Hospital: how an AI tool could improve patient safety
Hospital: how an AI tool could improve patient safety
Interview with Prof. Maik Kschischo, Professor of Biomathematics, University of Applied Sciences Koblenz
Sometimes, a hospital stay can proceed successfully without a hitch. At other times, there might be an unexpected turn of events if the patient exhibits complications.Early identification of these patients could prevent unnecessary suffering.A new research project intends to develop an AI-based tool that predicts a patient's risk of complications at an early stage.
Prof. Maik Kschischo
In this MEDICA-tradefair.com interview, Prof. MaikKschischotalks about events that trigger in-patient deterioration, explains how an AI tool could facilitate early detection and reveals the software development challenges his team might face.
Prof. Kschischo, you are developing an in-patient critical event prediction software tool. Can you tell us how it is meant to work?
Prof. Maik Kschischo: Critical events are complications that lead to a dramatic decline in the hospital patient’s condition, life-threatening situations or even death. This may include sepsis, an overlooked pre-existing condition or drug-drug interaction.
Our tool is designed to identify possible event indicators and triggers and alert hospital staff. Physicians of our project partner, the Marienhaus Hospital Group already analyze patient events retrospectively and look for likely triggers.
Our first step is to study the hospital information systems (HIS) to identify common triggers and make this identification easier for physicians. The second step is what makes our approach different and pertains to AI-assisted automated and prospective trigger identification that alerts physicians or medical staff to take a closer look at a patient.
What role does the ongoing coronavirus pandemic play in this setting?
Kschischo: The development presently takes place against the backdrop of the global corona pandemic and focuses on COVID-19 patients, of whom a percentage are in critical condition, admitted to the ICU and where death might occur despite full intensive care support. We must consider two aspects in this context: You have the SARS-CoV-2 infection itself and the burden on hospitals caused by the high admittance of people infected with the virus, which can lead to overlooking early warning signs of critical transitions.
Products and exhibitors around the use of data in health care
Discover more interesting products and exhibitors in the database of virtual.MEDICA 2020:
An AI-based tool that recognizes possible triggers for a worsening of a patient's condition and warns hospital staff in time could improve patient safety in the hospital.
How might this tool be used in practice?
Kschischo: In practice, we would integrate the trained tool in the HIS and allow it to continually comb through and analyze patient data. It is meant to calculate and identify if a patient’s condition is critical or if the data indicates event triggers.
At first, the tool might not identify all aspects, but hopefully it would detect what physicians spot, meaning it utilizes the available knowledge. If we let the system run and allow it to learn, it will gradually evolve and self-improve. For example, it will filter out specific event constellations that are presently not yet considered critical.
What's next for your development?
Kschischo: Right now, we are still in the early stages. The research question is whether this approach works the way we want it to work and whether its implementation is feasible. Having said that, we have hope and reason to believe it will.
First, we try to build a good data base that is secure and in compliance with applicable data protection laws. We include patient-generated health data of the Marienhaus Hospital Group and data from damedic, a startup from Cologne, Germany, that fosters AI innovation in healthcare. We also interview doctors and intend to collect and pool the existing knowledge about triggers.
We may subsequently conduct initial tests using hospital data. It will enable us to examine whether we can design an AI system that reliably identifies the triggers and does not constantly prompt false alarms or no alarms at all.
The next step is to make sure that physicians can interpret the tool’s statements. This must not be an AI black box model – as is common with AI systems – and the doctor must be able to understand why an alarm is triggered. The keyword here is "explainable AI", which means we must give physicians indications that make them appreciate why the situation is critical, otherwise the tool will not invite trust and inspire confidence. This is one of our biggest challenges.
What is the project development life cycle?
Kschischo: The project is scheduled to take 18 months, at which point we hope to have a prototype and in-house testing completed to launch an initial HIS test run. As part of this pilot study, we will hopefully also be able to evaluate the user experience. If all goes well, the Marienhaus hospitals would like to use the tool. We must then assess whether we can offer this option to other hospitals as well.
At that point, the coronavirus pandemic might be over, or we no longer have as many problems managing it. Right now, our tool development still focuses on COVID-19 patients, but we can essentially extend it to other events and illnesses. Amid the pandemic crisis, it would be instrumental to support the medical staff, but the goal is to also make it useful and valuable beyond the COVID-19 pandemic.
More topic-related exciting news from the editors of MEDICA-tradefair.com: