Uncovering Phenotypes in Sensorineural Hearing Loss: A Systematic Review of Unsupervised Machine Learning Approaches.

Abstract

Objectives: The majority of the 1.5 billion people living with hearing loss are affected by sensorineural hearing loss (SNHL). Reliably categorizing these individuals into distinct subtypes remains a significant challenge, which is a critical step for developing tailored treatment approaches. Unsupervised machine learning, a branch of artificial intelligence (AI), offers a promising solution to this issue. However, no study has yet compared the outcomes of different AI models in this context. The purpose of this review is to synthesize the existing literature on the application of unsupervised machine learning models to hearing health data for identifying subtypes of SNHL.

Design: A systematic search was performed of the following databases: MEDLINE, PsycINFO (Ovid version), EMBASE, CINAHL, IEEE, and Scopus as well as a search of grey literature using GitHub and Base, and manual search (Jan 1990-Mar 2024). Studies were included only if they reported on adult patients with SNHL and used an unsupervised machine-learning approach. Quality assessment was performed using the APPRAISE-AI tool. The heterogeneity of studies necessitated a narrative synthesis of the results.

Results: Seven studies were included in the analysis. Apart from one case-control study, all were cohort studies. Four different algorithms were used, with no study comparing the performance of more than one algorithm. Across these studies, only 2 distinct numbers of subtypes were identified: 4 and 11. However, the overall quality of the studies was deemed low, thus preventing definitive conclusions regarding model selection and the actual number of subtypes.

Conclusions: This systematic review identifies key methodological practices that need to be improved before the potential of unsupervised machine learning models to subtype SNHL can be realized. Future research in this field should justify model selection, ensure reproducibility, use high-quality hearing data, and validate model findings.