Estimation of Reference Intervals using Real-World Data

Background and Objectives Reference intervals play an essential role in the interpretation of laboratory test results. They typically represent the central 95% of values observed in a healthy reference population, providing the necessary context for interpreting numerical test results. Usually, apparently healthy individuals are specifically recruited within a study to determine the reference interval for an analyte of interest. These direct methods, however, face logistical, practical, and ethical challenges due to the specific recruitment, sample collection, extensive definition of the reference population, and ethical implications of blood draw in healthy individuals, especially children. This can lead to reference intervals of suboptimal quality as they may not be representative for the population of interest or may be derived from a small sample set, thus limiting clinical decision-making based on laboratory test results. Alternative approaches, so-called indirect methods, leverage routine measurements containing both physiological and pathological values and employ statistical algorithms to estimate reference intervals. Several indirect methods have already been developed. However, they are limited in terms of usability, applicability, robustness, and accuracy, which hinders their application in clinical practice. Therefore, the first major aim of this thesis is to develop an improved algorithm that outperforms existing methods. Additionally, an objective comparison and performance evaluation of these methods is lacking, which leads to uncertainty about their applicability in different use cases. Thus, the second aim of this thesis is to develop a benchmarking suite that enables such an objective comparison and evaluation. As most analytes vary with different covariates, like age, these changes should also be reflected in the corresponding reference intervals. Using direct methods to depict this age dependency exacerbates the aforementioned problems. Indirect methods would be the ideal approach to provide a solution for these issues. However, so far, no method exists that is free of subjective user input or manual tuning. Therefore, the third aim of this thesis is to develop an automated algorithm pipeline enabling the estimation of continuous, age-specific reference intervals. Methods As part of this thesis, we have developed a novel indirect method called refineR that uses routine measurements to estimate reference intervals. The algorithm utilizes a multi-level grid search with maximum likelihood optimization to find the model that best explains the observed data. In order to evaluate the performance of refineR and the other existing indirect methods, a benchmarking suite called RIbench was developed. It contains simulated datasets mimicking non-pathological distributions of ten biomarkers with various pathological distributions added to the data. The modular benchmarking suite enables the evaluation of existing but also future indirect methods and provides a scoring metric that can help enhance individual methods. Further, we have developed the first automated algorithm pipeline to estimate continuous, age-specific reference interval curves and percentile charts. The pipeline uses generalized additive models for location, scale, and shape in combination with discrete model estimates provided by the developed indirect method refineR and thus requires no subjective user input. Results The developed algorithm, refineR, solves many of the limitations of existing indirect methods and was shown to achieve results comparable to direct methods. Enabled by the development of RIbench, an extensive evaluation covering various test cases showed that refineR outperformed existing indirect methods for the majority of use cases. RIbench additionally provided answers to important questions regarding the application of indirect methods. For example, it offers guidance regarding the minimum sample size to achieve robust results, as well as the maximum tolerable proportion of pathological samples. The creation of the first automated algorithm pipeline for the estimation of continuous reference intervals using indirect methods proved that it is feasible to incorporate age dependency directly into the estimation. Beyond that, the pipeline was shown to achieve results similar to established and validated reference intervals derived from extensive pediatric studies. Conclusion and Discussion Altogether, the work carried out within this doctoral thesis has made significant contributions to the field of indirect reference interval methods and laboratory medicine. The developed refineR algorithm is already widely adopted and used in this field, fueled by its usability and superior performance. Proving the feasibility of incorporating the dependency on a continuous covariate like age in an automated way into the estimation of continuous reference intervals has laid the basis for future research in this field and will especially improve pediatric reference intervals. In conclusion, this doctoral thesis has contributed substantially to the enhancement of the establishment of reference intervals using real-world data. The herewith developed methods are an essential complement to existing methods, and thereby contribute to more precise reference intervals and overall to an improved interpretation of laboratory test results.