Time series are ubiquitous in the health context, ranging from ECG over EEG to data from wearable devices. We are working on particular methodological challenges in this domain such as modeling long-term dependencies, processing long sequences (such as Holter ECGs), handling missing data imputation as well as the modeling of (multimodal) patient trajectories.
Keywords: time series, long-range interactions, information extraction from long time series, missing data imputation, patient trajectories
Label scarcity is a central challenge for machine learning in general and for medical application in particular. We explore ways to alleviate the dependence on labeled training data such as self-supervised learning, incorporating symmetries/side-information and learning with weak labels.
Keywords: self-supervised learning, incorporation of symmetries/side-information, weak labels
Modern deep learning models are often perceived as opaque black boxes, which is a central barrier for bringing such tools into clinical applications. We explore interpretability methods, with particular focus on feature interactions and concept-based explanations. We try to measure and improve the robustness as well as other quality criteria for machine learning systems to increase the trustworthiness of machine learning systems.
Keywords: Explainable AI (XAI), interpretability, quality criteria, robustness, AI auditing
We aim to transfer state-of-the-art machine learning methods into the (bio)medical application space. Our main focus lies on physiological time series, in particular ECG data, but extends for example to ICU time series and beyond. In the biomedical domain, we are interested in inferring protein properties from its primary sequence. We are very open to other application domains within the broad domain of health data -- please reach out if your are interested in a collaboration.
Keywords: physiological time series, electrocardiography (ECG), intensive care unit (ICU), protein function