Development and evaluation of a multivariable prediction model for overall survival in advanced stage pulmonary carcinoid using machine learning

Abstract

Background

Evidence is limited on whether patients with advanced pulmonary carcinoid (APC) benefit from comprehensive pulmonary resection (CPR), chemotherapy, or radiotherapy. Existing prognostic models for APC are limited and do not guide treatment selection. This study aims to develop and evaluate a multivariable machine learning model to predict overall survival in APC patients and provide a web-based prognostic tool.

Methods

Clinical data of APC patients were obtained from SEER database. Propensity score matching reduced retrospective study bias. Kaplan-Meier analysis evaluated survival differences between CPR vs. nonCPR, chemotherapy (Chem) vs. no chemotherapy (nonChem), and radiotherapy (Radio) vs. no radiotherapy (nonRadio). Univariate and multivariate Cox regression identified survival-associated variables. Using these clinical variables, 91 machine learning models were developed to predict APC survival, and the best model led to a web-based prognostic tool.

Results

Among 1077 APC patients, 37.0 % underwent CPR, 30.2 % received chemotherapy, and 19.9 % received radiotherapy. After matching, overall survival was significantly improved in the CPR compared to the nonCPR. However, there were no significant differences in survival between the Chem and nonChem groups or between the Radio and nonRadio groups. Eight out of 13 clinical variables were significant prognostic variables. Models with eight variables reached a mean C-index of 0.770 and a 5-year AUC of 0.835. Using all 13 variables, a C-index of 0.785 and an AUC of 0.850 was achieved. An online tool (https://apcmodel.shinyapps.io/APCsp/) displays survival curves for different treatments.

Conclusion

The developed prognostic model enables individualized survival predictions and supports evidence-based treatment decisions for APC patients.