Authors: Gerard Encina-Llamas¹
¹University of Vic – Central University of Catalonia
Background: Hearing impairment is multifactorial, yet pure-tone audiometry remains the clinical gold standard despite its limited sensitivity to certain pathologies, including hidden hearing damage. Prior phenotyping approaches either relied solely on audiogram data or incorporated additional tests but were constrained by small sample sizes, leaving a substantial portion of the patient space unclassified.
Method: We analysed a large clinical dataset from Rigshospitalet (Denmark) spanning 1995–2022, comprising ~300,000 audiograms from over ~85,000 unique adult patients, including audiometric thresholds and speech audiometry results. We compared Principal Component Analysis (PCA) and Uniform Manifold Approximation and Projection (UMAP) for two-dimensional visualisation. Clustering algorithms — Gaussian Mixture Models, and Leiden community detection — were applied to a sensorineural hearing loss (SNHL) subset. Longitudinal trajectory analysis was also performed to characterise phenotypic progression over time.
Results: PCA captured major audiogram variance but proved limited for multimodal audiological data. UMAP better exploited both global and local data structures and was adopted as the basis for our Hearing Loss Map. Comparison with existing phenotype classifications revealed that up to 40% of data remained unclassified. Clustering analyses identified novel SNHL phenotypes not captured by conventional schemes, and trajectory analysis enabled longitudinal characterisation of hearing loss progression.
Conclusion: We propose this UMAP-based Hearing Loss Map as a navigational tool for the SNHL patient space, with potential applications in clinical decision-making, patient counselling, evaluation of novel treatments, and development of next-generation hearing technologies.
