Voxelwise statistical methods to localize practice variation in brain tumor surgery.

Area of Science:

• Neurosurgery
• Medical Statistics
• Radiology

Background:

• Neurosurgical decisions during brain tumor resections involve balancing residual tumor risk against functional deficits.
• Variations in surgical practices across institutions can arise from differing risk perceptions.

Purpose of the Study:

• To develop and evaluate statistical methods for localizing institutional differences in the extent of brain tumor resection.
• To identify specific brain regions affected by practice variations in surgical resections.

Main Methods:

• Generated synthetic brain tumor and resection data to simulate institutional practice variations.
• Investigated three statistical methods: permutation testing, Fisher's exact test, and Bayesian Markov chain Monte Carlo (MCMC).
• Assessed method performance using false discovery rate (FDR), receiver operating characteristic (ROC), and precision-recall curves on synthetic and retrospective patient data.

Main Results:

• Fisher's exact test accurately estimated FDR on synthetic data, while permutation testing was too liberal.
• Bayesian MCMC and Fisher's methods showed similar Area Under the Curve (AUC) values, outperforming permutation testing.
• Performance of Fisher's method degraded with smaller tumors, effect regions, lower resection extents, smaller cohorts, and less pronounced practice variations.

Conclusions:

• Voxel-based statistical methods can detect differences in surgical practices for brain tumor resections.
• Fisher's test is a rapid tool for localizing differences but may underestimate true FDR.
• Bayesian MCMC offers flexibility and extensibility, yielding comparable results to Fisher's test at a higher computational cost.