TL;DR
Researchers have created a 2D map of over 21 million biomedical research abstracts from PubMed, revealing the intricate structure and evolution of scientific fields. This map offers a novel way to explore research trends, identify emerging issues like gender bias, and even detect potentially fraudulent publications.
What Problem Does This Paper Address?
The sheer volume of scientific publications in biomedicine and life sciences makes it incredibly challenging to grasp the overall landscape of research, track its evolution over time, or understand how different fields relate to each other. Existing search engines are good for finding specific papers but don't offer a global, visual overview. This lack of a "bird's eye view" hinders the ability to identify emerging trends, understand disciplinary shifts, and uncover systemic issues within the scientific literature.
Methods
The core of this work is the creation of a 2D "map" of biomedical research. This was achieved by:
- Embedding Abstracts: Using a transformer-based large language model called PubMedBERT, each of the 21 million English abstracts was converted into a 768-dimensional numerical representation. This model is specifically trained on biomedical text, making it well-suited for this domain.
- Dimensionality Reduction: The high-dimensional embeddings were then reduced to two dimensions using t-SNE (t-Distributed Stochastic Neighbor Embedding). The authors employed a modified version of t-SNE tailored to handle very large datasets, ensuring efficient processing and better convergence.
- Labeling and Visualization: The 2D map was then annotated with labels derived from journal titles (e.g., "virology," "biochemistry," "radiology"). This allowed for the visualization of different scientific fields and their relationships. An interactive website was also developed for easy exploration of this map.
Datasets
- PubMed Database: The primary dataset used was a snapshot of the PubMed database containing approximately 21 million English-language biomedical research articles, primarily from 1970-2021.
- Retraction Watch Database: Used to identify and validate retracted papers.
Key Insights and Ideas
- Visualizing the Unseen: The 2D map provides an unprecedented visual overview of the entire biomedical literature, allowing for the identification of clusters of related research and the relationships between different disciplines.
- Emergence of COVID-19 as a Distinct Field: The COVID-19 literature formed a remarkably isolated cluster on the map, suggesting an unprecedented impact on the scientific landscape, creating a distinct field of study in a very short period.
- Shifting Dynamics in Neuroscience: The map revealed a clear separation within neuroscience between molecular/cellular and cognitive/behavioral research, and showed how the relative prominence of these subfields has shifted over time.
- Uneven Adoption of Machine Learning: The uptake of machine learning in biomedical research is not uniform, with some fields like radiology and psychiatry adopting it earlier and more rapidly than others.
- Granular Gender Bias Analysis: The map allows for a more nuanced understanding of gender bias, revealing heterogeneity within broad disciplines and highlighting specific sub-regions where gender ratios differ significantly.
- Identifying Suspicious Literature: Clusters of retracted papers, particularly those associated with "paper mills," were identifiable on the map, suggesting its potential as a tool for editorial investigation.
Key Results
- Map Creation: A 2D map of 21 million biomedical abstracts was successfully generated.
- COVID-19 Isolation: COVID-19 literature showed the highest "isolatedness" metric (80.6% in BERT representation), indicating its distinctness.
- Neuroscience Trends: The map visually demonstrated a shift in neuroscience research focus over time, with older papers concentrated in molecular/cellular areas and newer ones in cognitive/behavioral areas.
- Machine Learning Uptake: Radiology and psychiatry showed early ML adoption, while dermatology and gynecology showed very little until recently.
- Gender Distribution: The map revealed that female authors are more concentrated in fields like nursing, education, and psychology, and showed detailed within-discipline variations in gender ratios.
- Retracted Papers: Retracted papers were found to be concentrated in specific areas, particularly those related to cancer drugs, marker genes, and microRNA, suggesting potential paper mill activity.
Limitations and Future Work
- t-SNE Limitations: The authors acknowledge that t-SNE, while effective for visualization, may not preserve the exact relationships of the closest neighbors in the high-dimensional space. For finding the most similar papers, other tools might be better.
- Labeling Method: The journal-title-based labeling method is an approximation and overlooks interdisciplinary journals.
- Gender Inference: Gender inference from first names is approximate and has limitations, particularly with ambiguous names and biases in training data.
- Future Work:
- Combining textual similarity with citation information for a more comprehensive map.
- Exploring larger generative language models for potentially better representations.
- Continuously updating the interactive map with new PubMed data.
- Further investigating the identified clusters of suspicious literature.
Summary updated