BabSim.Hospital: forecasting hospital bed needs

Amid the coronavirus pandemic, capacity of intensive care units is a hot-button issue as this determines the number of severe COVID-19 cases hospitals can treat. The percentage of currently infected patients can deliver insights into the projected need for ICU beds down the road. The TH Köln developed a tool to help with this planning process.

In this MEDICA-tradefair.com interview, the team behind "BaBSim.Hospital" reviews the tool and explains how it can assist hospitals and government agencies.

You developed the "BaBSim.Hospital" tool. How does it work?

Eva Bartz: The tool facilities long-term capacity and resource planning for hospitals. It allows them to forecast medical supply needs, plan care and medical capacities, and validate the average length of stay for every ICU bed or isolation ward with or without mechanical ventilators.

Prof. Thomas Bartz-Beielstein: It is based on an idea by Dr. Tom Lawton, an intensive care specialist from the United Kingdom, who developed the tool to assist critical care resource planning. In early 2020, we modified and expanded this approach to adapt to the coronavirus pandemic.

The tool performs a so-called discrete event simulation. You could say it considers every patient who comes to the hospital and flows from one hospital ward to the next. Most other models do not work that way —they set up equations and try to approximate them. Our model is much closer to the individual patient, doctor, or bed. We would therefore also like to discuss and review it with other critical care physicians to optimize it for specific scenarios.

What inputs can users add and what is their output?

Bartz-Beielstein: We factor in both the average length of a patient’s stay in a ward and the probability of how long the patient actually stays for this length of time and the disease progression, meaning which route he takes after admission.

We use a trick: We take data provided by the physicians and use AI to optimize this data based on the currently available information of the Robert Koch Institute and the DIVI Intensive Care Register. The probabilities are personalized to reflect the age and gender of the patient. In doing so, AI calculates the risk of each individual patient.

Frederik Rehbach: We update the data provided by the Robert Koch Institute (RKI) and the DIVI daily. AI subsequently processes the data and adjusts the different parameters in the simulation. In doing so, new data is calculated daily for each region.

Bartz-Beielstein: The web interface allows users to indicate how they want to model the pandemic events. Users can set the R value, which enables them to see a comparison of the data from our model with the real-life data on the previous infectious incidents and subsequently as a projection of how the figures will develop under the set conditions within the specified period.

At some point during the development, did the model bear comparison with reality?

Bartz-Beielstein: It already did back in the spring during the first wave. We used data from the Oberbergische Kreis (a local district) here in the state of North Rhine-Westphalia, Germany. The model turned out to be so productive and reliable that we decided to enhance it. This is the major advantage our model has over others: we use local information and get accurate, local statements as a result.

Prof. Olaf Mersmann: Discrete event simulations and AI processes are complex technical systems. But unlike other models, ours is not a black box approach. We can explain the results to a physician or a crisis expert, allowing him to interpret the times and transition probabilities.

Bartz: Flexibility takes place on two levels: first, by adapting the simulation itself, which can take local events in a hospital or local district into account, and second, for users who want to assess different pandemic events. We simulate this by adjusting the R-value at the start time and the end time of each specific event. The user can assess the importance of initiating a lockdown immediately versus next week because the model shows how hospitals would fill their available beds with infected patients during this time and demonstrates the effects of earlier or later measures.

Are you planning to advance the tool?

Bartz-Beielstein: We run this model in addition to our regular work here at the TH Köln. Nearly 50 students are involved in this endeavor as part of their thesis or project. However, we also welcome corresponding research funds and grants to allows us to increase our focus on this subject.

We presently incorporate the impact of vaccinations – providing the option of entering the percentage of the vaccinated population and how it effects changes in infections. There are other conceivable additions, but this is something we can only address if we receive the respective feedback and support from policymakers and, most notably, physicians who are interested in this aspect. We would like to take additional steps and implement more functions.