A reply to two letters to the editor of Neuro-Oncology

We thank Dr. Peng Sun and colleagues as well as Ryan G. Rilinger and Andrew Dhawan for their interest in our recently published article.1 We appreciate the insightful comments regarding potential overfitting, data leakage, and model validation, as well as the additional questions raised in the subsequent letter, as these are critical considerations for translating high-dimensional biomarker studies into clinical practice. Dr. Sun et al. correctly point out the risk of model overfitting and optimistic bias if feature selection is not properly confined to training data.2 In our study, we employed a cross-validated, consensus-based feature selection strategy and repeated subsampling to identify a stable metabolite panel, aiming to mitigate overfitting. However, we acknowledge that the final selection of our “compact panel” was derived from the full discovery cohort, which included all sample groups. In retrospect, adopting fully nested cross-validation (with feature selection repeated within each fold) could further strengthen the rigor of performance estimation and help guard against potential information leakage. We appreciate this suggestion and are conducting additional analyses with strict nested cross-validation to ensure our model’s robustness.