Background: The emergence of promising biomarkers offers new opportunities to enhance risk prediction for lung cancer. The International Agency for Research on Cancer (IARC) recently developed the 13-pre-diagnostic-protein-based 3-year lung cancer INTEGRAL-Risk score for people with a smoking history, which demonstrated strong improvement over a smoking-based score.
Objectives: This abstract includes preliminary results from our work aiming to further improve lung cancer risk prediction by incorporating repeated measurements of pre-diagnostic proteins.
Methods: This study is underway within the framework of the Lung EArly Proteins (LEAP) Project funded by the US National Cancer Institute. Proteins on the INTEGRAL panel were measured in blood samples from the Carotene and Retinol Efficacy Trial (CARET) for model development. Finally results will later be validated using samples from the Prostate, Lung, Colorectal, and Ovarian (PLCO) Cancer Screening Trial. Participants in both CARET and PLCO were restricted to current or former smokers with no history of lung cancer prior to enrollment and at least one available pre-diagnostic serum sample. Incident lung cancer cases up to a pre-defined timepoint and non-cases were selected in a 1:1 ratio, with non-cases sampled at baseline using a stratified design. Time-varying Cox regression models were fit in the CARET development set to examine associations between longitudinal protein measurements and lung cancer risk. Individual INTEGRAL-Risk scores were calculated at baseline and at each blood-draw time point. We compared the predictive performance (measured by the area under the ROC curve, AUC), with that of the baseline single-timepoint model. 
Results: We analyzed 532 lung cancer cases and 532 non-cases from the CARET cohort; 61.3% (326/532) of cases and 81.6% (434/532) of non-cases provided more than one serum samples. Among the 21 repeatedly measured proteins on the INTEGRAL panel, nine proteins (TRAIL-R2, WFDC2, SYND1, MMP12, CEACAM5, CDCP1, SCGB3A2, CXL17, LAMP3) were associated with an increased risk of lung cancer, with adjusted hazard ratios (HR) per Z-score unit ranging from 1.21 (95% confidence interval [CI]: 1.07, 1.36) to 6.12 (95% CI: 4.04, 9.29). Two proteins including SCF (0.87, 95% CI: 0.76, 0.99) and FASLG (0.75, 95 % CI: 0.65, 0.87) were inversely associated with lung cancer risk. The AUC for the Cox model including only the baseline single-timepoint INTEGRAL risk score (logit-transformed) was 0.46 (95% CI: 0.42, 0.50). The time-varying Cox model including logit-transformed INTEGRAL risk scores based on rates of change during follow-up yielded an AUC of 0.73 (95% CI: 0.68, 0.78).
Conclusions: Preliminary results from the LEAP study suggest that repeated pre-diagnostic protein measurements may improve lung cancer risk prediction compared with a single-timepoint measurement.