Uncertainty-aware modeling of hemolysis in rotary blood pumps using reduced-order models and Bayesian parameter estimation Parameterschätzung

In extracorporeal membrane oxygenation (ECMO), a life-support therapy used in patients with severe cardiac or respiratory failure, blood is circulated outside the body through an artificial circuit that includes a membrane oxygenator and a rotary blood pump (RBPs). While ECMO can be life-saving, it exposes blood to artificial surfaces and elevated mechanical stresses, particularly in RBPs, which operate under supraphysiological shear conditions and are a primary source of red blood cell (RBC) damage. A key complication is shear-induced hemolysis, the mechanical destruction of RBCs in high-shear blood flows. This damage releases hemoglobin and contributes to adverse events such as thrombosis, inflammation, and organ dysfunction. To better understand and reduce hemolysis, numerical models have become an essential tool in the design and evaluation of blood-contacting devices. However, current in-silico approaches face major limitations. They are computationally expensive, often lack validation against clinical data, and typically neglect biological and experimental variability. As a result, their ability to reliably predict hemolysis across realistic clinical conditions remains limited. The overarching aim of this dissertation is to develop a computationally efficient, uncertainty-aware in-silico framework that can predict hemolysis across the full clinical operating range of RBPs and explicitly account for key sources of uncertainty. To make large-scale predictions feasible, a reduced-order model was developed based on non-intrusive polynomial chaos expansion. This approach enables rapid yet accurate simulation of flow conditions across the entire clinical operating space. Building on this, stress-based hemolysis predictions were validated to in-vivo data from 580 ECMO patients. While the results showed good agreement with cohort-level trends, they also revealed substantial inter- and intra-patient variability, pointing to the limits of purely mechanical models. To investigate the source of this variability, a controlled experimental study was conducted using a custom-built shearing device. Repeated hemolysis measurements under identical conditions showed unexpectedly high intra-donor variability, challenging assumptions underlying standard calibration practices and highlighting inherent uncertainty in experimental data. To address this, a Bayesian parameter estimation framework was introduced that incorporates this variability directly into model calibration. The resulting probabilistic model provides not only hemolysis predictions but also quantified uncertainty, increasing robustness, interpretability, and credibility in both clinical and regulatory contexts. Together, these contributions help to establish a more comprehensive framework for hemolysis modeling by combining computational efficiency, clinical validation, and uncertainty quantification. Building on this foundation, future applications such as in-silico clinical trials or patient-specific modeling across diverse patient conditions may become more feasible.