AMMR-Surv: Learning Metabolic and Morphologic Responses Across Lesions from Longitudinal PET/CT for Survival Prediction in Non-Small-Cell Lung Cancer

Lesion responses to chemoradiotherapy can differ within a patient, complicating prognostic assessment in locally advanced non-small-cell lung cancer. We developed AMMR-Surv to integrate prototype-based metabolic response learning, asymmetric CT comparison, and event-aware multi-lesion aggregation. This retrospective study included 126 patients from ACRIN 6668/RTOG 0235 with paired fluorodeoxyglucose positron emission tomography/computed tomography (FDG-PET/CT) before and after treatment. Overall survival was measured from post-treatment examination. Five-fold nested patient-level cross-validation was repeated five times, with prototype counts selected within each outer training partition. Aggregated out-of-fold Uno’s concordance was 0.719 (95% confidence interval [CI], 0.670–0.772), with a 24-month area under the curve of 0.765 and an integrated Brier score of 0.1747. Paired concordance differences were 0.007 (95% CI, −0.021 to 0.046) versus generic longitudinal modeling and 0.059 (95% CI, 0.005–0.114) versus clinical variables with simple PET measurements. Across three runs on 120 common lesion sites, aligned prototype-assignment agreement ranged from 93.33% to 96.67%. These results support further evaluation of lesion-response representations, although superiority over generic longitudinal modeling remains uncertain. The small, selected cohort limits generalizability, and earlier fixed-four-prototype analyses remain exploratory because that configuration was selected using test performance. Independent validation is required before clinical application.