FAU information systems expert receives scholarship from the Daimler and Benz Foundation
How long are patients expected to stay in hospital, how often do they need to be transferred, and what treatments are required? Answering these questions as reliably as possible is an important aspect of organizing work in the healthcare sector. Dr. Sven Weinzierl, an information systems expert at Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU), is developing an AI-supported model that aims to accurately make such predictions while ensuring full transparency in the decision-making processes. The postdoctoral researcher at the Chair of Digital Industrial Service Systems has been awarded 40,000 euros in funding for his research as part of a scholarship from the Daimler and Benz Foundation.
Age, gender, lifestyle, pre-existing medical conditions, but also current vital signs such as heart rate, blood pressure, or body temperature – all this data provides information about the likely course of treatment for patients. Hospitals and other healthcare facilities are increasingly using this information to organize their workflows and plan their resources. “Models that process this data must be both accurate and self-explanatory so that clinical staff can trust them,” says Sven Weinzierl.
Transparency vs. accuracy
Until now, such modeling has been carried out primarily using statistical methods such as regression analysis. These so-called white box models have proven effective because they are transparent and easy to interpret. Their disadvantage: They only capture linear effects, i.e., proportional relationships between input and target variables – for example, between age and the probability of being transferred to the intensive care unit. “However, patient pathways are very complex and determined by a multitude of different and temporary data,” Weinzierl explains. “Such white box models are therefore always somewhat inaccurate.”
Deep learning models would deliver much more accurate results – such as large language models (LLMs), which are used in ChatGPT, for example. The major drawback: LLMs are structured in such a way that it is impossible to understand how they arrive at their results, which is why they are also referred to as black box models. Weinzierl: “However, transparency is essential in an area as sensitive as healthcare – not least for reasons of legal certainty. Just think of the prioritization involved in triage.”
Middle ground between black and white
In his new research project, Sven Weinzierl aims to combine the best of both worlds by integrating white box models with elements from deep learning. The goal is to develop a tool that serves both to plan individual treatment paths and to aggregate patient data in a way that supports the work and resource planning of the entire facility.
This is to be achieved using neural generalized additive models (neural GAMs) – a flexible statistical method that uses AI algorithms to learn from data and automatically capture nonlinear relationships. “A large part of my work will involve further developing established neural GAMs because they are actually designed for static data,” explains Weinzierl. “This may be sufficient for processing structured data, that is, mainly what is contained in patient records. However, integrating sequential data, such as current vital signs, requires adapting the models.”
After successful adaptation, the neural GAMs are intensively trained and tested using clinical datasets that have been released for research purposes. Weinzierl will also invest considerable effort in the interface between humans and models: “For practical use, the complexity of the data must be reduced to the essentials, while at the same time the tool must, of course, provide reliable information at all times,” he says. Whether this will result in visualization in the form of a dashboard has not yet been determined. In any case, however, the tool should help make the management of patient flows in healthcare facilities easier, more precise, and faster.
With his project “Neural generalized additive models for interpretable predictions in patient pathways,” Sven Weinzierl prevailed among 332 applicants and received one of 12 scholarships that the Daimler and Benz Foundation awards each year to selected postdoctoral researchers and assistant professors. The funding amount per scholarship is 40,000 euros for a period of two years and can be used flexibly to finance research assistants, technical equipment, research trips, or participation in conferences. The aim of the funding is to strengthen the autonomy and creativity of the next generation of researchers and to pave the way for their careers during the productive phase after completing their doctoral degrees. The funding program is open to applicants from all academic disciplines.
Further information:
Dr. Sven Weinzierl
Chair of Digital Industrial Service Systems
sven.weinzierl@fau.de