Screening patients with sensorineural hearing loss for vestibular schwannoma using a Bayesian classifier

Area of Science:

• Computational Neuroscience
• Medical Diagnostics
• Machine Learning in Healthcare

Background:

• Selecting patients with asymmetrical sensorineural hearing loss for investigation presents clinical challenges due to symptomatic patient numbers versus low vestibular schwannoma incidence.
• Existing clinical and audiological decision-support screening protocols have limitations in balancing sensitivity and specificity for vestibular schwannoma detection.

Purpose of the Study:

• To develop and validate a diagnostic model using a neural network generalization for detecting vestibular schwannomas from clinical and audiological data.
• To compare the performance of the developed model against six previously published screening protocols.

Main Methods:

• Utilized probabilistic complex data classification via a neural network generalization (Gaussian Process Ordinal Regression Classifier).
• Trained and cross-validated the classifier on clinical and audiometric data from 129 patients with vestibular schwannoma and an equal number of matched controls.
• Assessed diagnostic performance using receiver operator characteristic plots.

Main Results:

• The Gaussian Process Ordinal Regression Classifier achieved an area under the curve of 0.8025.
• At 95% sensitivity, the model demonstrated a specificity of 56%, which is 30% higher than the closest performing audiological protocols.
• Previously published audiological protocols showed sensitivities ranging from 82-97% with specificities from 15-61%.

Conclusions:

• The Gaussian Process Ordinal Regression Classifier enhances the flexibility and specificity of vestibular schwannoma screening.
• Prospective application could potentially reduce 'normal' magnetic resonance (MR) scans by up to 30% without reducing detection sensitivity.
• Further improvement is possible with additional data domains, and findings require validation on larger datasets.