LungFlag™, a Machine-Learning (ML) Personalized Tool for Assessing Pulmonary Complications a Community Setting, Demonstrates Comparable Performance in Flagging Non-Small Cell Lung Cancer.

Clinical trials targeting heavy smokers have shown to reduce mortality and paved the way for the USPSTF to recommend lung cancer screening based  on a combination of age, smoking history and current or recent smoking status.  However, less than 27% of Americans diagnosed with lung cancer meet the original USPSTF criteria for screening, leaving many high-risk individuals unable to access screening.

Using a risk-based selection score resulted in higher sensitivities compared to criteria using dichotomized age and smoking history. Application of ML-based risk models to cancer screening cohorts has shown increase in screening efficiency expressed by increased compliance, resources yield and increased detection of early stages in particular. The classifiers developed for that purpose used training data that are random samples of the screen-eligible population, with same distribution of the key data elements.

Therefore, ML-based models may be vulnerable to sex-and race-based bias arising from historical bias in access to healthcare as well as biased training data.  Demonstrating fairness in the predictions of ML-based models is a prerequisite to their acceptance by clinicians, patients and policy-makers.  We assessed the clinical performance of LungFlag based on sex and race, two key demographic subgroups at risk for disparate outcomes due to bias.