Summary
Immune-related pneumonitis (IRP) is a rare adverse event (AE) associated with ICIs that may result in treatment interruption or death if not managed properly. IRP prediction is critical to identify at-risk patients and estimate background IRP risk for single-arm trials. ML modeling was used on electronic health records (EHR) to predict IRP risk in patients with aNSCLC using various IRP definitions.
Methods
Patients with curated aNSCLC treated with ICIs between Jan 2015 - May 2022 were identified in a US-based oncology EHR database (ConcertAI®, Cambridge, MA). Survival XGBoost models were trained with 4-fold cross-validation. Due to lack of consensus on the IRP definition using RWD, putative IRPs were derived from a mix of human-curated AEs and ICD diagnosis (dx) codes (preexisting [<90 days of ICI start] codes excluded). Three IRP definitions were used: expert-curated (IRP1, n=237 IRP events), literature-derived (IRP2, n=445), and a broad definition (IRP3, n=1213). Risk factors were identified using SHapley Additive exPlanations (SHAP) values and marginal hazard ratios.
Results
The study cohort had 8928 patients. Validation cumulative/dynamic area under the ROC curve for IRP1-3 was 0.68, 0.70, and 0.71, respectively. Common across IRP definitions were predictors related to underlying lung conditions, comorbidities, polypharmacy, systemic inflammation, and de novo metastases (Table 1). Chemoradiation was a risk factor for the narrow definition (IRP1), while labs suggestive of organ dysfunction were predictive for broader definitions (IRP2-3). Unsupervised clustering of SHAP resulted in 4 patient clusters. Higher-risk clusters were driven by polypharmacy (IRP1-3), reduced platelets (IRP1), and comorbidities (IRP2-3).
Conclusions
ML on different IRP definitions from RWD differentiated risk factors of pneumonitis from those inherent to lung disease among aNSCLC patients.
