Ackground: There is a need to operationalize existing clinical data to support precision medicine in progressive hearing loss (HL). By utilizing enlarged vestibular aqueduct (EVA) and its associated inner ear abnormalities as an exemplar, we model data from a large international cohort, confirm prognostic factors for HL, and explore the potential to generate a prediction model to optimize current management paradigms.
Methods: An international retrospective cohort study. Regression analyses were utilized to model frequency-specific HL and identify prognostic factors for baseline average HL severity and progression. Elastic-net regression and machine learning (ML) techniques were utilized to predict future average HL progression based upon routinely measurable clinical, genetic, and radiological data.
Results: Higher frequencies of hearing were lost more severely. Prognostic factors for HL were the presence of incomplete partition type 2 (coefficient 12.95 dB, P=.011, 95% CI 3.0-22 dB) and presence of sac signal heterogeneity (P=.009, 95% CI 0.062-0.429) on magnetic resonance imaging. Elastic-net regression outperformed the ML algorithms (R2 0.32, mean absolute error 11.05 dB) with coefficients for baseline average hearing level and the presence of sac heterogeneity contributing the most to prediction outcomes.
Conclusion: Incomplete partition type 2 and endolymphatic sac signal heterogeneity phenotypes should be monitored closely for hearing deterioration and need for early audiological rehabilitation/cochlear implant. Preliminary prediction models have been generated using routinely collected health data in EVA. This study showcases how international collaborative research can use exemplar techniques to improve precision medicine in relatively rare disease entities.