Toward a Computable Phenotype for Determining Eligibility of Lung Cancer Screening Using Electronic Health Records

Area of Science:

• Medical Informatics
• Public Health
• Radiology

Background:

• Lung cancer screening (LCS) using low-dose computed tomography (LDCT) can reduce mortality but faces barriers due to high false-positive rates.
• Electronic health records (EHRs) offer a potential data source for identifying eligible individuals for LCS.
• Developing accurate computable phenotype (CP) algorithms is crucial for enhancing LCS utilization.

Purpose of the Study:

• To develop and validate CP algorithms using EHR data to identify individuals eligible for LCS.
• To improve the accuracy and efficiency of LCS eligibility determination in real-world settings.
• To leverage both structured and unstructured EHR data for enhanced CP performance.

Main Methods:

• A cohort of 5,778 individuals undergoing LDCT for LCS between 2012-2022 was analyzed.
• CP rules based on LCS guidelines (USPSTF 2013 & 2020) were developed using structured EHR data and natural language processing of clinical notes.
• Manual chart reviews of 453 individuals validated the CP rules, assessing performance using F1 score, specificity, and sensitivity.

Main Results:

• An optimal CP rule integrating structured and unstructured data was developed, adhering to 2013 and 2020 LCS guidelines.
• The CP rule incorporated age, smoking status, and pack-years, achieving F1 scores of 0.75 (2013) and 0.84 (2020).
• Inclusion of unstructured data improved F1 scores by up to 9.2% (2013) and 12.9% (2020) compared to structured data alone.

Conclusions:

• EHR-based CP algorithms are effective in identifying individuals eligible for LCS.
• The study highlights the need for improved smoking documentation in EHRs.
• AI techniques and integrated data significantly enhance CP performance, supporting clinical decision-making and optimizing patient care for lung cancer screening.