Using Machine Learning to Assist Clinicians to Diagnose Auditory Processing Disorder in Children
Abstract
The ability to perceive and understand auditory information is critical for child development. Some children struggle to process sound signals and associate meaning with them, despite having normal hearing sensitivity. These children are suspected of having Auditory Processing Disorder (APD). Unfortunately, children are often misdiagnosed or identified late, after they have already experienced academic failure. Reasons may include complex and subjective assessments; associated comorbidities; testing time; and a lack of clinical training. Machine Learning (ML) algorithms have been successfully employed in other areas to overcome such challenges. This thesis explores the use of ML algorithms to analyze pediatric APD assessment data.
First, a systematic literature review was performed of studies using ML to analyze auditory brainstem responses (ABRs). Next, data collected from 134 school-age children suspected of having APD by the National Centre for Audiology at Western University were used to develop two ML models. The first model aimed to identify abnormal ABRs recorded from children suspected of APD, and the second aimed to categorize data into clinically-relevant subgroups. Both models were found to make predictions that agreed with expert clinical knowledge at an accuracy level of 92% and 90%, respectively.
This work has both clinical and research implications. Categorizing children into subgroups may allow clinicians to intervene selectively. Additionally, the ML workflows derived here can be adopted into future studies. Lastly, this study may provide a basis for multi-centre collaborations to better understand APD and develop a diagnostic gold standard.