Predicting Emerging Themes in Rapidly Expanding COVID-19 Literature With Unsupervised Word Embeddings and Machine...

Area of Science:

• Computational biology
• Medical informatics
• Network science

Background:

• Synthesizing vast scientific literature, like COVID-19 research, is challenging.
• Computational pipelines can improve efficiency by analyzing network features and temporal trends.
• Evidence-based responses to global health threats require effective information assimilation.

Purpose of the Study:

• To demonstrate tracking new knowledge via temporal changes in unsupervised word embeddings.
• To predict emerging research themes using machine learning on evolving word associations.
• To enhance the understanding of rapidly evolving scientific corpuses.

Main Methods:

• Extracted medical entities from >150,000 COVID-19 abstracts monthly.
• Constructed entity networks using word embeddings and cosine similarity.
• Applied machine learning for link prediction and community detection to track themes.

Main Results:

• Detected thromboembolic complications as an early theme (Aug 2020).
• Observed shifts towards long COVID (Mar 2021) and neurological symptoms (Jun 2021).
• Link prediction models achieved 0.87 AUC; identified key research themes.

Conclusions:

• Machine learning prediction of emerging research links can guide scientific inquiry.
• Tracking semantic relationships over time reveals dominant themes in biomedical literature.
• This approach aids in proactively identifying and responding to critical health issues.