Work

Explainable machine learning for predicting grade ≥3 treatment-related adverse events in resectable non-small cell lung cancer receiving neoadjuvant chemoimmunotherapy: predominant contribution of hematologic toxicity

Year: 2026

Type: article

Abstract: Background: Neoadjuvant chemoimmunotherapy improves outcomes for resectable non-small cell lung cancer (NSCLC), but can cause severe treatment-related adverse events (TRAEs). Pretreatment models for predicting grade ≥3 TRAEs during neoadjuvant chemoimmunotherapy remain limited. This study aimed to develop and externally validate an explainable machine learning model for predicting grade ≥3 TRAEs in patients with resectable NSCLC receiving neoadjuvant chemoimmunotherapy. Methods: This multicenter retrospective study included a discovery cohort (n=953) and an independent external validation cohort (n=143). The discovery cohort was stratified into training (n=668) and internal test (n=285) sets. Pretreatment candidates included routine clinical and laboratory variables and first-cycle treatment characteristics. Stability-oriented Boruta selection was used for variable selection, followed by variance inflation factor (VIF) assessment. Eight machine learning (ML) algorithms were tuned using 10-fold cross-validation in the training set. Discrimination, calibration, decision curve analysis (DCA), and explanations based on SHapley Additive exPlanations (SHAP) were evaluated. Secondary analyses assessed hematologic, nonhematologic, and immune-related adverse event (irAE) endpoints. Results: Grade ≥3 TRAEs occurred in 496 of 1,096 patients (45.3%); 464 patients (42.3%) experienced grade ≥3 hematologic toxicity, and 446 (89.9% of patients with grade ≥3 TRAEs) reached the composite endpoint exclusively because of hematologic events. The final random forest model incorporating 12 pretreatment predictors achieved areas under the receiver operating characteristic curve (AUCs) of 0.782 and 0.742 in the internal test and external validation cohorts, respectively. When the locked overall-TRAE model was applied to toxicity subtypes, discrimination was maintained for grade ≥3 hematologic toxicity (AUCs, 0.764 and 0.732) but was substantially weaker for grade ≥3 nonhematologic toxicity (0.638 and 0.631) and any-grade irAEs (0.435 and 0.535). The hematologic toxicity-specific model achieved AUCs of 0.777 and 0.711 in the internal test and external validation cohorts, respectively. Conclusions: The pretreatment random forest model showed moderate external discrimination for overall grade ≥3 TRAEs, primarily reflecting hematologic toxicity.

Source: Translational Lung Cancer Research

DOI: 10.21037/tlcr-2026-0984