This letter examines the interpretation and clinical translation of a meta-analysis of radiomics-based artificial intelligence (AI) for predicting pathological response after neoadjuvant immunochemotherapy in resectable non–small cell lung cancer. We highlight that the conventional-response comparator combines PERCIST and RECIST 1.1 assessments from the same 36-patient cohort, whereas the AI estimates arise from unmatched cohorts; consequently, the reported denominator and Z tests do not establish comparative superiority. We further consider how restriction to patients who reached resection and variation in imaging timepoints narrow the clinical estimand and limit inference about earlier treatment redirection. Clarification of the RECIST and error-direction examples is also warranted. We propose same-patient comparisons in treatment-initiation cohorts using fixed imaging times, locked thresholds, explicit primary-tumor and nodal labels, and calibration and net-benefit analyses at prespecified clinical thresholds.