Transformer AI Model Predicts Survival in Rare Lung Neuroendocrine Cancers

Pulmonary neuroendocrine carcinomas (pNECs) are a rare and heterogeneous group of lung malignancies, spanning a spectrum from slow-growing, low-grade carcinoids to the highly aggressive small-cell lung carcinoma and large-cell neuroendocrine carcinoma. Prognostic tools for these tumors are scarce, and the models that do exist have rarely been validated in external cohorts. A new retrospective two-center cohort study from China, published in Frontiers in Oncology, reports that a Transformer-based artificial intelligence model can outperform conventional statistical methods in predicting survival for these patients in most tasks—though with important caveats about where it excels and where it does not.

The research team, led by investigators at Peking Union Medical College Hospital in collaboration with the medical technology company Infervision, assembled records from 823 patients who underwent surgical resection of a pNEC between 2001 and 2023. The cohort was split into two parts: 374 patients from the first center, randomly divided in an 8:2 ratio into training and internal test sets, and an independent external validation cohort of 449 patients from a second center. This external validation is the study’s methodological backbone, because it tests whether a model built at one institution can transport to patients from a different hospital and population.

The team compared three algorithms head-to-head: logistic regression, a decision tree, and TabPFN, a Transformer-based tabular foundation model that adapts the attention architecture behind modern language models to structured clinical data. Predictions targeted overall survival and disease-free survival at one-, three-, and five-year landmarks, and the models were judged by three complementary yardsticks—receiver operating characteristic analysis for discrimination, calibration analysis for whether predicted probabilities match observed outcomes, and decision curve analysis for clinical usefulness. To make the model interpretable, the researchers applied SHAP (SHapley Additive exPlanations) analysis, which quantifies how much each input feature pushes an individual prediction up or down.

The results were nuanced rather than triumphant. TabPFN achieved the highest area under the curve (AUC) in most tasks in both test sets, but logistic regression won in several. In the internal test set, logistic regression performed best for one-year overall survival with an AUC of 0.885, while TabPFN led the remaining tasks, including three-year overall survival at 0.880, overall disease-free survival at 0.810, and one-year disease-free survival at 0.892. In the external cohort, TabPFN remained the top performer for one- and three-year overall survival and for overall, one-, and three-year disease-free survival, with AUCs ranging from 0.747 to 0.802—yet logistic regression again performed best for five-year overall and disease-free survival. In other words, the Transformer’s advantage was real but task-dependent, and less consistent at the five-year mark.

Calibration told a humbler story. When the models were evaluated in the external center, their calibration was less satisfactory, meaning the raw probabilities they emitted did not always align with what actually happened to patients there. The authors are explicit that prospective multicenter validation and calibration refinement are needed before the model can be used more broadly in clinical practice.

The interpretability analysis offered clinically recognizable results. SHAP analysis identified disease stage, pathological subtype, and age as the key prognostic predictors. Notably, the Transformer-based model was more accurate for low-grade subtypes, the indolent end of the pNEC spectrum. Because SHAP assigns each feature a quantified contribution for every patient, the model’s reasoning can be inspected rather than taken on faith, which the authors present as support for individualized treatment decisions.

For rare diseases like pNEC, where no single center accumulates large numbers of cases, the two-center design is a meaningful step beyond single-institution proof-of-concept studies. Demonstrating that a Transformer-based model retains discrimination in an independent institutional cohort addresses one of the central hurdles facing prognostic models built on small, rare-disease datasets.

Still, the caveats deserve emphasis. This was a retrospective cohort of surgically resected patients, so the findings apply to that population rather than to all pNEC presentations. The five-year predictions, where logistic regression performed best, remain a task where the Transformer did not lead. And the external calibration shortfalls mean the model currently ranks risk better than it sizes it. As a demonstration that Transformer architectures can generalize across institutions for a rare thoracic malignancy, the study is a genuine advance; as a clinical tool, it remains a promising prototype awaiting prospective trials.

Subject of Research: Transformer-based artificial intelligence for prognostic prediction of pulmonary neuroendocrine carcinoma in a two-center retrospective cohort study

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

Article References: Liu, L., Liang, L., Wang, D., Zhang, J., Zhao, K., Guo, C., Huang, C., Li, S., Gao, Y., & Chen, Y. (2026). Prognostic prediction of pulmonary neuroendocrine carcinoma using a transformer-based artificial intelligence model: a two-center retrospective cohort study. Frontiers in Oncology, 16, Article 1948262. https://doi.org/10.3389/fonc.2026.1948262

Image Credits: AI Generated

DOI: 10.3389/fonc.2026.1948262

Keywords: pulmonary neuroendocrine carcinoma, Transformer, TabPFN, artificial intelligence, prognosis prediction, overall survival, disease-free survival, external validation, SHAP, retrospective cohort study, thoracic oncology, machine learning