Ensemble principal component analysis-based multi-omics for radiation pneumonitis prediction in stage III non-small...

Conventional univariate feature-selection pipelines for radiation pneumonitis (RP) prediction are prone to overfitting in small radiotherapy cohorts, while alternative dimensionality-reduction approaches remain insufficiently evaluated. This study compared a Principal Component Analysis (PCA)-based ensemble pipeline comprising variance pre-filtering, PCA, and LASSO with a conventional pipeline comprising Mann-Whitney U testing, Spearman correlation filtering, and ElasticNet. The analysis included 73 retrospectively analyzed patients with stage III non-small cell lung cancer (training, n = 52; testing, n = 21). A total of 214 features, including 107 radiomic and 107 dosiomic features, were extracted from planning CT images and three-dimensional dose distributions. Both pipelines used five-model ensemble logistic regression with inverse-frequency weighting. Performance was evaluated using testing AUC, the training-to-testing AUC gap, 10-fold nested cross-validation, paired bootstrap comparison of AUC differences with 2,000 resamples, Brier score, and bootstrap feature-selection stability. The PCA-based model achieved a testing AUC of 0.765 with a training-to-testing AUC gap of 0.020, compared with an AUC of 0.684 and a gap of 0.181 for the conventional model (ΔAUC = + 0.081, p = 0.265). The PCA-based model also achieved a nested cross-validation AUC of 0.742. PCA-based dimensionality reduction produced a smaller training-to-testing AUC gap and greater feature-selection stability than conventional univariate filtering, indicating improved internal robustness and reduced overfitting in this small-cohort setting.