Adjusting Twitter Data as a Source for Blood Donation Analysis: BDT-UC Dataset and BERT Implementations

Collecting data related to Blood Donation (BD) from Twitter often faces challenges related to noise. This study addresses this issue through automated classification, implementing Bidirectional Encoder Representation from Transformers (BERT) models. We begin by creating an unsupervised Twitter dataset focused on blood donation. Subsequently, we introduce the Blood Donor Tweet-User Classification Dataset (BDT-UC), a multi-class labeled dataset that considers user roles in blood donation. BDT-UC was used to train and validate BERT models with two approaches: fine-tuning the models on BDT-UC and customizing the models with additional layers. The results demonstrate consistent user classification in our proposed dataset. This dataset paves the way for more efficient BD-related data analysis by categorizing data and reducing noise from Twitter data.Medical relevance: This work showcases the possibility of adapting Twitter as a reliable source for blood donation-related data, with a categorization implementation of 78% in accuracy.