Work

Prognostic prediction of pulmonary neuroendocrine carcinoma using a transformer-based artificial intelligence model: a two-center retrospective cohort study

Year: 2026

Type: article

Abstract: Background Pulmonary neuroendocrine carcinomas (pNECs) are a rare and heterogeneous group of malignancies, ranging from low-grade carcinoids to high-grade small-cell lung carcinoma (SCLC) and large-cell neuroendocrine carcinoma (LCNEC). Prognostic tools for these tumors are scarce and have rarely been validated in external cohorts. Objectives This study aimed to develop a Transformer-based prognostic prediction model for pNEC and validate it in an independent external cohort, to improve survival estimation and support individualized treatment decisions. Methods This retrospective two-center cohort study included 823 patients with surgically resected pNEC (2001-2023). Models were developed in Center 1 (n=374; randomly split 8:2 into training and internal test sets) and externally validated in an independent Center 2 cohort (n=449). Three algorithms—logistic regression (LR), decision tree (DT), and the Transformer-based tabular foundation model TabPFN—were compared for estimating overall survival (OS) and disease-free survival (DFS) at 1-, 3-, and 5-year landmarks, with evaluation by receiver operating characteristic, calibration, and decision curve analyses, and SHapley Additive exPlanations (SHAP) analysis for interpretability. Results TabPFN achieved the highest area under the curve (AUC) in most tasks in both test sets, with task-dependent exceptions: in the internal test set, logistic regression performed best for 1-year OS (AUC 0.885), while TabPFN performed best for the remaining tasks, including 3-year OS (0.880), overall DFS (0.810), and 1-year DFS (0.892); in the external test set, TabPFN remained best for 1- and 3-year OS and for overall, 1-, and 3-year DFS (0.747-0.802), whereas logistic regression performed best for 5-year OS and DFS. External calibration was less satisfactory. Disease stage, pathological subtype, and age were the key prognostic predictors, and TabPFN was more accurate for low-grade subtypes. SHAP analysis quantified the contribution of each feature and made the model interpretable. Conclusions The Transformer-based model improves survival estimation in pNEC, with task-dependent advantages over conventional methods and confirmed transportability to an independent institutional cohort. Prospective multicenter validation and calibration refinement are needed before broader clinical use.

Source: Frontiers in Oncology

DOI: 10.3389/fonc.2026.1948262