Thema der Dissertation:
Data-driven computational approaches for the prediction of chemo-protein based food allergenicity and adverse effects Thema der Disputation:
Biased and Imbalanced Data in Biomedical Machine Learning: Problems and Solutions
Data-driven computational approaches for the prediction of chemo-protein based food allergenicity and adverse effects Thema der Disputation:
Biased and Imbalanced Data in Biomedical Machine Learning: Problems and Solutions
Abstract: Machine learning (ML) models have become central tools in computational biomedicine, chemoinformatic, and systems biology research, enabling the analysis of increasingly complex biological and chemical data. These data-driven methods learn patterns and relationships from real-world data and may therefore reproduce biases and distortions present in the training data. This challenge is particularly relevant in biomedicine, where data is complex, heterogeneous, and influenced by diverse biological and technical factors. Bias can be introduced through human factors in laboratory and clinical measurements, differences in instruments and reagents, and the underrepresentation of patient populations. Class imbalances are common in biomedical datasets due to factors such as limited patient recruitment, unequal disease prevalence, and bias in data collection. As a result, machine learning models may become biased toward the majority class, even though the minority class is often the primary class of clinical interest.
This talk examines these challenges in the context of biomedical ML and discusses possible approaches to address them. Examples include open data and code sharing to improve reproducibility and facilitate more diverse datasets, sampling techniques to mitigate class imbalances, and appropriate performance measures for robust and reliable model evaluation.
This talk examines these challenges in the context of biomedical ML and discusses possible approaches to address them. Examples include open data and code sharing to improve reproducibility and facilitate more diverse datasets, sampling techniques to mitigate class imbalances, and appropriate performance measures for robust and reliable model evaluation.
Zeit & Ort
29.09.2026 | 10:00
Seminarraum 053
(Fachbereich Mathematik und Informatik, Takustr. 9, 14195 Berlin)