Authors: Hector Gabriel Corrale de Matos¹, Jan-Willem Wasmann², Kátia de Freitas Alvarenga¹, Lilian Cássia Bórnia Jacob¹
¹Department of Speech Language Pathology and Audiology of the University São Paulo Campus Bauru; ²Radboud University Medical Centre
Background: Computational reproducibility relies on open data access. Ethical and legal restrictions on patient privacy often hinder the sharing of raw audiometric data. Differential privacy offers a mathematical framework to share aggregate statistical patterns while guaranteeing individual anonymity. Its use in clinical audiology data utility remains underexplored.
Method: We generated a synthetic dataset (N=100) simulating a clinical population with two distinct clusters: young adults with normal hearing and older adults with hearing loss. The Laplace mechanism was applied to mask simulated variables (age and air thresholds from 250 to 8000 Hz) across a range of privacy budgets (ϵ) from 0.1 to 1. Utility loss was compared against the synthetic baseline sample through analysis of data distribution, group differences, effect sizes, statistical power, and Pearson correlation (r) between age and the quadritonal mean.
Results: Laplacian noise significantly altered the data distribution to a non-linear trajectory, with parametric assumption violations likely when ϵ < 0.5. These noise levels resulted in negligible effect sizes and reduced age-hearing correlations. Statistical utility was equilibrated at ϵ = 0.8, where correlations (r > 0.72) and effect sizes (d > 0.8) were maintained, alongside high post hoc statistical power (>90%) and significant group differences similar to baseline.
Conclusion: Differential privacy is a viable model for audiometric reproducibility. Selecting a balanced privacy budget allows the sharing of reproducible datasets that preserve clinical validity without compromising patient privacy.
