Abstract
Speech-based automatic detection of Alzheimer's disease (AD) and depressionhas attracted increased attention. Confidence estimation is crucial for atrust-worthy automatic diagnostic system which informs the clinician about theconfidence of model predictions and helps reduce the risk of misdiagnosis. Thispaper investigates confidence estimation for automatic detection of AD anddepression based on clinical interviews. A novel Bayesian approach is proposedwhich uses a dynamic Dirichlet prior distribution to model the second-orderprobability of the predictive distribution. Experimental results on thepublicly available ADReSS and DAIC-WOZ datasets demonstrate that the proposedmethod outperforms a range of baselines for both classification accuracy andconfidence estimation.