Predicting progression-free survival (PFS) in first-line (1L) immune checkpoint inhibitor (ICI)-treated patients (pts) with advanced non-small cell lung cancer (aNSCLC): Machine learning (ML) application in real-world data

Anti-programmed cell death (ligand)-1 ICIs are the standard of care 1L treatment for pts with aNSCLC without actionable oncogenic driver mutations. Heterogeneity in ICI utilization and outcomes is observed in real-world pts. By leveraging ML in electronic health records (EHRs), we identified predictors influencing key clinical outcomes, and developed a nomogram to predict likelihood of median PFS in 1L ICI-treated aNSCLC pts for potential risk stratification for interventions.

1L ICI-treated aNSCLC pts without ALK/ROS1/KRAS/BRAF/EGFR alterations were identified in a US oncology EHR database (ConcertAI; Jan 2015-Feb 2023). Survival ML models were trained on 112 clinical and demographic features with 5-fold nested cross-validation. The top predictors, determined by the best-performing ML model through SHapley Additive exPlanations and clinical judgement, were used to create a Cox proportional hazard (CPH) nomogram of median PFS. All models were evaluated using the concordance index (c-index). Patients were categorized as having high and low risk of progression/death at median PFS according to median risk predicted by the nomogram.

The study cohort had 4668 pts (median PFS: 6.1 months; 3811 events). The CPH nomogram predicting 6-month PFS had a c-index of 0.60 with the top 10 predictors identified using the XGBoost model (c-index: 0.62). Nomogram predictors included increased number of metastases, ECOG PS, cough suppressant use, WBC counts, and NLR (Table). Median probability of 6-month PFS for low- and high-risk groups was 35.9% and 19.4%, respectively.

Future research should validate these findings and evaluate the opportunity to guide clinical practice to optimize outcomes in aNSCLC pts treated with 1L ICIs.