Conference Agenda

The intensive care unit (ICU) is a critical and costly asset within hospital settings, necessitating effective management strategies, especially given the influx of emergency cases and the unpredictable nature of patient stays. ICU admissions arise from elective surgeries, further complicating capacity planning. Leveraging artificial intelligence (AI) and analytics as decision support tools presents a promising avenue to address these challenges. However, the practical implementation of such systems remains limited, hindered by factors including digitization gaps and skepticism surrounding AI transparency. This work presents the development and validation of a transparent scoring model utilizing AI and analytics to provide decision support for management in hospitals. Focused on the decision of post-operative ICU transfer for elective surgery patients, our model aims to aid physicians, especially those with less experience, while enhancing capacity planning efficiency through efficient and effective scheduling of elective patients. A comprehensive experiment involving physicians is conducted to evaluate the practical relevance and usability of our transparent scoring model. The results demonstrate that our clinical decision support system accurately predicts ICU requirements, thus optimizing resource utilization and enhancing patient care. In our evaluation data set, approximately 90% of patients are classified correctly. By addressing the gap between AI innovation and practical implementation, this research contributes to the advancement of AI and analytics in healthcare, offering tangible benefits for hospitals striving to navigate resource scarcity and improve patient outcomes.

Optimizing patient-to-room and nurse-to-patient assignments is crucial for efficient hospital workflows, high-quality care, and patient and staff satisfaction. Integrating both assignment problems enables the optimization of additional objectives that depend on the interaction of the two assignment problems. For example, minimizing the walking distances of nurses or assigning the minimum number of nurses to patients in the same room to mitigate negative effects, such as the spread of infections between rooms by nurses or the disturbance of patients. Existing literature tackles the static version of this integrated problem, assuming full prior knowledge of patient and nurse parameters. However, real-world hospital operations are rife with uncertainties, including patient no-shows, emergency admissions, fluctuating length of stays, and unforeseen nurse absences. Enhancing predictability and forecasting reliability necessitates accounting for stochastic variations within the planning horizon.

We have developed a decision support model that addresses the dynamic patient-to-room and nurse-to-patient assignment. The model is presented as a mixed integer optimization problem. We present an efficient heuristic to solve the assignment problem under data uncertainty. We conduct computational experiments on real-world and artificially generated instances. A comparative analysis against the static problem formulation underscores the efficacy and superiority of our dynamic extensions.

Most activities in hospitals require the presence of the patient. Delays in patient transport can therefore cause disruptions and costly downtime in many different areas and departments.This makes patient transport planning a central operational problem in hospitals that should be integrated with other problems such as operating room planning or material transport planning.

This talk presents the first literature review of Operations Research approaches for improving non-emergency patient transport in hospitals. We structure the different patient transport problems considered in the literature according to several main characteristics and introduce a four-field notation for patient transport problems that allows for a concise representation of different problem variants. We then analyze the relevant literature with respect to different aspects related to the considered problem variant, the employed modeling and solution techniques, as well as the data used and the level of practical implementation achieved. Based on our literature analysis and semi-structured interviews with hospital practitioners, we provide a comparison of current hospital practice and the existing literature on patient transport, and we identify research gaps and formulate an agenda for relevant future research in this area.

The basic idea of Federated Machine Learning is to collect models rather than data centrally. We study the feasibility and the potential of the rather young research area, which has been proposed by Google in 2016, for digital Covid-19 diagnosis based on German multicenter data of 3,670 patients. Therefore, we compare Federated Machine Learning to traditional testing methods such as antigen or PCR (Polymerase chain reaction) tests on economical and operational dimensions. We aim to inform essential decisions regarding the choice of diagnostic methodology during the progression of a pandemic. Accordingly, we run a time dependent simulation for Federated Machine Learning to digitally diagnose Covid-19 and find a significant potential. The federated deep learning model with six clients and full access to all datapoints achieves an F1-score of 89.7 percent. For comparison, the centrally trained model reaches up to 92 percent. Our results highlight the potential of applying Federated Machine Learning to Covid-19 diagnosis. The study might therefore function as a benchmark for hospital managers to contribute to future research while maintaining governance of their data.