Abstract / Summary
Abstract Background To develop, as a proof-of-concept analysis, a machine learning model predicting early radiological treatment response in an elderly, single-center esophageal cancer (EC) cohort, and to test whether adding clinical factors to CT radiomics improves prediction. Methods We retrospectively reviewed 189 patients with EC treated with definitive radiotherapy. Radiomic features were extracted from baseline contrast-enhanced CT. Following TRIPOD recommendations, a LASSO-penalized logistic regression (LASSO-LR) on the radiomic feature set was pre-specified as the primary model, with the number of predictors capped at five (events per variable [EPV] ≥ 10; 50 events). Performance was estimated by 5 × 5 nested cross-validation and in an independent test set (8:2 split; 151 training, 38 test), with out-of-fold calibration; a sensitivity analysis repeating the full pipeline within each outer fold and a 200-resample feature-stability analysis were also performed. Five classifiers across clinical-only, radiomics-only and multimodal feature sets, including support vector machines (SVM), were exploratory comparisons only; incremental value was tested by bootstrap paired comparison (5,000 resamples). Results The cohort comprised 189 patients (126 responders, 63 non-responders). The pre-specified LASSO-LR primary model retained five radiomic predictors, selected from 21 LASSO features by absolute coefficient, and achieved a fully nested cross-validation AUC of 0.554 ± 0.108, obtained when univariable filtering, feature selection, and model fitting were re-executed within each outer training fold; a partially nested (development) estimate of 0.637 ± 0.105 (optimism 0.148; apparent 0.785), derived after univariable filtering had been performed once on the complete training cohort, and an independent test-set AUC of 0.716 in a very small test cohort ( n = 38) are also reported; out-of-fold calibration gave a Brier score of 0.216 (intercept 0.328, slope 1.502). Exploratory clinical-only, radiomics-only and multimodal SVM models reached test AUCs of 0.603, 0.732 and 0.745; the multimodal model showed no statistically significant incremental value over clinical-only (ΔAUC + 0.142, 95% CI − 0.077 to 0.364, P = 0.212) or radiomics-only (ΔAUC + 0.012, 95% CI − 0.033 to 0.064, P = 0.520) models. Multimodal discrimination was similar across age strata (AUC 0.739 vs. 0.786). Conclusions This proof-of-concept study indicates that a pre-specified LASSO-LR model based on five radiomic features achieves limited and unstable discrimination (fully nested AUC 0.554 ± 0.108; test-set AUC 0.716 in a very small test cohort) for early radiological response in esophageal cancer. Exploratory SVM modeling reached a numerically higher but statistically non-significant multimodal AUC, providing no evidence of incremental value over radiomics or clinical factors alone. Prospective multicenter validation is required before any clinical application. Clinical trial number Not applicable.