Label efficient phenotyping for Long COVID using electronic health records

Area of Science:

• Infectious Diseases
• Medical Informatics
• Public Health

Background:

• Long COVID presents a significant global health challenge due to its varied symptoms and inconsistent coding.
• Accurate identification of Long COVID cases is essential for risk prediction and effective patient management.
• Current methods relying solely on specific electronic health record (EHR) codes are insufficient for comprehensive Long COVID phenotyping.

Purpose of the Study:

• To develop and validate an efficient phenotyping algorithm for Long COVID using EHR data.
• To improve the accuracy of Long COVID classification beyond existing coding practices.
• To leverage both labeled and unlabeled EHR data for enhanced case identification.

Main Methods:

• Introduction of the LAbel-efficienT Long COVID pHenotyping (LATCH) algorithm.
• LATCH synthesizes a small set of gold-standard labels with a large unlabeled EHR dataset containing numerous features.
• Internal and external validation of the LATCH algorithm's performance.

Main Results:

• LATCH demonstrated superior performance compared to methods using only the U09.9 Long COVID EHR code.
• Downstream analysis revealed increased healthcare utilization among Long COVID patients, peaking beyond four months post-infection.
• The algorithm effectively utilizes both labeled and unlabeled data for improved classification.

Conclusions:

• The LATCH algorithm offers a more robust approach to classifying Long COVID cases.
• Enhanced classification provides critical insights into Long COVID-related healthcare utilization patterns.
• Findings can inform clinical practice and public health strategies for managing Long COVID's long-term effects.