1561P - Targeted screening methodologies to select high risk individuals: LungFlag performance in Estonia Lung Cancer Screening Pilot

Background

In Estonia, based on WHO data - LC accounted for 809 new cases and 669 deaths in 2020. The first stage in the implementation of screening was a feasibility pilot conducted in 2021 to provide low-dose computed tomography (LDCT) screening for high-risk groups chosen either by USPSTF 2021 smoking criteria and/or PLCOm2012_norace(>1.5% cutoff). LungFlag™ (LF) is a machine learning risk score for Lung Cancer that uses EMR data as input, it was used retrospectively in the second phase to assess risk among those individuals that performed LDCT in the pilot.

Methods

The goals of the study were to evaluate whether LF can be used on existing pilot EMR Data collected, and the potential value of LF in this setting and as a potential tool to select individuals for screening. LF was run on EMR data of individuals who performed LDCT in the pilot, minimal mandatory information of age, sex and smoking status data was required. Top 5% of individuals at risk were selected by each method and performance was compared by AUC, ranking and average age of flagged individuals. Note since LF was used retrospectively performances were limited by selected cohort.

Results

Total data included 3,708 individuals classified as high-risk by either USPSTF criteria or PLCO, out of those 3,695 were included in the comparison analysis. A total of 3,444 individuals performed LDCT identifying 31 LC cases. Average smoking years were 40, over 75% classified as current smokers. LF scores were able to be calculated for all individuals. LF ranked the risk in the study population better than the PLCO model (AUC of 0.697 vs 0.674) and significantly better than USPSTF criteria. Average age of top 5% flagged individuals by LF was over 2 years younger than PLCO (68.5 vs. 71).

Conclusions

LungFlag was able to work on Estonian EMR data, based on the pilot data and follow up it outperforms PLCO in ranking individuals that were later on diagnosed with lung cancers and improved the False Positive rate, flagging younger population. External retrospective studies suggest that LF might be even a better classifier if it was allowed to independently select high-risk people for screening from the ever smokers population in Estonia in on-going basis – additional prospective data would be necessary to prove that.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

E. Choman.

Funding

F. Hoffmann-La Roche Ltd, Basel, Switzerland.

Disclosure

All authors have declared no conflicts of interest.