Multiple Machine Learning Algorithms for Predicting Postoperative Residual Tumor in Pituitary Neuroendocrine Tumors

To develop and validate an interpretable machine learning (ML) model for predicting radiologically confirmed postoperative residual tumor (RT) in pituitary neuroendocrine tumors (PitNETs).

We retrospectively analyzed 584 surgically treated patients with PitNETs from a single center, using a random 7:3 training-test split and internal cross-validation. The primary endpoint was radiologically confirmed postoperative RT, assessed using MRI obtained within 48 h after surgery and longitudinal follow-up MRI. Candidate variables were selected using LASSO regression, with nine ML algorithms developed. Discrimination, calibration, clinical utility, interobserver agreement, and SHAP-based interpretability were assessed.

Postoperative RT occurred in 72 patients (12.3%). Sensitivity analysis showed high agreement between 48-h MRI and follow-up confirmation (96.4%; Cohen's kappa = 0.837). Interobserver agreement was high for retrosellar invasion (kappa = 0.825), internal carotid artery invasion (kappa = 0.960), and Knosp grade (quadratic weighted kappa = 0.994). Eight predictors were retained: maximum sagittal diameter, maximum coronal diameter, age, Knosp grade, internal carotid artery invasion, sphenoid sinus pneumatization, tumor consistency, and retrosellar invasion. In the internal test set, the random forest model achieved high discrimination (AUC = 0.945; accuracy = 0.938), with performance comparable to LightGBM and XGBoost in DeLong testing. LightGBM had the lowest Brier score.

Random forest was selected for the web tool after internal evaluation, with discrimination comparable to LightGBM, XGBoost, and SVM. Its generalizability to other centers and future patient cohorts remains unproven because evaluation used only a single-center random split and internal cross-validation. External and temporal validation with calibration assessment is needed before clinical implementation.