If you’ve seen any American medical drama, you’ve probably watched a scene where someone comes in after a massively destructive event walking and talking normally. Then, about halfway through the episode, the person suddenly dies.
Not all emergency rooms are that tragic, but it can’t be denied that we have a doctor shortage or that patients are sometimes undertriage. At the same time, doctors who second guess their decisions may end up over-triaging a patient, resulting in expensive and unnecessary tests. Each of these scenarios leads to mental, physical, and financial ramifications for the patient and the doctor.
Artificial Intelligence may be able to help alleviate some of the frustrations.
Artificial intelligence is now being considered to help better triage hospital emergency patients. The models are constructed to estimate Emergency Severity Index Version 4 (ESI-4), which is then compared to the gold standard.
That should be exciting news, but before we discuss some of the research happening, we need to discuss some important considerations.
Models like this can sometimes discover that more deaths occur when a particular doctor, nurse, or hospital system is assigned to a case. These models do not necessarily have a full picture and can unintentionally cause the administration to act without fully investigating. When staff feels attacked by a computer, it can often lead to more issues in a technologically innovative environment.
Methods for AI-Based Triage
In 2017, a study was conducted to determine which methods would most accurately triage a patient. Patients can be triaged from level 1 to level 5 and selected randomly from the patients available. Due to the low level of admittance levels, 1 and 5 were omitted from the study. For levels 2-3, neural networks performed fairly well. However, in this study, a simple decision tree performed better at level 4. This particular study was conducted on only patients with acute abdominal pain.
What does this mean for your AI project? Be open to testing multiple model types. One model type may not be enough to get a good result. Also, the model constructed may not generalize well. This model may not perform well on patients without acute abdominal pain because they weren’t used in the initial study. Ensure that you are constantly revisiting any assumptions that are relevant in your experimentation and asking questions about the data you are using to train the model.
In 2019, a different study was conducted to compare which models better-triaged patients. It had a similar outcome. The traditional machine learning models outperformed the deep learning models (neural networks). This study differs from the first study because it was conducted on publicly available data from the 2007–2015 National Hospital and Ambulatory Medical Care Survey (NHAMCS).
NHAMCS collects a nationally representative sample of visits to non-institutional general and short-stay hospitals, excluding federal, military, and Veterans Administration hospitals, in the 50 states and the District of Columbia. The survey has been conducted annually since 1992 by the National Center for Health Statistics (NCHS). This means that the model was built on a more generalized and well-curated dataset.
What does this mean for your AI project? A generalized model may be possible but your real-time hospital data will be much messier than a curated survey. When structuring your experiment, this will need to be taken into account.
Still have questions about how to build the best triage model for your situation? Reach out to our machine learning team to see how we can help!