Original Research
Monday, November 13, 2023
01:00 PM–02:30 PM
Abstract
Introduction: As the utilization of DPP4 inhibitors and SGLT2 inhibitors increases in diabetes treatment, there is an increased focus on their potential adverse events. Twitter, being a platform facilitating the exchange of opinions and information on diverse subjects, offers a distinct advantage in collecting adverse event data related to drugs. It enables us to capture the wide range of perspectives and experiences of drug users, including the prompt identification of previously unreported adverse events.
Research Question or Hypothesis: This study aimed to leverage deep learning techniques and Twitter data for the analysis of side effects associated with diabetes medications.
Study Design: Cross-sectional study
Methods: We gathered Twitter posts pertaining to DPP4 inhibitors and SGLT2 inhibitors and employed a BERT-based binary classification model to filter out spam and promotional tweets. Subsequently, we developed two classifiers based on RoBERTa and Distilbert models, employing supervised learning to classify tweets containing adverse drug events (ADEs). To extract relevant terms associated with ADEs, we utilized a text summarization extractor utilizing the Seq2Seq model and combined them with Meddra PT terms utilizing a Word2Vec embedding model combined with the Attention algorithm.
Results: Using RoBERTa and DistilBERT ADE classifiers, we attained impressive accuracies of 0.9689 and 0.9429, respectively. Our NER model demonstrated accuracies of 29.36% (RoBERTa) and 32.29% (DistilBERT) for extracting PT terms of adverse reactions. To enhance results, we employed an Attention-based Seq2Seq summarization approach, which yielded 84% of summaries with non-zero Rouge scores. For connecting PTs, we employed a supervised learning model with FastText, BERT, CBOW, RN, CNN, and Attention embeddings, achieving accuracies of 11.8%, 32.2%, 47.1%, 49%, 17%, and 54.7% respectively. Notably, the count-based embedding method outperformed the others.
Conclusion: Deep learning enabled real-time collection of adverse tweets, offering diverse insights from drug users. Sentence summarization technology holds promise for proactive integration into future clinical treatment planning.
Presenting Author
Yun-Kyoung Song Ph.DCollege of Pharmacy, Seoul National University
Authors
Se Hun Oh BS
Daegu Catholic University
Dong Young Park BS
Daegu Catholic University