Can radiomics combined with clinical data predict checkpoint inhibitor pneumonitis?

Abstract

e14555 Background: Immunotherapy (ImT) is increasingly utilized in patients with advanced non-small cell lung cancer (aNSCLC). Checkpoint-inhibitor pneumonitis (CIP) is an uncommon and potentially life-threatening adverse event (AE) associated with ImT. For patients with aNSCLC, CIP signs and symptoms can be misclassified in the setting of underlying lung disease and incidence may be underreported. The ability to identify patients at risk for CIP prior to ImT could prevent significant morbidity. Previously, we reported that radiomics, a datamining technique that extracts patterns from medical imaging, had identified texture features on pre-ImT CT that correlate with CIP. In this study, we hypothesized that such features combined with clinical data can predict CIP. Methods: In an IRB-approved database of 129 patients with aNSCLC treated with nivolumab, 9 controls were identified with clinical diagnosis of CIP. For all patients, uninvolved lung in the last pre-ImT CT was mined for textures associated with CIP. In the 9 controls, 3 features were identified that correlated with CIP. Imaging and patient charts were reviewed for signs and symptoms of CIP within the first 6 months after ImT administration. Non-controls with progressive clinical symptoms or radiographic findings consistent with CIP and unexplained by disease progression, infection or other medical interventions were treated as CIP-positives. Lastly, the area under the receiver operating curve (AUC) for predicting CIP was built progressively combining uncorrelated laboratory, imaging, and clinical data. Results: In all, 25 patients (19%) were identified with any-grade CIP. A model based on pre-ImT platelets (PLT) and neutrophil-to-lymphocyte ratio (NLR) had an AUC for predicting CIP of 0.648. The radiomics features associated with CIP were average gray value (p = .002), 3D normalized entropy (p = 0.022), and kurtosis (p = 0.011). When combined these 3 features had an AUC of 0.735. Combining both radiomics and laboratory data produced an AUC of 0.779. This increased to 0.793 with the addition of ECOG performance status and sex. Finally, incorporating smoking status yielded an AUC of 0.802. Paired sample difference in AUC between the final model and that including only PLT and NLR was significant (p = 0.003). Conclusions: In patients with aNSCLC treated with ImT, CIP is an uncommon but potentially serious AE. The ability to identify patients at higher risk for CIP could prevent significant morbidity. A model combining radiomics of pre-ImT imaging, laboratory, and clinical data showed better prediction for CIP than laboratory data alone. Future directions include validation of this model on an external cohort and with patients on different ImT regimens.