Machine learning-based models for preoperative prediction of pituitary adenoma consistency: a systematic review and meta-analysis

TL;DR

This paper reviews machine learning models for predicting pituitary adenoma consistency before surgery, showing high accuracy and potential for improving surgical planning.

Contribution

A meta-analysis of ML models for predicting pituitary adenoma consistency, providing pooled diagnostic performance metrics.

Findings

Pooled AUC of 0.92 indicates excellent predictive accuracy of ML models for PA consistency.

Algorithms like Random Forest and SVM showed strong performance with high sensitivity and specificity.

Leave-one-out analyses confirmed model robustness and low publication bias.

Abstract

The consistency of pituitary adenoma (PA) significantly impacts surgical difficulty and the extent of resection. Machine learning (ML) and radiomics have emerged as quantitative tools to predict tumor firmness from MRI-derived features. This systematic review and meta-analysis aimed to synthesize the diagnostic performance of ML-based models for preoperative prediction of PA consistency.

PubMed, Embase, Scopus, and Web of Science were searched through September 2025. Studies developing or validating ML or deep learning (DL) models for predicting PA consistency were included. Pooled estimates of area under the curve (AUC), accuracy (ACC), sensitivity (SEN), specificity (SPE), and diagnostic odds ratio (DOR) were calculated with 95% confidence intervals (CIs).

Nine studies with 1,621 patients were analyzed. Algorithms included Extra Trees (ET), Random Forest (RF), Support Vector Machine (SVM), k-Nearest Neighbors (kNN), Logistic Regression (LR), Artificial Neural Network (ANN), and hybrid DL architectures. The pooled AUC was 0.92 (95% CI: 0.86–0.98), ACC 0.86 (95% CI: 0.79–0.92), SEN 0.80 (95% CI: 0.71–0.87), SPE 0.85 (95% CI: 0.80–0.89), and DOR 19.27 (95% CI: 10.27–36.17). Leave-one-out analyses confirmed robustness, and Egger’s tests indicated no significant publication bias.

ML-based models demonstrate excellent pooled diagnostic accuracy in predicting PA consistency preoperatively, underscoring their value for individualized surgical planning. Future multicenter studies with standardized imaging and external validation are needed to optimize clinical translation.

The online version contains supplementary material available at 10.1007/s00701-026-06775-w.

Figures