Predictive Modeling of Nonfunctioning Giant Pituitary Neuroendocrine Tumor Resection: A Multi-Planar Perspective

Abstract

Background

Giant pituitary neuroendocrine tumors (GPitNETs) are challenging tumors with low rates of gross total resection (GTR) and high morbidity. Previously reported machine learning (ML) models for prediction of pituitary neuroendocrine tumor extent of resection (EOR) using preoperative imaging included a heterogenous dataset of functional and nonfunctional pituitary neuroendocrine tumors of various sizes leading to variability in results.

Methods

A retrospective study of 100 large nonfunctioning GPitNETs (≥3 cm diameter, >10 cm³ volume) was conducted to develop predictive models for GTR or EOR based on 5 variables: tumor diameter, shape, revised Knosp grade, and modified Hardy classifications for sellar and extrasellar invasion. Model performance was assessed using receiver operating characteristic-area under the curve (AUC) and confusion matrix metrics.

Results

The median preoperative tumor volume was 17.35 cm³ (interquartile range: 12.4–27.0). The median EOR was 97.6% (interquartile range: 84.9–100), and GTR was achieved in 49% of patients. The most predictive variables were the modified Hardy classification for extrasellar extension and Knosp grade (AUC of 0.771 and 0.713, respectively). Among the constructed ML models, the extreme gradient boost algorithm had the highest predictive capability, with an internal validation AUC of 0.86, while the external validation sensitivity, specificity, positive, and negative predictive values were 84%, 77%, 78%, and 82%, respectively.

Conclusions

Utilizing preoperative imaging parameters in a 3-dimensional manner proves highly valuable in predicting the EOR for nonfunctioning GPitNETs. These predictions can be easily calculated using an online open-access application: http://emoryskullbase.shinyapps.io/giant_pituitary_adenoma_resection/.