WCLC 2025 - Posters & ePosters-P1.04.36 LungFlagTM Risk Prediction Validation on Canadian Ever Smokers Pre-Classified as High Risk for Lung Cancer

Pdf Summary

This study validates the LungFlagTM machine learning model for lung cancer risk prediction in a Canadian population already classified as high-risk due to smoking history. LungFlag uses routine outpatient laboratory data—including blood counts, chemistry tests, comorbidities, spirometry, and PLCOm2012 risk scores—to identify individuals at elevated risk. The research leveraged two Canadian datasets: the Lung Screening Population (LSP) cohort (individuals aged 55-80 with ≥30 years smoking history) and the PanCan longitudinal trial cohort (aged 50-75 with ≥2% six-year lung cancer risk per PLCOM2008). Participants included 3,051 ever-smokers, with 2,958 meeting eligibility criteria; 49 developed lung cancer within a 2-year prediction horizon.<br /><br />The study applied LungFlag to these cohorts, comparing its predictive accuracy against established models—PLCOm2012 and USPSTF2013 criteria—using sensitivity, specificity, predictive values, and area under the curve (AUC). LungFlag demonstrated comparable or slightly superior performance, with AUCs of 0.707 for LungFlag versus 0.693 for PLCOm2012 in the 2-year risk window among people aged 50-80 who had smoked. Among participants who did not meet traditional screening eligibility but included four lung cancer cases, LungFlag achieved an AUC of 0.764, suggesting potential to identify cancers missed by existing criteria.<br /><br />Demographically, lung cancer cases were older, had longer smoking durations (average 42.9 vs. 36.5 years), and higher COPD prevalence (35% vs.14%) compared to controls. The 2-year LungFlag risk horizon performed comparably to longer horizons, indicating practicality for clinical application.<br /><br />In conclusion, LungFlag is a feasible tool for lung cancer risk prediction in Canadian electronic health record data and shows non-inferior accuracy to standard clinical risk models, even within a pre-selected high-risk group. Further prospective studies are recommended to assess LungFlag as an independent classifier to refine screening eligibility and improve early lung cancer detection.<br /><br />This work was supported by Roche, BC Cancer Foundation, and Terry Fox Research Institute.

Asset Subtitle

Sukhinder Atkar-Khattra

Meta Tag

Speaker Sukhinder Atkar-Khattra

Topic Screening and Early Detection

Keywords

LungFlag

lung cancer risk prediction

machine learning model

Canadian population

high-risk smokers

PLCOm2012

USPSTF2013 criteria

lung cancer screening

electronic health records

lung cancer early detection