Representation of EHR data for predictive modeling: a comparison between UMLS and other terminologies

Area of Science:

• Health Informatics
• Computational Biology
• Machine Learning

Background:

• Electronic health records (EHRs) are increasingly used for predictive disease modeling.
• Mapping raw clinical data to standard terminologies aids data aggregation and reuse.
• The impact of different terminologies on predictive model performance, especially with machine learning, is not well understood.

Purpose of the Study:

• To systematically investigate how various medical terminologies affect predictive model performance.
• To compare the effectiveness of different EHR data representations for disease risk prediction.
• To evaluate the utility of Unified Medical Language System (UMLS) and other terminologies in machine learning models.

Main Methods:

• Projected EHR diagnoses data onto Unified Medical Language System (UMLS), CCS, CCSR, ICD-9, ICD-10, and PheWAS.
• Evaluated prediction performance on two tasks: heart failure risk in diabetes patients and pancreatic cancer risk.
• Utilized logistic regression and recurrent neural network models for performance assessment.

Main Results:

• Unified Medical Language System (UMLS) provided optimal Area Under the Receiver Operating Characteristics (AUROC) for logistic regression in both prediction tasks.
• For recurrent neural networks, UMLS performed best for pancreatic cancer prediction and second best for heart failure prediction.
• Terminologies with larger vocabularies and finer granularity generally correlated with improved prediction accuracy.

Conclusions:

• The choice of terminology significantly impacts predictive model performance.
• Unified Medical Language System (UMLS) is a highly effective terminology for EHR-based predictive modeling.
• Finer-grained terminologies are associated with enhanced predictive accuracy, informing future model design.