When a patient arrives at the emergency unit at a Danish hospital, one of the many decisions to take is which ward the patient should be admitted to. The decision is taken, among other things, based on an assessment of the length of stay at the hospital. At Aalborg University Hospital, for example, there is a ward for patients where you expect a maximum length of stay of 48 hours.

If, on the other hand, a longer stay is expected, the patient will immediately be admitted to a special ward. In a new research project, researchers from the Department of Computer Science, Aalborg University, in collaboration with the North Denmark Region, have developed an artificial intelligence model to predict the duration of the individual patient's hospitalization to support the staff's decision about where the patient should be moved from the emergency unit.

- The aim of the project is in the first instance to reduce time and resources spent on moving patients around between wards. In addition, fewer moves will be an advantage for the individual patient, since being moved between wards is typically a source of stress, Emil Riis Hansen explains. He is a PhD student at the Department of Computer Science, Aalborg University, and the main author of the scientific article that describes the model.

No system today

Today, there is no system that can support the individual clinician in assessing the duration of a hospital stay. Statistical data may give an overview of, for example, the months during which wards are typically under pressure, but the concrete assessment of the duration of a specific patient's hospitalization is entirely up to the clinician who sees the patient in the emergency unit.

- It is important to emphasize that our method does not replace the staff's assessment but can be used to support the decision in a specific case. A system cannot see the patient, so we need a clinician to take a decision based on the overall picture, says Emil Riis Hansen.

A sequence of events

The method developed in the research project assumes that a patient course consists of a sequence of events in a given order. These events can be tests and measurements, scans, medication, etc. Each event yields a result, and each result is assigned a value – high, medium, or low – based on medical reference values ​​relating to demographics such as age and gender. By training a so-called BERT Natural Language Processing (NLP) model, one can then predict the duration of a hospital stay.

- It is important to be able to handle the sequence in a patient process, as it reflects the clinician's view of the patient. If, for example, the patient has initially received pain-relieving medication, but is then sent for a scan, this may indicate that the condition is more serious than first thought, Emil Riis Hansen explains.

Improvement of existing models

The fact that the new model is a sequence model meaning that it can take into account the sequence of events in a patient process is an improvement compared to models developed in previous research projects.

- If you want to use machine learning to solve a problem, you typically look for the basic models that you know will often work. The problem is that these models - which have also been used in other research projects in this domain - work on data contained in rows and columns in a table. There are at least two problems with that: You don't get the temporal dimension, and if there is missing data in a field in the table, e.g. a blood pressure value, then you have to insert some other value for the model to be able to run - which means that a certain degree of imprecision is introduced. We avoid that with the model we have developed. Overall, our current experiments suggest that sequence models can be more accurate than traditional machine learning models, Emil Riis Hansen explains.

Big data

In connection with his PhD studies, Emil Riis Hansen has worked in the Business Intelligence and Analysis department of the North Denmark Region. This has given him access to anonymised medical data from more than 45,000 patients from the region.

- It has been fantastic to have access to such a large amount of medical data, which we have been able to use to train the model. However, there are additional data types that could be very relevant to include to make the model even more reliable. It could be imaging data, the clinicians' notes, and probably the most important type of data: the staff on duty in the emergency unit when the patient entered. The reason is that the level of experience of the clinician who first sees the patient can have an impact on how quickly and precisely a decision is taken about the next step in the process, says Emil Riis Hansen, who presented the results from the research project on Tuesday 13 June, 2023, at a large international conference in Slovenia on the use of artificial intelligence in healthcare.