Predicting Emerging Themes in Rapidly Expanding COVID-19 Literature with Dynamic Word Embedding Networks and Machine Learning

Pal, R.; Chopra, H.; Awasthi, R.; Bandhey, H.; Nagori, A.; Gulati, A.; Kumaraguru, P.; Sethi, T.

Background. COVID-19 knowledge has been changing rapidly with the fast pace of information that accompanied the pandemic. Since peer-reviewed research is a trusted source of evidence, capturing and predicting the emerging themes in COVID-19 literature are crucial for guiding research and policy. Machine learning, natural language processing and dynamical networks have the potential to enable rapid distillation and prediction of actionable insights for ending the pandemic. Objective. We hypothesized that emerging COVID-19 research trends can be captured and predicted from networks constructed upon language features. Further, we aimed to detect communities in these networks and used centrality measures to track and predict emerging network modules as dominant themes in a given time period. The goal of our study was to make our findings publicly available as an explainable AI dashboard for researchers and policymakers. Methods. Abstracts from more than 95,000 peer-reviewed articles from the WHO curated COVID-19 database were used to construct word embedding models. Named entity recognition was used to refine the terms. Cosine similarity between the terms was then used to construct dynamical networks in order to understand the temporal trend of emerging associations over months and visualized as alluvial diagrams. Finally, temporal link prediction between diseases for the subsequent month based on their trends of occurrence in the previous six months was carried out to predict the emergence and disappearance of associations in the rapidly changing pandemic scenario. Results. Community detection upon dynamical networks clearly demonstrated the emergence of thromboembolic complications as a cluster and dominant theme between March and August 2020. Forecasting of top-K influential entities further allowed prediction of future trends, such as the emergence of psychiatry theme as a central node by February 2021. XGBoost modeling in our proposed temporal link prediction framework achieved an AUC-ROC score of 0.855 for predicting new dis(associations) one month in advance. Visualization of the underlying word-embedding models allowed interactive querying to choose novel keywords and extractive models summarized the research relevant to the keyword, allowing faster knowledge distillation. Conclusion: We provide an explainable AI approach for querying, tracking and predicting novel insights in COVID-19 peer reviewed literature. The EvidenceFlow web-application is publicly available and emerging trends are updated on a monthly basis. Such approaches will be crucial to understand and pre-empt actionable research such as vaccine strategies in the ongoing pandemic.