AudioGene 9.0: novel ensemble machine learning classification of 23 classes of autosomal non-syndromic hearing loss (deafness)

As the field of personalized genomic medicine expands and grows new technology is needed to observe and understand individuals, so individuals can be treated and diagnosed for their specific needs. Technologies like sequencing may be used to identify an individual’s health problems through their variants from sequencing their genome. Some diagnosis such as genetic hearing loss is a health problem where distinct genes contribute to hearing loss or deafness of an individual. However, there are other technologies to identify known distinct deafness-causing gene. In 2014, Kyle Taylor presented a machine learning model called AudioGene4.0 to predict these distinct genes through the information of individual’s audiogram and corresponding age. This software tool allowed clinician-scientists to generate a reasonable or general hypothesis of potential known deafness-causing genes to focus their scope of variants existing in subregions of known genes. However, AudioGene4.0 was plagued with challenges related to the existing audiometric data used to train AudioGene4.0. Therefore, a new approach was developed called AudioGene9.0 using multiple specialized ‘sub-models’ trained based upon partitions from characteristics related to the audiometric dataset. These results from these sub-models are then combined into one additional ensemble classifier (model) to predict the 3 most likely disease-causing genes for a given unknown audiogram. Our data showed that performance increased from 70.93% to 79.64% from AudioGene4.0 to 9.0.