Individualized Prediction of Early Post-Treatment Mortality Risk in Stage I Non-Small Cell Lung Cancer: A Machine Learning-Based Approach

Abstract

In stage I non-small cell lung cancer (NSCLC), the STARS/ROSEL pooled analysis showed worse overall survival (OS) but equivalent cancer-specific survival with surgical resection compared to stereotactic body radiation therapy (SBRT), with other studies demonstrating interaction effects between early post-treatment mortality and age. Therefore, appropriate patient selection is essential to optimize survival outcomes. We utilize predictive modeling using machine learning (ML) of mortality risk at 90 days post-treatment in stage I NSCLC patients. Patients with T1-T2aN0M0 NSCLC, diagnosed from 2004 to 2015, were identified using the National Cancer Database and divided into separate surgery and SBRT cohorts. All patients were censored 90 days after receipt of definitive treatment. Training and validation sets were formed by a randomized 80%/20% split. Features (age, histology, tumor size, and comorbidity score) were pre-specified based on predictors of post-treatment mortality identified in literature. Three different ML approaches were tested: 1) a Cox proportional hazards (CPH) linear model, 2) a survival Support Vector Machine (SVM) model, and 3) a gradient-boosted regression tree (GBRT) model using Cox proportional hazard loss. Hyperparameter tuning was performed manually using grid search techniques. Model performance was evaluated using Harrell’s concordance index (CI). Feature importance and interaction effects were modeled using Shapley additive explanations (SHAP) values applied to the GBRT model. The open-source scikit-survival library in Python 3.7 was used for all analyses. We identified 118,087 stage I NSCLC patients (81.0%) receiving surgery and 27,659 receiving SBRT (19.0%). Within the surgery cohort, the GBRT model had the best performance (CI=0.689), as compared to the SVM model (0.673) and CPH model (0.672). Within the SBRT cohort, similarly, the GBRT model performed the best (CI=0.591) compared to the SVM model (CI=0.572) and CPH model (0.570). Based on feature importance rankings, age had the highest importance in both the surgery and SBRT cohorts. When using SHAP values to model nonlinear interaction effects between age and treatment received, an inflection point was identified between age 73 and 74, indicating GBRT model predictions of generally greater 90-day post-treatment mortality with SBRT in patients age 18-73 and greater risk with surgery in patients age 74+. We describe a ML-based approach for analyzing comparative 90-day post-treatment mortality risk between surgery and SBRT in stage I NSCLC. We found an inflection point in the interaction between age and treatment modality, underscoring the need for patient selection when assessing treatment options. Our findings may aid individualized decision-making in early stage NSCLC.