The NET score: an interpretable AI-assisted prognostic score for mortality risk in lung neuroendocrine tumors

Lung neuroendocrine tumors (NETs), including typical and atypical carcinoids, show heterogeneous outcomes. The existing nomograms are often complex and insufficiently validated, limiting their bedside use. A simple, reproducible model based on routine pathology is therefore needed. This study aimed to develop and internally validate the NET score, an interpretable, AI-assisted tool to estimate individualized mortality risk. In a retrospective cohort of resected pulmonary carcinoids, candidate predictors were screened using LASSO regression to reduce overfitting and identify key variables: nodal status, mitotic index (>2 per 2 mm2), necrosis, and Ki-67 (>5%). Selected variables were incorporated into a logistic regression model to generate a point-based score (0-8). The model estimates cumulative mortality risk across follow-up, not fixed-time survival. Internal validation included bootstrap and cross-validation to assess discrimination and calibration. The final model included LODDS > -0.5 (3 points), mitotic index >2 (2 points), necrosis (2 points), and Ki-67 (1 point). Discrimination was moderate (bootstrap-corrected AUC: 0.70). Risk groups were defined as low (0-2 points, ≤5%), intermediate (3-4 points, 8-12%), and high (≥5 points, ≥18%). Kaplan-Meier curves demonstrated progressive survival stratification across risk groups. The NET score is a practical and interpretable prognostic tool for lung NETs. It supports risk communication and clinical decision-making while maintaining transparency. External validation is required. Combining AI-based variable selection with a simple scoring system represents a pragmatic approach to prognostic modeling in rare thoracic malignancies.