AI predicts a patient's risk of postoperative complications

Whether it is a routine surgery or a personalized surgical intervention that perhaps has never been done before: There is always a residual risk involved. That’s why hospitals monitor and supervise patient care before, during, and after surgery to be ready for immediate intervention if needed. If postoperative complications could be predicted, it would allow medical professionals to monitor high-risk patients more closely and effectively.

In this MEDICA-tradefair.com interview, Prof. Bettina Jungwirth talks about the use of artificial intelligence (AI) in anesthesiology and in the intensive care unit and explains how it benefits patients and hospitals.

Prof. Jungwirth, you study the use of AI in anesthesiology. Can you explain what that might look like?

Prof. Bettina Jungwirth: The goal of AI in anesthesiology and critical care medicine is to use big data to sustainably improve patient outcomes via personalized perioperative treatment. This necessitates individualized risk prediction, which has already been addressed in several studies. The objective is to effectively predict postoperative nausea and vomiting or patient deterioration in the recovery room, or to identify patients at risk for postoperative delirium.

Anesthesia per se has become very safe, making anesthesia-related complications rare these days. This allows us to focus on the postoperative period also known as the recovery period and complications that may occur after surgery. The idea is to use risk prediction models to identify and avoid these aspects at an early stage.

What types of data would you use?

Jungwirth: Anesthesiologists monitor vital signs such as blood pressure, heart rate or oxygen saturation, and also include respiratory and laboratory parameters. They might also track process parameters: How did the patient end up at the hospital and what was his journey? This refers to diagnostic imaging and data pertaining to medication administration, dosages, and volumes.

As you can see, it involves different types of data. That’s why it is key – if not essential – to structure the data to ultimately make it usable for AI.

Does specific data indicate a higher risk for surgical complications?

Jungwirth: We already consider parameters such as pre-existing conditions, patient age or the type of surgery the patient is scheduled for in common risk scores. However, I do not think this is comprehensive enough to accurately predict the risk for an individual patient. We should include more routine stats such as lab data, vital signs, and process data to facilitate an accurate risk prediction.

What is the ultimate benefit?

Jungwirth: Accurate risk prediction is always our ultimate goal and, based on this, subsequent prevention thanks to personalized and targeted therapy. There is still no respective scientific evidence at the moment. However, at this juncture, we presume that individualized treatment will enable us to either avoid complications altogether or at least assess which patient needs closer postoperative monitoring based on the risk prediction.

During the SARS-CoV-2 pandemic, we aimed to ensure the adequate capacity of intensive care beds. AI-driven risk models would not only allow us to predict postoperative complications but also study patients in the hospital wards and determine whether there is a relevant risk of them ending up in the intensive care unit. And if so, we could establish the length of their stay in critical care. In recent months, German hospitals have intentionally delayed elective surgeries with a likely need for post-acute care to free beds for the most critically ill COVID-19 patients. AI methods would give hospitals and wards a great tool for more reliable capacity planning.

Are you planning to implement AI techniques in Ulm?

Jungwirth: Absolutely. We are presently building the basic IT infrastructure that we need to collect structured, electronic data. This is a PDMS, a patient data management system, meaning an electronic anesthesia and intensive care chart that automatically records data in a way that makes it usable for AI methods.

During my tenure at the Technical University of Munich (TUM), I had already initiated AI research. Today we have forged a collaboration with the Technical University of Munich, which also benefits our research team at the University Hospital of Ulm. We have built an extensive database that comprises anesthesiological and perioperative data in Munich. It maps the patients before, during, and after general anesthesia including up to the intensive care unit. We take this data to develop our predictive models, i.e., the risk calculator for complications or stay in an intensive care unit. Subsequently, we plan to team up with other hospitals and clinics for multicenter research trials to assess whether our models help improve postoperative outcomes.

What are the prerequisites for AI use in anesthesiology?

Jungwirth: Data quality is the most challenging aspect of AI methods. You need data that is clean and machine-readable. Unfortunately, the consensus is that 70 to 80 percent of all health care data is unstructured data. This makes it unusable for AI methods. In other words, clinical free-text notes are disastrous.

If we plan to conduct multicenter research trials or share data among hospitals – including on a global scale – we simply must speak the same language. This data could then also be integrated into very large databases. It is an essential prerequisite for us to use artificial intelligence methods in a way that we actually prefer to use them.