Assessing Clustering Replicability in Sensorineural Hearing Loss: Insights from the UK's Largest Audiogram Cohort

Abstract

Despite nearly 20% of the global population experiencing hearing loss, there remains limited insight into the underlying subtypes of its most prevalent cause, sensorineural hearing loss (SNHL). This understanding is crucial for effective therapeutic and preventative strategies. A recent study using a Gaussian Mixture Model (GMM) identified 10 distinct SNHL phenotypes in a large US cohort, highlighting the potential of unsupervised machine learning to provide a data-driven solution to this task. Rigorous validation of these models is essential; however, it is limited due to several factors, including the absence of ground truth labels for model evaluation, restricted data access, and the lack of a standardized reporting framework for comparing results across clustering studies. Here, we apply GMM to a UK database of 109,854 audiograms, revealing 9 phenotypes, partly overlapping prior findings. Notably, our study cohort is characterized by advanced age, a higher proportion of female participants, and more severe hearing impairments. We observed instability in the GMM model when subjected to variations in the original dataset. To enhance practices, we propose a Clustering Replicability Framework, ensuring robustness in unsupervised machine learning-driven health research for safe clinical translation. Words: 187