An explainable machine learning model demonstrated utility in predicting the risk of high-grade treatment-related adverse events in patients with resectable non–small cell lung cancer (NSCLC) receiving neoadjuvant chemoimmunotherapy.  

“We see this tool primarily as a potential aid for pretreatment risk stratification and closer blood count monitoring,” said presenting author Fei-Hang Zhi, of Shanghai Pulmonary Hospital, School of Medicine, Tongji University, China, at the International Association for the Study of Lung Cancer (IASLC) 2026 World Conference on Lung Cancer (WCLC; Abstract OA11.04). 

Currently, no established framework exists for predicting treatment-related adverse events in patients receiving neoadjuvant chemoimmunotherapy. Zhi pointed to existing tools such as CARG and CRASH, which were developed primarily for older adults receiving chemotherapy and thus may not be applicable to this patient population.  

The researchers sought to develop a pragmatic, interpretable model using routine pretreatment data to identify patients at high risk for adverse events before starting chemoimmunotherapy and to support monitoring and perioperative planning.  

The model was developed using a cohort of 953 patients from Shanghai Pulmonary Hospital with stage IB to IIIB resectable NSCLC who received neoadjuvant chemoimmunotherapy. The cohort was divided into training (70%) and testing (30%) sets, with 10-fold cross-validation. The model was then externally validated in an independent cohort of 143 patients from three centers.  

Overall, 86.4% of patients underwent surgery. Squamous cell carcinoma was the most common histology (65.7%), followed by adenocarcinoma (22.0%) and other histologies (12.3%). Most patients received standard-intensity chemotherapy (70.6%) and more than two treatment cycles (79.8%).  

In the total analytic population, 45.3% of patients experienced a grade 3 or higher treatment-related adverse event during neoadjuvant chemoimmunotherapy. Neutropenia and leukopenia were the most common events.  

Study and Model Findings 

The researchers found an association between pathologic response and treatment-related adverse events. Major pathologic response rates were higher among patients who experienced high-grade treatment-related adverse events in both the discovery (P = .005) and validation (P = .033) cohorts.  

Among patients who experienced severe treatment-related adverse events, female sex, adenocarcinoma histology, and elevated baseline calcium levels were association with non-major pathologic response.  

Feature selection using multiple machine learning techniques identified 10 final predictors of high-grade treatment-related adverse events: chemotherapy dose intensity, calcium level, white blood cell count, absolute neutrophil count, albumin level, potassium level, platelet count, basophil count, cN stage, and aspartate aminotransferase level.  

Among the eight models evaluated, the Extra Trees model performed best, achieving an area under the curve (AUC) of 0.800 (95% confidence interval [CI] = 0.748–0.848) and an accuracy of 73.8% in the test set. The Random Forest and Gradient Boosting models had the next-highest AUCs, at 0.795 and 0.790, respectively. The Naive Bayes model had the lowest AUC, at 0.690, as well as the lowest specificity.  

In the external validation cohort, the Extra Trees model maintained predictive performance, with an AUC of 0.728 (95% CI = 0.644–0.808) and an accuracy of 67.8%.  

SHapley Additive exPlanations (SHAP) analysis was used to assess the contribution of individual features and revealed nonlinear, clinically interpretable risk patterns. Dose intensity was the dominant predictive feature. Baseline hematologic measures also demonstrated nonlinear effects, underscoring their contribution to predicted risk.  

“Our model showed moderate discrimination in both the internal and external validation [sets]. Its predictive performance was largely driven by the severe pathologic events,” Zhi said.  

The researchers used the model to develop an online risk calculator for severe treatment-related adverse events during neoadjuvant chemoimmunotherapy for NSCLC. Using the 10 selected baseline features, the calculator estimates a patient’s risk and assigns a risk classification.  

“Importantly, this model is intended to support risk-adapted monitoring rather than to mandate chemotherapy dose reduction or treatment modification,” Zhi added. He noted that the researchers plan to conduct more refined analyses and identify more effective predictors to improve the model’s performance.  

Expert Insights 

Discussant Yu Zhi Zhang, MBBS, PhD, of the Department of Histopathology at Royal Brompton and Harefield Hospitals in London, compared the study population with that of CheckMate 816, which evaluated neoadjuvant nivolumab plus chemotherapy in patients with resectable NSCLC. Dr. Zhang noted that the rate of grade 3 or higher treatment-related adverse events was higher in the current study than in CheckMate 816 (45.3% vs 33.5%). 

 He also commented on the variables associated with non-major pathologic response. “Potentially this investigation into the treatment-related side effects has got some utility for our oncology colleagues in evaluating potential treatment response,” he said. 

Dr. Zhang commended the researchers for developing the open-source calculator as a practical tool with potential application in patient care.  

However, he questioned whether the study population was representative of the broader population of patients with resectable NSCLC receiving neoadjuvant chemoimmunotherapy. He also questioned how the model would compare with other recognized variables and biomarkers associated with treatment-related adverse events, including tumor mutational burden and tumor-infiltrating lymphocyte density in the primary tumor bed.

DISCLOSURES: The presenting author and discussant reported no relevant financial relationships. 

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO^®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.