Mining Primary Care Electronic Health Records for Automatic Disease Phenotyping: A Transparent Machine Learning...

Area of Science:

• Health Informatics
• Machine Learning
• Clinical Epidemiology

Background:

• Developing transparent machine learning (ML) frameworks for automated patient identification from electronic health records (EHRs) is crucial.
• Existing methods often lack transparency or require extensive feature engineering.

Purpose of the Study:

• To develop a transparent ML framework for automated patient identification from EHRs using a concise set of features.
• To evaluate the framework's efficacy in phenotyping patients with rheumatoid arthritis (RA) and ankylosing spondylitis (AS).

Main Methods:

• Linked multiple EHR sources (primary, secondary, specialist) between 2002-2012, creating a unique dataset.
• Treated patient identification as a text classification problem, developing a transparent disease-phenotyping framework.
• Framework includes patient representation generation, feature selection, and optimal phenotyping algorithm development to address data imbalance.

Main Results:

• Applied to 9657 patients (1484 RA, 204 AS), the framework achieved high accuracy and positive predictive values.
• RA identification: 86.19% accuracy, 88.46% PPV.
• AS identification: 99.23% accuracy, 97.75% PPV, comparable to expert methods.

Conclusions:

• The developed framework is an efficient tool for identifying patients with specific conditions from EHRs.
• This approach can support clinical decision-making processes.
• The transparent nature of the framework enhances its potential for widespread adoption.