The classification of Arabic tweets around COVID-19 is a crucial task to understand public discourse, information exchange, and the dissemination of misinformation detection, ultimately supporting effective policy-making and public health strategies. However, dealing with multi-label Arabic tweets classification is a challenging task, due to many challenges, such as tweets containing a lot of noisy data, and the frequent overlap of tweets across multiple biomedical classes. Moreover, machine learning models fail to capture the word context within the tweet. In this work, we build vector representations able to capture contextual information within Arabic social clinical text. These representations address the specific challenges of Arabic COVID-19 tweets, including the specific jargon of clinical domain, and the derivational structure of the Arabic language. Our approach consists of fine-tuning various pretrained transformer models, including AraBERT, ArBERT, AraELECTRA, MarBERT, CamelBERT, Roberta, and Qarib, using COVID-19-MLM dataset. Experiments show that ArBERT model achieved the best performance, by achieving an F1-micro of 81.4% and a Jaccard score of 78.6%. Highlighting the effectiveness of transformer-based representations in this context.