Sven Herrmann1,2, Mathias Dietz1,2
1Department für Medizinische Physik und Akustik, Universität Oldenburg, Oldenburg, Germany; 2Cluster of Excellence “Hearing4all”, Universität Oldenburg, Oldenburg, Germany
Background: Given the complexity of most brain and body processes, it is often not possible to relate experimental data from an individual to the underlying subject-specific physiology or pathology. Computer simulations of these processes simulate experimental data but rarely address the inverse problem, i.e. to identify a pathology from experimental data. Even if the inverse problem is addressed, typically a single model parameter is varied to fit previously recorded experimental data. After the fit, confidence intervals are given in the units of the experimental data but usually not for the model parameters that are the ultimate interest of the diagnosis. Our long-term goal is that models allow for a quantitative diagnosis and guide the diagnostic process.
Methods: We propose a likelihood-based fitting procedure, operating in the model parameter space and providing confidence intervals for the parameters under diagnosis.
Results: The procedure is capable to run in parallel with the measurement and can adaptively set stimulus parameters to the values that are expected to provide the most diagnostic information. By predefining the acceptable diagnostic tolerance, i.e. the confidence intervals, the experiment continues until the goal is reached. As an example, the procedure is tested with a simplistic three-parameter auditory model and a psychoacoustic binaural tone in noise detection experiment. The model-based measurement steering provided 80% more information for a given number of trials, compared to a conventional maximum likelihood measurement.
Conclusion: We conclude that model-based experiment steering is possible and has at least theoretical advantages over sequential measure-and-fit approaches. Practical problems and the lack of sufficiently accurate models are going to prohibit most diagnostic applications for the time being.
The goal is to find the model instance (folder from shelf y) that has the maximum likelihood to have generated the experimental data set d. The stimulus that leads to the smallest variance of parameter estimates is presented next. The process is repeated until the termination criterion is met.