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

OBJECTIVES

Selecting patients with asymmetrical sensorineural hearing loss for further investigation continues to pose clinical and medicolegal challenges, given the disparity between the number of symptomatic patients, and the low incidence of vestibular schwannoma as the underlying cause. We developed and validated a diagnostic model using a generalisation of neural networks, for detecting vestibular schwannomas from clinical and audiological data, and compared its performance with six previously published clinical and audiological decision-support screening protocols.

DESIGN

Probabilistic complex data classification using a neural network generalization.

SETTINGS

Tertiary referral lateral skull base and a computational neuroscience unit.

PARTICIPANTS

Clinical and audiometric details of 129 patients with, and as many age and sex-matched patients without vestibular schwannomas, as determined with magnetic resonance imaging.

MAIN OUTCOME MEASURES

The ability to diagnose a patient as having or not having vestibular schwannoma.

RESULTS

A Gaussian Process Ordinal Regression Classifier was trained and cross-validated to classify cases as 'with' or 'without' vestibular schwannoma, and its diagnostic performance was assessed using receiver operator characteristic plots. It proved possible to pre-select sensitivity and specificity, with an area under the curve of 0.8025. At 95% sensitivity, the trained system had a specificity of 56%, 30% better than audiological protocols with closest sensitivities. The sensitivities of previously-published audiological protocols ranged between 82-97%, and their specificities ranged between 15-61%.

DISCUSSION

The Gaussian Process ORdinal Regression Classifier increased the flexibility and specificity of the screening process for vestibular schwannoma when applied to a sample of matched patients with and without this condition. If applied prospectively, it could reduce the number of 'normal' magnetic resonance (MR) scans by as much as 30% without reducing detection sensitivity. Performance can be further improed through incorporating additional data domains. Current findings need to be reproduced using a larger dataset.