Detecting hearing loss in high-risk neonates using machine learning

Abstract

Deafness is one of the most commonly occurring birth conditions in children worldwide creating an increasingly significant global health concern. Failure to early identify hearing loss and provide subsequent intervention services will likely have negative consequences on language, cognition, and socio-emotional development. Current approaches in detecting neonatal hearing loss are limited specifically in developing countries such as South Africa. Machine learning offers an opportunity to create models which could predict the likelihood of a hearing loss occurring in high-risk neonates allowing for early identification and intervention to occur. Thus, the main aim of the current study was to use predictive modelling to predict the likelihood of hearing loss in high-risk neonates. The study sample comprised of 12 044 male and female hearing and deaf and/or hard-of-hearing South African children who either formed part of the HI HOPES or universal newborn screening programme implemented at the Netcare Hospital Group. A nonexperimental, predictive modelling design was employed for the purpose of the current study. Predictive variables used in the current study included mode of delivery, prematurity, gestational age, family history of hearing loss, extracorporeal membrane oxygenation (ECMO), in-utero infections, craniofacial anomalies, physical findings, syndromes associated with hearing loss, neurodegenerative disorders, cultural-positive infections, meningitis, maternal and/or infant HIV infection, and ototoxic medication. The results from several Chi-Square (X 2 ) analyses showed significant correlations between each birth type (i.e., natural, elective caesarean, emergency caesarean), prematurity, family history, ECMO, in-utero infection, craniofacial anomalies, physical findings, syndromes associated with hearing loss, cultural-positive postnatal infections, meningitis, maternal and/or HIV infection, and ototoxic medication. The predictive models for hearing loss in high-risk neonates were developed using logistic regression and random forest (RF) classifiers. The major predictors of neonatal hearing loss determined by both models were prematurity, family iv history, cultural-positive infections, and meningitis. The final reduction of error rate for the logistic regression was 90% with a prediction rate of 92%. In contrast, the random forest performed slightly poorer with an out-of-bag error rate of 14.8% and a prediction rate of 88%. The results of the current study demonstrated that machine learning algorithms can be used as potential tools for the evaluation and prediction of hearing loss in high-risk neonates.