Machine Learning to Predict Hearing Preservation After Middle Cranial Fossa Approach for Sporadic Vestibular Schwannomas

Abstract

Objective: Predict hearing preservation after middle cranial fossa approach for vestibular schwannomas.

Study design: Application of machine learning algorithms, including classification and regression trees and random forest models to observational data.

Setting: Single-tertiary referral center.

Participants: Patients (n = 144) with a previously untreated sporadic vestibular schwannoma who underwent microsurgical resection by middle cranial fossa approach between November 2017 and November 2021.

Interventions: Middle cranial fossa approach.

Main outcome and measures: Hearing preservation, defined by postoperative word recognition score of 50% or greater and pure tone average below 50 dB HL or less than 10% reduction in word recognition score. Model performance was evaluated with classification accuracy in an independent validation sample. Variable importance for the random forest model is reported according to entropy, a measure of mean decrease in model accuracy incurred by excluding each variable from the model.

Results: Hearing preservation was achieved in 60% of patients (86 of 144) overall. The classification and regression tree model identified preoperative pure tone average with a cut point of 30 dB HL, and more posterior tumor position to be the most important prognostic features for hearing preservation. Model accuracy was 0.68. The random forest model demonstrated perfect accuracy (1). Baseline pure tone average, word recognition score, and anteroposterior tumor position were among the most influential features for hearing preservation prediction.

Conclusion: Machine learning algorithms have the potential for accurate prediction of hearing preservation rates after middle fossa approach for vestibular schwannomas at a single institution. These models have the capacity for continued refinement with ongoing addition of data.