Latent class statistical learning algorithms for the indirect estimation of reference intervals

• Applicant: Dr. Tobias Hepp
• Subject Area: Epidemiology and Medical Biometry/Statistics
• Term: since 2023
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 517012999

Reference intervals representing the range of physiological test results to be expected in a healthy population play an essential part in routine clinical decision making. In order to obtain reliable and accurate estimates for these diagnostic limits, direct estimation procedures usually rely on prospective study designs with extensive prescreening to filter out all pathologic samples. Unfortunately, this approach involves considerable effort and is not always possible, e.g. in the context of pediatric reference intervals due to strict recruitment regulations. As a consequence, "indirect" estimation strategies that focus on extracting the distribution of “healthy” samples from laboratory databases that also include unlabeled pathologic cases are becoming increasingly important. Currently, however, even advanced indirect methods are severely limited in terms of simultaneously adjusting the extracted reference intervals for important patient characteristics such as age. In order to overcome these limitations, different approaches for the implementation and optimization of latent class distributional regression / conditional finite mixture models were investigated by the applicant in preliminary work. The aim of this project is to enable the reliable identification of latent distributions and the simultaneous estimation of the dependence structure of their parameters on one or more covariates in the presence of unobserved heterogeneity by further developing and extending this model framework. This requires, among other things, modification and/or new development of the algorithms used to optimize the model parameters as well as different regularization strategies to address the challenges arising from the high flexibility of the model.

• DFG Programme: Research Grants
• Co-Investigator: Privatdozent Dr. Jakob Zierk