Researchers in Denmark have developed a new algorithm that predicts an individual patient’s risk of mortality in the ICU. Their work, recently published in the journal Digital Health, demonstrates that the algorithm outperforms current non-computational methods of estimating mortality. Algorithms of this nature can help direct resources where they are needed most to best improve patient outcomes, and help catch problems early.
In the ICU, doctors and nurses already use various metrics in order to estimate an individual’s chance of survival, so as to determine the best treatment course and deliver optimal care. Yet, these metrics can be inaccurate in practice.
The latest research was performed using data from over 230,000 ICU patients, along with an 23 years of medical history. The algorithm utilizes neural networks to weigh the various factors in a patient’s medical history. At first, the algorithm was not very predictive. But, then the researchers included measurements and tests made in the first 24 hours of admission to the ICU, and this resulted in significant improvements to accuracy. The algorithm ultimately makes three predictions: (1) the risk of a patient dying in the hospital, which could be any number of days after admission, (2) risk of the patient dying within 30 days of admission, and (3) risk of the patient dying within 90 days of admission.
The Danish team found that age and length of previous hospital visits were two of the most important variables their algorithm identified. The length of medical history was also important, because they found diagnoses from many years earlier had an important effect on predicting survival after admission to the ICU.
“Excessive treatment is a serious risk among terminally ill patients treated in Danish intensive care units. Doctors and nurses have lacked a support tool capable of instructing them on who will benefit from intensive care. With these results we have come a significant step closer to testing such tools and directly improving treatment of the sickest patients,” said Professor Anders Perner from the Department of Clinical Medicine and the Department of Intensive Care, Rigshospitalet, in a press release.
“We ‘train’ the algorithm to remember which previous diagnoses have had the greatest effect on the patient’s chances of survival. No matter whether they are one, five or 10 years old. This is possible when we also have data from the actual admission, such as heart rate or answers to blood tests. By analyzing the method, we are able to understand the importance it attaches to the various parameters with regard to death and survival,” said Professor Søren Brunak from the Novo Nordisk Foundation Center for Protein Research.
Here’s a short video from University of Copenhagen about the research:
The study in Digital Health: Survival prediction in intensive-care units based on aggregation of long-term disease history and acute physiology: a retrospective study of the Danish National Patient Registry and electronic patient records.