Using machine learning to identify health outcomes from electronic health record data

Area of Science:

• Computational Medicine
• Health Informatics
• Machine Learning Applications

Background:

• Electronic health records (EHRs) are rich data sources for identifying health outcomes but present significant challenges for computable phenotyping.
• Existing methods often struggle with the complexity and varied formats of data within EHRs, limiting accurate health outcome identification.

Purpose of the Study:

• To critically evaluate the utility of machine learning (ML) for identifying health outcomes from EHR data.
• To define four common scenarios for ML application in EHR-based phenotyping, considering diagnostic criteria and data formats.

Main Methods:

• Reviewed conditions where ML excels, focusing on diagnostic criteria complexity and data storage formats (structured vs. unstructured).
• Defined four joint scenarios by considering diagnostic criteria and data format dimensions.
• Illustrated scenarios with examples and discussed ML applications in recent studies, accounting for EHR data limitations.

Main Results:

• ML is particularly useful for health outcomes with complex, vague, or subjective diagnostic criteria, modeling intricate decision-making processes.
• ML, through natural language processing and image recognition, can extract and structure information from unstructured EHR data (text, images).
• Four distinct scenarios were identified, detailing ML's potential use in both ideal and real-world EHR data conditions.

Conclusions:

• Machine learning holds significant potential to enhance the accuracy and efficiency of health outcome identification from EHRs.
• The effectiveness of ML in EHR phenotyping is condition-dependent, particularly benefiting complex criteria and unstructured data.
• Future research should focus on improving the transportability of ML algorithms for multi-site EHR data to promote wider adoption.