Background: Healthcare professionals (HCPs) are on the frontline of fighting the COVID-19 pandemic. Recent reports have indicated that, in addition to facing an increased risk of being infected by the virus, HCPs face an increased risk of suffering from emotional difficulties associated with the pandemic. Therefore, understanding HCPs’ experiences and emotional displays during emergencies is a critical aspect of increasing the surge capacity of communities and nations. Methods: In this study, we analyzed posts published by HCPs on Twitter to infer the content of discourse and emotions of the HCPs in the United States (US) and United Kingdom (UK), before and during the COVID-19 pandemic. The tweets of 25,207 users were analyzed using natural language processing (NLP). Results: Our results indicate that HCPs in the two countries experienced common health, social, and political issues related to the pandemic, reflected in their discussion topics, sentiments, and emotional display. However, the experiences of HCPs in the two countries are also subject to local socio-political trends, as well as cultural norms regarding emotional display. Conclusions: Our results support the potential of utilizing Twitter discourse to monitor and predict public health responses in emergencies.

Twitter data analysis gives valuable insights into various aspects of society, such as consumer opinions, political sentiments, brand reputation, and more. This information can help businesses and organizations make informed decisions, track the success of marketing campaigns, and identify emerging trends. Additionally, Twitter data analysis can also aid in research fields such as social sciences and humanities by providing a large, real-time dataset of human behaviour and language. Sentiment Analysis uses natural language processing and machine learning algorithms to categorize tweets as positive, negative, or neutral based on the sentiment expressed in the text. The previous sentiment analysis literature has cited several drawbacks especially on frequency-based vectorization models like Bag of Words, TF-IDF, and traditional word embeddings that generally cannot capture semantic relationships and contextual dependencies in short and noisy Twitter data. The proposed work comprises two phases. In the first phase, text pre-processing, vectorization, word embedding and feature selection is performed using the Frequency Co-occurrence Matrix and Fisher's score algorithm. In the second phase, the Multi stacked BiLSTM is implemented to perform classification as positive, negative and neutral. The performance of the proposed work achieves an accuracy and mean squared error rate as 98% and 0.01% respectively.