Improving accuracy in classifying neonatal hearing losses

Mr Andrew Geyl1

1Sydney Children’s Hospital Randwick, NSW, Australia

When assessing neonates using the evoked Auditory Brainstem Response (ABR) determining the nature of a hearing loss — conductive, sensorineural/retrocochlear or mixed — is of utmost importance. In most cases, clinicians need to determine only if a hearing loss is sensorineural or conductive. Tympanometry is useful, but bone conduction thresholds are generally thought to be the gold standard.
However, for the neonatal population ABR bone conduction thresholds can be difficult to obtain. Babies tend to be “non-compliant”, and the vibrator itself causes significant electrical interference. Often, the bone conduction thresholds obtained for neonatal ABRs are quite limited – both due to limited time available before the baby starts to protest, and the limited gain before interference is too great.
Further, we show that, in our experience, the relationship between neonatal air and bone conduction thresholds is not necessarily 1:1. This means that first normalising air conduction thresholds and bone conduction thresholds and then applying a correction factor will not always produce an accurate estimation of the air conduction – bone conduction gap (ABG), especially at higher intensities.
Instead, using historical data neonatal ABR data we can significantly improve both accuracy and range by using a combined approach:
• At intensities less than or equal to 45 dBHL, the normal (ie non-conductive) ABG can linearly modelled as a function of air conduction threshold.
• At intensities greater than or equal to 45 dBHL, we use logistic regression to create a hearing loss classifier model as a function of wave V latencies (sensitivity and specificity both ~ 0.95)
The combined approach not only improves a clinician’s accuracy but can also provide statistical confidence intervals for the diagnosis.


Andrew Geyl is an audiologist with more than twelve years experience in assessing neonatal ABR, with an interest in data analysis.