Abstract 194P
Background
Despite the current success of immune checkpoint inhibitors (ICI), there is an ongoing search for new and better biomarkers to identify which non-small cell lung cancer (NSCLC) patients are likely to respond. While upfront evaluation of biomarkers might provide prognostic information, it does not provide information on the actual tumor response. Serum tumor markers (STM) are known to reflect tumor activity and might therefore be useful in early response prediction. This study aims to compare a wide range of prediction methods applied to multiple sequentially measured serum tumor markers regarding their accuracy in predicting non-response in NSCLC patients receiving ICI therapy.
Methods
Bi-weekly measurements of CYFRA, CEA, CA-125, NSE, and SCC were performed in a cohort of 412 NSCLC patients treated with ICIs. Disease progression was determined using RECIST criteria in combination with clinical assessment. Nine prediction models were assessed to predict treatment response at 6-months based on STM measurements taken during the first 6-weeks of therapy: logistic regression, quadratic discriminant analysis (QDA), LASSO, random forest, bagging, boosting, support vector machines, and (recurrent) neural networks. All methods were applied to six different biomarker combinations ranging from two to five STMs per combination. Patients were randomly divided over a training (75%) and validation cohort (25%). Model performance was compared based on the sensitivity, while model training aimed at 95% specificity to ensure a low false-positive rate.
Results
In the validation cohort, boosting provided the highest sensitivity across most STM combinations (2.4% to 68.8%). When all five STMs were included, QDA provided the highest sensitivity (68.8%) while the specificity dropped to 46%. Boosting applied to CYFRA and CEA achieved the highest sensitivity (59.4%) on the validation data while maintaining a specificity >95%.
Conclusions
Non-response in NSCLC patients treated with ICIs can be predicted by combining multiple sequentially measured STMs in a prediction model. Boosting provided the best overall performance, however, clinical use of the preferred model is subject to further validation.
Legal entity responsible for the study
The authors.
Funding
Has not received any funding.
Disclosure
S. Burgers: Non-Financial Interests, Personal, Advisory Board: Roche, Bristol Myers Squibb; Financial Interests, Personal, Other, Investigator Initiated Study: MSD. All other authors have declared no conflicts of interest.