AI-derived longitudinal and multi-dimensional CT classifier for non-small cell lung cancer to optimize neoadjuvant chemoimmunotherapy decision: a multicentre retrospective study

Background: Neoadjuvant chemoimmunotherapy (NACI) has significantly improved survival in patients with resectable non-small cell lung cancer (NSCLC). However, with the currently available methods (PD-L1, RECIST), it is difficult to predict who will benefit from treatment before therapy and who will achieve pathological complete response (pCR) before surgery. Non-invasive methods to predict treatment response to NACI could be used to further personalize treatment. Methods: In the multicenter retrospective study, we enrolled 534 patients with NSCLC who received NACI followed by surgical resection at three Chinese hospitals between January 2019 and December 2024 (193 patients from Center A, 193 patients from Center B, and 148 patients from Center C). We developed and validated Lung Cancer Neo-adjuvant Immuno-Chemotherapy Response Predictor (LC-NICER), a CT-based Artificial Intelligence (AI) system that integrates longitudinal radiomics (tumor texture), deep learning (microenvironmental context), and habitat imaging (tumor and peritumoral subregional dynamics) to predict pCR by analyzing tumor spatiotemporal heterogeneity. The LC-NICER system consists of two complementary predictive models: LC-NICER_α, a pretreatment model that identifies patients likely to benefit from NACI to guide personalized therapy, and LC-NICER_δ, a preoperative model that evaluates tumor regression and resection feasibility to inform surgical planning. This study is registered atClinicalTrials.gov(NCT06285058). Findings: Among the 386 patients from Center A and B, 308 were randomly assigned to the training dataset and 78 to the internal validation dataset, following an 8:2 split ratio. The 148 patients from Center C formed an independent and external test dataset. LC-NICER prediction system demonstrated excellent performance with the area under the curves (AUCs) of 0.950 (0.927–0.970) in the training cohort, 0.889 (0.796–0.961) in the internal validation cohort, and 0.870 (0.803–0.927) in the external test cohort. The LC-NICER_αachieved an accuracy of 0.722 (0.668–0.772) before therapy, while LC-NICER_δshowed significantly improved accuracy of 0.831 (0.800–0.861) before surgery. Notably, LC-NICER outperformed current clinical standards, with absolute accuracy improvements of 10% over PD-L1 testing (0.622 [0.564–0.683], p = 0.002) and 18% over RECIST 1.1 criteria (0.651 [0.610–0.693], p = 0.008). For easy clinical utility and research reproducibility, we developed and openly published a software. Interpretation: As a non-invasive AI system for predicting NACI response in NSCLC, LC-NICER may offer future clinical personalized therapeutic strategies, accelerate adaptive clinical trials, and optimize treatment decisions, potentially reducing reliance on invasive procedures. Funding: This work was supported by the Noncommunicable Chronic Diseases-National Science and Technology Major Project 2024ZD0531100, 2024ZD0531101, the National Natural Science Foundation of China 82472062,62102103, 82473298,82202148,82502479, the National Key Research and Development Program of China 2023YFC2508603, the Natural Science Foundation of Guangdong Province of China 2024A1515011672, Regional Innovation and Development Joint Fund of National Natural Science Foundation of China U22A20345, Guangdong Provincial Key Laboratory of Artificial Intelligence in Medical Image Analysis and Application (2022B1212010011).