Machine learning modeling of real-world primary resistance and hyperprogression in immune checkpoint inhibitor (ICI) treated advanced non-small cell lung cancer (aNSCLC) patients.

Abstract

e21127 Background: Many ICI-eligible patients display primary resistance (10-res), with some experiencing hyperprogressive disease (HPD). Machine learning (ML) on real-world data (RWD) can improve our clinical and mechanistic understanding related to ICI response. Methods: XGBoost models were built on ConcertAI’s oncology clinicogenomic database, comprising structured EMR, curated records from unstructured documents, and NGS reports, to predict 10-res and HPD in aNSCLC patients treated with ICI or chemotherapy (CT) within their 1st three lines between January 2015 and June 2021. HPD was defined as either tumor progression, treatment (Tx) discontinuation, or death within 49 days from Tx start (index date). 10-res was defined as any of these events after 49 days but within 180 days from Tx start. Models were trained on a clinical (N = 7647, ICI = 3265, 10-res = 3909, HPD = 2076), and a clinicogenomic (CG) (2304, 1271, 1084, 596) cohort. A 60:20:20 split was used for training, validation, and testing. Baseline period was defined as 180 days prior to the index date. Prognostic and predictive factors were identified from Shapley Additive Explanation (SHAP) values. Results: The clinical model was superior at predicting HPD; the CG model was better at predicting 10-res (AUROC = 0.67 for both). Total metastases (≥ 2) and lower levels of hemoglobin (HGB), and lymphocytes, higher ALP were risk factors for both HPD and 10-res. Elevated WBC and bilirubin, higher heart rate, lower body temperature, and history of anti-inflammatory medications were risk factors for HPD. Non-ICI-based regimens, low TMB, lower PD-L1 expression, and mutations in EGFR or KEAP1 were risk factors for 10-res. Mutation in KRAS was protective of 10-res. In addition to PD-L1 and TMB, the models identified prior lung radiation, smokers, higher baseline comorbidities, and positive MMR status as predictive of lower risk of 10-res from ICI but not in CT. STK11 mutation was predictive of higher risk of 10-res from ICI vs. CT. Non-adenocarcinoma histology was predictive of higher risk of HPD from ICI vs. CT. Conclusions: ML on RWD generated evidence to support both established and emerging prognostic and predictive markers for ICI response.[Table: see text]