts.org/abstract/deep-learning-social-media-analysis-demonstrated-gender-specific-disparity-in-side-effects-from-rheumatoid-medications/