Smart Expert System Assists Medical Diagnostics

A current research project develops a system designed to support critical care physicians in the future. The research project "A Learning and Interoperable Smart Expert System for Pediatric Intensive Care Medicine (ELISE)" uses data collected via machine learning algorithms to assist diagnostic decision-making.

In this interview, Dr. Antje Wulff and Dr. Thomas Jack explain why the project was launched and describe the improvements that can be achieved with the system.

Dr. Wulff, why was this project initiated?

Dr. Antje Wulff: The original impetus behind it was a budding collaboration between the Peter L. Reichertz Institute for Medical Informatics and the Department of Pediatric Cardiology and Intensive Care Medicine at the Hannover Medical School. The objective was to collect and analyze the data of children in pediatric intensive care. As part of my master’s thesis, the idea came up to detect a specific inflammatory syndrome in children using rule-based expert systems and vital signs and laboratory values. This initial research project was so successful over several years that we received funding by the Federal Ministry of Health (BMG) last year to continue and expand the project.

Dr. Jack, what prompted you to launch the project?

Dr. Thomas Jack: We were one of the first pediatric intensive care units that had created a so-called Patient Data Management System (PDMS). This opened the door and made us one of the first pediatric intensive care units in Germany to facilitate comprehensive digital data storage. The system merged and stored all the patient’s vital parameters including vital signs, medication data, and lab data in one digital system. To further tap into the potential of this setting, we looked for ways to expand this perspective and came up with the research project idea.

How would this smart expert system for pediatric intensive care medicine work?

Wulff: Our goal is to create a system that supports healthcare professionals in their daily work. More specifically, we would initially add a demonstrator module to the patient data management system: the smart decision support system. Based on machine learning algorithms, the system could then detect certain diseases and alert the medical staff. It is important to note that this is not an autonomous system, meaning it will not automatically start therapy or immediately administer medication. It is primarily designed to make a probable diagnosis.

Jack: Our project focuses on the detection of Systemic Inflammatory Response Syndrome (SIRS) and the diagnosis of severe (multiple) organ dysfunction syndrome and resulting organ failure - a problem that frequently occurs in the intensive care unit. SIRS shows similar symptoms to sepsis and is often associated with various diseases or interventions in critical care settings. SIRS is often not detected immediately, which is where the system would assist a diagnosis thanks to systematic and automated detection. The system always inquires whether the automated diagnosis was correct and learns autonomously based on the feedback. This self-learning function is the ultimate goal, which cannot be put into effect immediately. For now, the system is designed to support treatment through an early diagnosis, since many intensive care decisions are time critical. For example, if the patients rapidly develop kidney or liver failure during the post-operative period, early medication adjustments to reflect the situation could improve patient outcome. 

What technologies does the project study or utilize?

Wulff: We basically have two principal components: rule-based processes and machine learning methods. Both methods are meant to play a role in the system, most likely resulting in a hybrid system. It is also important for the development to focus on open and standardized interfaces. The goal is to use the so-called interoperability standard, meaning to present the collected data in a way that allows access, exchange, and cooperative use of data in a coordinated manner, within and across organizational boundaries. This aspect is very important to us because if you don’t adhere to these standards, you inevitably create point solutions that only work in one hospital.

What are the next steps?

Jack: We are currently in the middle of the first half of the project and are working closely to implement further aspects of SIRS and organ dysfunctions into rule-based machine learning. This includes renal or hematological dysfunction, for example. We continue to train the system to recognize more criteria for the diseases. Every new identified rule means we conduct a new study to show that the system interprets the data correctly and accurately captures the target criteria.

What is the project objective?

Wulff: The ELISE project wants to develop decision support concepts to determine how much critical care physicians can benefit from these types of systems. To help us with this, we use various diseases that are associated with SIRS and sepsis, including organ dysfunction and signs of impending organ failure.

Jack: Ultimately, our systems are meant to offer on-site decision-making support for physicians as they cope with the stressful intensive care environment. This also provide an immediate benefit for patients.

Can you give a progress report at this juncture?

Jack: We conducted a study that proves the feasibility of the system’s implementation. We have taken a path that may soon prompt PDMS systems to also gradually turn into diagnostic systems that increase patient safety in the intensive care unit. Thanks to the multidisciplinary collaboration and the opportunities afforded by our funding, the project will relatively quickly find its place in routine critical care operations - of course, while complying with all current data protection requirements and Medical Device Regulations.

Wulff: We conducted a clinical trial prior to winning funding, which means we now have a prototype decision support system with real data. This allowed us to compare the data that was analyzed by the system with the diagnosis made by the on-site physician. It already indicated the system works reliably and might possibly be even more accurate than the physician. Needless to say, this is just an initial intermediate result, which now requires an evidence-based evaluation.