In March 2020, the World Health Organisation declared that COVID-19 was a new pandemic. This deadly virus spread and affected many countries in the world. During the outbreak, social media platforms such as Twitter contributed valuable and massive amounts of data to better assess health-related decision making. Therefore, we propose that users’ sentiments could be analysed with the application of effective supervised machine learning approaches to predict disease prevalence and provide early warnings. The collected tweets were prepared for preprocessing and categorised into: negative, positive, and neutral. In the second phase, different features were extracted from the posts by applying several widely used techniques, such as TF-IDF, Word2Vec, Glove, and FastText to capture features’ datasets. The novelty of this study is based on hybrid features extraction, where we combined syntactic features (TF-IDF) with semantic features (FastText and Glove) to represent posts accurately, which helps in improving the classification process. Experimental results show that FastText combined with TF-IDF performed better with SVM than the other models. SVM outperformed the other models by 88.72%, as well as for XGBoost, with an 85.29% accuracy score. This study shows that the hybrid methods proved their capability of extracting features from the tweets and increasing the performance of classification.

The spread of COVID-19 has affected more than 200 countries and has caused serious public health concerns. The infected cases are on the increase despite the effectiveness of the vaccines. An efficient and quick surveillance system for COVID-19 can help healthcare decision-makers to contain the virus spread. In this study, we developed a novel framework using machine learning (ML) models capable of detecting COVID-19 accurately at an early stage. To estimate the risks, many models use social networking sites (SNSs) in tracking the disease outbreak. Twitter is one of the SNSs that is widely used to create an efficient resource for disease real-time analysis and can provide an early warning for health officials. We introduced a pipeline framework of outbreak prediction that incorporates a first-step hybrid method of word embedding for tweet classification. In the second step, we considered the classified tweets with external features such as vaccine rate associated with infected cases passed to machine learning algorithms for daily predictions. Thus, we applied different machine learning models such as the SVM, RF, and LR for classification and the LSTM, Prophet, and SVR for prediction. For the hybrid word embedding techniques, we applied TF-IDF, FastText, and Glove and a combination of the three features to enhance the classification. Furthermore, to improve the forecast performance, we incorporated vaccine data as input together with tweets and confirmed cases. The models’ performance is more than 80% accurate, which shows the reliability of the proposed study.