AI in the ER: How Artificial Intelligence can transform Emergency Medicine as we know it.

Almost everyone has been or knows someone that has been in the ER. According to the CDC, 45.8% of US citizens visited the ER in 2016 alone [1]. It is a vital part of every hospital, serving as an entry point for patients, assessing the urgency of their symptoms, conducting necessary diagnostics and giving initial treatment. That is why it is crucial for this department to operate smoothly. Yet it is not always so, especially in the time of the recent pandemic outbreak. So, how can AI contribute to alleviate some of the problems?

Assessment

At the entry of every Emergency Room, the patients coming in need to be triaged. It’s an initial assessment to gauge who is in need of immediate intervention and who can wait and make way for those. There are several ways and systems used by hospitals worldwide but all of them require time in order to retrieve and analyse relevant patient data. Moreover, it’s an initial assessment so every scale and algorithm has its shortcomings in identifying the outliers. That’s where Deep Learning comes in. By using available medical data, Korean researchers have created a Deep Learning network that is able to significantly outperform the existing traditional assessment algorithm (0.935 vs 0.785 AUROC).[2]

Detection

Sometimes, the symptoms may not be clearly visible to the human eye but are detectable for a neural network. Atrial fibrillation is a potentially dangerous condition where a patient’s arterial chambers have fast, irregular and ineffective contractions. It can be a symptom of other diseases such as pneumonia, cardiac arrest, or poisoning. A gold standard of detecting this condition is taking ECG and analysing it by a medical professional, which means it depends on equipment and staff availability. However, Chinese researchers have proposed another solution. By analysing video of the patients’ faces they were able to detect changes in the volume of the facial microvessels and detect patients who had atrial fibrillation in real time, acquiring sensitivity of 95.6% and specificity of 96.2%, respectively. [3]

Diagnosis

One of the most dangerous conditions is an acute intracranial hemorrhage with a 30-day mortality rate of 44%. It can be a result of a variety of conditions and events such as a head injury, hypertension, drug abuse or blood-clotting disorders. [4] A fast and accurate detection is crucial, because these patients often require life-saving neurosurgery and brain tissue is very susceptible to the lack of oxygen. A golden standard in intracranial hemorrhage detection is a computer tomography of the head (Head CAT scan), but it is difficult to do it accurately and swiftly, for a number of reasons (radiologist workload, poor signal-to-noise ratio, artifacts). It might be about to change, though, as researchers have put forward a deep learning model able to detect an intracranial hemorrhage, which is about as good as the experts from the American Board of Radiology. However, while physicians had unlimited time to examine the image, the time it took the network to do it was only one second. [5]

Treatment

Okay, but what about actually treating patients? Is there something AI can do to speed up the actual treatment? As it turns out, there is. Intravenous access is a standard way of delivering drugs and fluids to the hospital patients. In order to enable this, a medical professional needs to find a suitable vein (typically in the forearm), puncture the skin and get to the vessel with an over-the-needle catheter, minding that it doesn’t rupture. It takes time, expertise and sometimes several tries to do it successfully. In the ER, time is an essential asset and most drugs can not be administered before the IV access is set up. Fortunately, researchers have recently published an article in Nature Machine Intelligence journal, where they demonstrate a deep-learning robotic guidance for this. It reduced failed access attempts by sixfold and increased first-stick success rates from 53% to 88%, compared to standard manual access (gains were observed across various physiological conditions). [6]

In conclusion

Artificial Intelligence can significantly improve workflow, condition detection, diagnosis and treatment of ER patients. We should look forward to these advancements as these are upgrades that can super-charge our physicians, giving them more focus and time to actually save lives.

If you want to read more about AI, check out our site at Tooploox: https://www.tooploox.com/ai

References

[1] Rui P, Kang K, Ashman JJ. National Hospital Ambulatory Medical Care Survey: 2016 emergency department summary tables. 2016. Available from: https://www.cdc.gov/nchs/data/ahcd/nhamcs_emergency/2016_ed_web_tables.pdf

[2] Kwon J-m, Lee Y, Lee Y, Lee S, Park H, Park J (2018) Validation of deep-learning-based triage and acuity score using a large national dataset. PLoS ONE 13(10): e0205836. https://doi.org/10.1371/journal.pone.0205836

[3] Yan BP, Lai WHS, Chan CKY, et al. High-Throughput, Contact-Free Detection of Atrial Fibrillation From Video With Deep Learning. *JAMA Cardiol.* 2020;5(1):105–107. https://doi.org/10.1001/jamacardio.2019.4004

[4] Intracranial Hemorrhage https://emedicine.medscape.com/article/1163977-overview

[5] Kuo W, Hӓne Ch., Pratik Mukherjee P., Malik J., Yuh EL. Expert-level detection of acute intracranial hemorrhage on head computed tomography using deep learning. Proceedings of the National Academy of Sciences Nov 2019, 116 (45) 22737–22745; https://doi.org/10.1073/pnas.1908021116

[6] Chen, A.I., Balter, M.L., Maguire, T.J. et al. Deep learning robotic guidance for autonomous vascular access. Nat Mach Intell 2, 104–115 (2020). https://doi.org/10.1038/s42256-020-0148-7