Introduction: Radiomics have shown promise in improving prognostication in metastatic non-small cell lung cancer (mNSCLC) subjects treated with immunotherapy (IO). However, ensuring generalizability across different centers still represents an open challenge to clinical adoption. We sought to develop and test the generalizability of a radiomics model aimed at predicting risk of progression in IO treated subjects with mNSCLC.

Methods: Pretreatment CT scans of IO treated mNSCLC subjects and with known outcome data were collected from a single institution (Discovery cohort) to develop the model. Radiomics features were extracted from the segmentation of the largest lung tumor lesion. The 8 most predictive radiomics features were selected using a least absolute shrinkage and selection operator (LASSO) cox regression and combined using a survival random forest algorithm. The radiomics risk model was trained via cross-validation using censored progression-free survival (PFS) data. To determine what learned features were predictive of IO outcome, we tested the model in a cohort of mNSCLC subjects treated with 1L chemotherapy (Chemo cohort). To test model generalizability, we used a publicly available retrospective cohort of pretreatment CT scans of mNSCLC treated with IO and with known PFS data from an independent institution (External cohort). Risk models were evaluated by splitting the data into high and low risk groups, and evaluating the hazard ratios (HR) and log rank test p-values between the predicted risk groups.

Results: The Discovery cohort included 108 mNSCLC subjects who received IO as 1L therapy. Sixty-seven (62%) subjects received a combination of IO and chemotherapy. The cohort (51% female) had a median age of 68 (range 25->89) years and a median PFS of 11.5 (95% CI 8.3-15.8) months. The chemo-cohort of 55 patients (45% female) had a median age of 65 (range 45-82) years and a median PFS of 10.3 months (95% CI 5.3-12.9). The External cohort included 174 mNSCLC subjects (52% female) who underwent IO as first (33%), second (54%) or subsequent (13%) line of therapy. Nine (5%) subjects received IO combination therapies. These subjects had a median age of 68 (range 38->89) years and a median PFS of 2.7 (95% CI 2.2-3.6) months. The model cross-validated on the discovery cohort produced a HR of 1.79 (95% CI 1.08-2.95), p-value=0.024. In the Chemo-cohort, the HR was of 1.93 (95% CI 1.07-3.47), p-value=0.026. On the independent test cohort the HR was of 1.45 (95% CI 1.04-2.03), p-value=0.029. When focusing on subjects treated with IO as 1L therapy, the HR was of 1.74 (95% CI 0.92-3.32), p-value=0.089.

Conclusions: The derived radiomics model showed promising risk stratification and generalizability capabilities in IO treated subjects. However, results from the Chemo cohort suggest the model might be predictive of overall risk of disease progression rather than response to IO therapy.

VIEW THE PUBLICATION (P2.11A.27)

VIEW THE POSTER