知镜 Cognispect · 医学文献与研究工作台

Predicting immunotherapy response in patients with non-small cell lung cancer: superior performance of a positron emission tomography radiomics-based stacking ensemble model

Zhang L, Zeng Z, Sheng P, et al.
Quantitative imaging in medicine and surgery · 2026-09-29 · PMID 42997730

文献摘要

Background The efficacy of immunotherapy among patients with non-small cell lung cancer (NSCLC) varies, and thus reliable predictive biomarkers of response are needed. This study aimed to develop and validate a stacking ensemble learning model based on pretreatment positron emission tomography (PET) radiomics to predict response to immunotherapy in patients with NSCLC. Methods In this retrospective study, 45 patients with NSCLC treated with immunotherapy were enrolled. Pretreatment Digital Imaging and Communications in Medicine-format PET images were preprocessed and normalized. Tumor volumes were segmented using three methods: manual delineation, fixed standardized uptake value (SUV) thresholding with a standardized uptake value of 2.5 (SUV2.5), and adaptive thresholding (40% of the maximum SUV). Radiomics features were extracted, followed by a multistep feature selection process involving mutual information, Spearman correlation, and least absolute shrinkage and selection operator (LASSO) regression. Predictive models constructed included a stacking ensemble model [with k-nearest neighbors (KNN), logistic regression (LR), and random forest (RF) as base learners], a clinical feature-based model, and a deep learning model. Model performance was evaluated on an independent test set according to the area under the curve (AUC), accuracy, precision, recall, and F1 score. Results On the test set, the stacking ensemble model achieved the highest predictive performance using features from the SUV2.5 segmentation dataset, with an AUC of 0.933 [95% confidence interval (CI): 0.846-0.994]. This significantly outperformed the best single model (LR; AUC=0.896) and a deep learning model (AUC =0.925). The clinical feature-based model showed moderate predictive capability, with extreme gradient boosting (XGBoost) yielding the highest AUC of 0.877 (95% CI: 0.644-1.000). Conclusions The proposed PET radiomics stacking ensemble model, which leverages features from multiple segmentation strategies, demonstrated superior predictive performance for immunotherapy response in NSCLC compared to single models and other benchmarks. This approach offers an effective strategy for enhancing robustness in small-sample radiomics studies and holds promise as a noninvasive decision-support tool that can aid in personalized immunotherapy stratification.

在知镜精读原文来源DOI 原文

知镜提供医学文献检索、阅读笔记、人体图谱、统计绘图与研究复现工具。