TL;DR
This paper introduces a method to track how biomedical knowledge evolves over time by using temporal concept embeddings derived from scientific literature. By analyzing changes in the semantic relationships between medical concepts, the authors demonstrate a way to automatically identify shifts in understanding and treatment of diseases.
What Problem Does This Paper Address?
Scientific literature is a cornerstone of medical progress, but it's a constantly moving target. New discoveries are made, and existing knowledge can become outdated or even disproven. Researchers and practitioners face a significant challenge in staying abreast of these rapid developments. While NLP methods can extract information like entities and relationships, they often focus on a static snapshot of knowledge. This paper addresses the need to understand the evolution of biomedical knowledge, specifically how the meaning and relationships of medical concepts change over time, which is crucial for interpreting the reliability and context of scientific information.
Methods
The core of this work lies in generating temporal concept embeddings. Instead of traditional word embeddings, the authors focus on biomedical concepts (e.g., "headache" or "pain in the head" are mapped to a single concept ID). These concept IDs are then treated as tokens.
- Concept Normalization: Medical abstracts are processed to identify and map mentions to concept IDs from the Unified Medical Language System (UMLS) using MetaMap.
- Temporal Splitting: The MEDLINE corpus is divided into time-based subsets (primarily 5-year intervals) to capture knowledge at different points in time.
- Incremental Embedding Training: FastText (in Skip-gram negative sampling mode) is used to train embeddings for each time period. Crucially, embeddings from a previous period are used to initialize the embeddings for the next, mimicking how human knowledge builds incrementally.
- Semantic Change Measurement: Cosine similarity is used to measure the semantic shift between concept pairs across different time periods. A decrease in similarity might indicate a shift in treatment or understanding, while an increase could signify a strengthening of a known relationship.
Datasets
- MEDLINE Repository: A large bibliographic database of life sciences literature, used for training the temporal embeddings. The authors utilize titles and abstracts from MEDLINE articles.
- UMLS Metathesaurus: A comprehensive biomedical knowledge base used for concept normalization and for selecting known drug-disease relationships for evaluation.
Key Insights and Ideas
- Concepts over Words: Focusing on concepts rather than individual words is more robust for capturing biomedical meaning, as a single concept can be expressed in multiple ways.
- Incremental Learning Mimics Human Knowledge: The sequential training of embeddings, where each period's embeddings are initialized by the previous one, is a powerful analogy for how human knowledge is acquired and refined over time.
- Diachronic Shifts in Treatment: The method can reveal significant changes in how diseases are treated, showing a decline in the association of older treatments with a disease and a rise in newer ones.
- Temporal Embeddings vs. Co-occurrence: Temporal concept embeddings are shown to be more robust than traditional co-occurrence metrics (like PPMI) in capturing semantic shifts, especially when concepts don't frequently co-occur in a specific time window. This is attributed to the embeddings' ability to leverage historical context.
Key Results
The paper presents several case studies demonstrating the effectiveness of their approach:
- Minoxidil: The embeddings show a clear shift in Minoxidil's association from "high blood pressure" to "hair loss" around 1988, aligning with FDA approval for the latter.
- Microprolactinoma: The study illustrates a transition in treatment methods, moving from surgical interventions like Transethmoidal Hypophysectomy to Dopamine Agonists (Cabergoline, Bromocriptine) after the late 1970s.
- White Blood Cell Cancer: The embeddings capture the shift from older chemotherapy agents (Mitobronitol, Myelobromol) to targeted therapies like Tyrosine Kinase Inhibitors (Imatinib) around 2001.
- Hepatitis-C: The analysis highlights the transition from older antiviral treatments (Polyethylene Interferon Alpha) to newer Direct Antiviral Agents (DAA) like Telaprevir, with a significant increase in similarity observed from 2011.
- Drug-Disease Pairs: A quantitative analysis of known drug-disease pairs from UMLS shows that temporal concept embeddings generally provide higher similarity scores for known relationships compared to randomly generated negative pairs, and these scores evolve over time, reflecting changes in medical practice.
Limitations and Future Work
- Negative Polarity: The current approach assumes higher similarity equates to stronger use, which isn't always true. Future work could incorporate negative polarity detection.
- Specific Concept Effects: Sometimes, only specific sub-concepts (e.g., derivatives of dopamine agonists) show the expected effect, rather than the broader concept. This suggests a need to explore graph embeddings alongside text-based temporal ones.
- Ambiguity in Concept Mapping: Identifying the correct UMLS concept for a given mention can be challenging, requiring careful manual selection and potentially leading to more detailed analysis.
- General Literature vs. Domain Expertise: While the method is promising, interpreting the results often requires domain knowledge, and the current examples were manually selected. Future work could aim for more automated discovery.
Summary updated