Copyright: Large Language Models for Electronic Health Records (Case No. 2024-216)

Intro Sentence:

UCLA researchers from the Department of Computational Medicine have developed a novel model for tabulating electronic health records.

Background:

Electronic Health Records (EHR) provide healthcare systems with insights into health histories. Machine learning models have been developed to use EHR for inference tasks based on specific diagnosis and prognosis. Existing models require pre-training using large and diverse datasets to achieve complex tasks. The widespread adoption of these models across different domains is limited by the tabular form of EHR data and the inaccessibility of textual clinical data due to privacy concerns. Traditional machine learning approaches are hampered by the heterogeneous, or inconsistent nature of EHR data, which includes numerical and categorical information. These techniques suffer from inefficient data usage and sparse data representation. There remains an unmet need for a machine learning model that efficiently deciphers heterogeneous electronic health records to improve the storage, retrieval, and analysis of patient data. 

Innovation:

UCLA researchers from the Department of Computational Medicine have developed a Multiple Embedding Model for Electronic Health Records (MEME) for efficient data processing. This model utilizes a multiple embedding strategy, where different components of the EHR are encoded separately, to bridge between tabular EHR data and existing Natural Language Processing techniques. The model encodes different components of EHR data separately, allowing for efficient data processing while maintaining the inherent meaning of the data. This innovation can revolutionize modern healthcare as it leverages patient data encoded in traditional EHR and allows ML models to predict future events with high precision. 

Potential Applications: 

•    Emergency department decision support 
•    Hospital resource management 
•    Clinical research 
•    Personalized treatment plans 
•    Patient prognostics and diagnostics

Advantages: 

•    Integration with modern ML models
•    Improved accuracy
•    Increased data processing efficiency 
•    Enhanced adaptability 

Development-To-Date:

The first demonstration of the invention is complete 

Reference:

UCLA Case No. 2024-216

Lead Inventor:

Jeffrey Chiang 
 

Patent Information:
For More Information:
Joel Kehle
Business Development Officer
joel.kehle@tdg.ucla.edu
Inventors:
Jeffrey Chiang
Simon Lee