Classifying Pseudogout Using Machine Learning Approaches With Electronic Health Record Data.

Area of Science:

• Rheumatology
• Medical Informatics
• Artificial Intelligence

Background:

• Pseudogout, an acute subtype of calcium pyrophosphate (CPP) deposition disease, is challenging to identify in large datasets due to its episodic nature and lack of specific billing codes.
• Current methods for identifying pseudogout in electronic health records (EHRs) are limited, impacting research and clinical care.

Purpose of the Study:

• To evaluate a novel machine learning (ML) approach for classifying pseudogout using EHR data.
• To develop and assess an algorithm that improves the accuracy of pseudogout identification compared to existing methods.

Main Methods:

• A cohort of 900 patients was curated from EHR data (1991-2017) based on billing codes or natural language processing (NLP) mentions of pseudogout or chondrocalcinosis.
• Gold standard chart review was performed to confirm definite or probable pseudogout.
• A combined algorithm integrating NLP, ML topic modeling, and synovial fluid laboratory results (CPP crystals) was developed and evaluated.

Main Results:

• The combined algorithm achieved a sensitivity of 42% and a positive predictive value (PPV) of 81% for pseudogout.
• This algorithm identified 50% more pseudogout cases than relying solely on the presence of CPP crystals.
• Traditional billing codes demonstrated a low PPV of 22% for pseudogout identification.

Conclusions:

• Combining NLP, ML, and laboratory data significantly enhances the PPV for identifying pseudogout in EHRs compared to billing codes alone.
• This novel ML approach offers a more effective strategy for classifying pseudogout, a condition often underidentified due to its clinical characteristics.
• The developed algorithm holds promise for improving the study of pseudogout in large patient populations.