Text Phrase Mapping to Medical Taxonomy via Embedding Spaces
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for mapping text phrases to medical language taxonomies face challenges due to the complexity of coding systems, limited scalability, and the inability to generalize across different versions and private coding mechanisms, especially when dealing with large controlled vocabularies like SNOMED CT.
Innovation Solution
A computer-implemented method and system that maps text phrases to a taxonomy by converting them into word embeddings and node embeddings using techniques like GloVe, fastText, node2vec, and neural networks, enabling a mapping function to translate between these spaces, allowing for more generalizable and efficient data interchangeability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If symbolic natural language processing techniques (e.g., cTAKES with annotated lookup tables) are used to map clinical entities to controlled vocabulary, then mapping accuracy for known terms is improved, but scalability and generalization ability deteriorate due to reliance on predefined symbolic representations that cannot adapt to vocabulary updates or private coding mechanisms
Solution Approach 1:
The patent replaces symbolic NLP mechanisms with neural network-based embedding models (word2vec, GloVe, fastText) that learn semantic representations from data. This substitution eliminates dependency on predefined symbolic lookup tables, enabling the system to handle vocabulary updates and private coding mechanisms without requiring manual reconfiguration of symbolic mappings.
Solution Approach 2:
The patent transforms discrete symbolic representations into continuous vector embeddings, changing the fundamental parameter representation from categorical to continuous space. This allows smooth interpolation and generalization between terms, enabling the system to adapt to new vocabulary items and private coding mechanisms through continuous parameter adjustment rather than discrete symbolic matching.
2Adaptability or versatility
If machine learning approaches with word embeddings are used to map medical concepts to taxonomy, then scalability is improved, but effectiveness deteriorates when applied to large controlled vocabularies like SNOMED CT due to the large size of vocabularies and small number of terms per concept
Solution Approach 1:
The patent segments the mapping task into multiple components: (1) generating word embeddings for terms, (2) generating concept embeddings for taxonomy nodes, and (3) learning a projection function between these embedding spaces. This segmentation allows the system to handle large vocabularies by processing them in manageable chunks through pre-computed embeddings while maintaining precision through the learned projection function.
Solution Approach 2:
The patent introduces an intermediary projection function that maps word embeddings to concept embeddings in the taxonomy space. This intermediary layer decouples the large vocabulary from the structured taxonomy, allowing the system to leverage pre-computed word embeddings while maintaining precise mappings to the controlled vocabulary through the learned projection function.
3Quantity of substance
If multiple coding systems and private coding mechanisms are used to encode clinical information, then comprehensive data capture is improved, but system complexity and difficulty of data analysis worsen due to the sheer number of different coding systems with no standardized mappings
Solution Approach 1:
The patent creates a universal embedding space that can represent terms from multiple different coding systems and vocabularies simultaneously. The neural network models generate embeddings that capture semantic meaning independent of the source coding system, enabling a single unified representation to handle diverse coding mechanisms without requiring separate processing pipelines for each system.
Solution Approach 2:
The patent introduces embedding vectors as an intermediary representation that mediates between different coding systems and the target taxonomy. This intermediary layer translates diverse coding system terms into a common semantic space, eliminating the need for complex direct mapping dictionaries between multiple coding systems while maintaining comprehensive data capture capability.
Data Source
AI summary
There is provided systems and methods for mapping of text phrases to a taxonomy. In an aspect, the method including: receiving the corpus and the taxonomy; mapping the text phrases in the corpus to a set of word embeddings in a word embedding space, where each sequence of word embeddings corresponds to individual words in one of the text phrases; vectorizing the taxonomy to a set of node embeddings in a node embedding vector space; mapping the set of word embeddings to the set of node embeddings using a mapping function, the mapping function outputting points in the node embedding space associated with sequences in the word embeddings; and outputting the mapping function.