h1

TY  - THES
AU  - Einollahzadeh Samadi, Moein
TI  - Uncertainty quantification of binary hybrid models in life sciences
PB  - Rheinisch-Westfälische Technische Hochschule Aachen
VL  - Dissertation
CY  - Aachen
M1  - RWTH-2025-06656
SP  - 1 Online-Ressource : Illustrationen
PY  - 2025
N1  - Veröffentlicht auf dem Publikationsserver der RWTH Aachen University
N1  - Dissertation, Rheinisch-Westfälische Technische Hochschule Aachen, 2025
AB  - As clinical information becomes increasingly accessible and quantitative modeling techniques evolve, the potential for AI-assisted decision-making in healthcare is more important than ever. However, ensuring the reliability of AI-based decisions remains challenging, largely due to the need for representative datasets that span diverse patient populations and conditions to mitigate algorithmic bias. This challenge is compounded by the nature of clinical data, which often takes the form of discrete or binary endpoints (e.g., throughout presence or absence of a diagnosis) and is collected selectively from individuals deemed necessary for specific tests, leading to sparse data representation. This thesis presents novel contributions to the development and validation of learning strategies for hybrid models employing binary data. The primary focus is on tree-structured hybrid models, which facilitate the transfer of domain knowledge to data-driven learning in life sciences applications. Three main contributions are: i) a learning algorithm, DICE, for binary hybrid models that addresses the curse of dimensionality in binary feature spaces, facilitates the assessment of the extrapolation range of hybrid models in binary space, and introduces a design-of-experiment strategy to identify new data with maximum information content for reducing uncertainty in extrapolation. Additionally, DICE introduces the concept of a binary database by showcasing a proof-of-concept in a biomedical application for mortality estimation of COVID-19 ICU patients. ii) A hybrid modeling framework in a multi-hospital study for interpretable and generalizable ICU mortality predictions is proposed, offering consistent explainability of mortality causes for mechanically ventilated influenza and pneumonia patients across diverse healthcare settings. iii) NoiseCut, a Python package for hybrid modeling specialized in noise-tolerant supervised learning on binary-encoded data, is introduced, offering a novel approach to mitigating overfitting in sparse datasets.
LB  - PUB:(DE-HGF)11
DO  - DOI:10.18154/RWTH-2025-06656
UR  - https://publications.rwth-aachen.de/record/1015954
ER  -