Authors: Abirami Kirubarajan, Sameer Masood
During a busy night shift, you are reading a series of chest X-rays when you overhear a colleague remark, “This will all be replaced by computers soon, anyways”. You think back to a few Tweets on artificial intelligence in medicine and wonder if her comment has merit.
There has been so much hype about the topic of Artificial Intelligence (AI) in recent years, much of which is optimistic and some of which is cautionary. Moreover, there have been many different terms and possibilities attributed to it – but what is AI? How is it currently used in Emergency Medicine? And what could the future hold?
In this month’s post, we will provide an introduction to artificial intelligence and its potential applications in emergency medicine. Much of the findings will be presented from our recently published scoping review in the Journal of American College of Emergency Physicians Open: “Artificial intelligence in emergency medicine: a scoping review”.(1)
What is artificial intelligence?
Simply put, AI is a broad subset of computer science that attempts to mimic human cognitive ability. AI can include speech recognition, predictions, and problem-solving. A popular subset of AI is machine learning, which includes the ability to improve itself autonomously, without human input. Deep learning is a subset of machine learning which is based on artificial neural networks. Natural language processing, which involves the analysis and manipulation of human language, can include all three levels of artificial intelligence, machine learning, and deep learning.
Broad categories of AI can include supervised learning, unsupervised learning, reinforcement learning, and natural language processing.
AI has received considerable buzz and media attention for its applications in medicine given its ability to outperform human expertise in clinical predictions, diagnosis, and monitoring. The emergency department (ED) may be uniquely situated to benefit from AI, due to the diversity of patient information, the necessity of balancing probabilities for risk stratification, and the rapid decision-making.
However, if you don’t have a computer science background, it can be difficult to understand which AI interventions are promising and which require further study. The media “hype” can also be difficult to decipher.
What is the current state of AI research?
From a recent scoping review, a total of 150 articles specific to emergency medicine and artificial intelligence were found.(1) Of the 150 studies, 37 interventions were aimed at improving diagnosis within the ED. Radiology was a particular area of interest, with 19 studies focused on how to improve diagnostic imaging.
The snapshot of the current field of literature was promising, however the majority of included studies were retrospective in nature, with only 3 prospective controlled trials. AI interventions were better able to diagnose acute cardiac events, identify hyperkalemia, risk-stratify patients in triage, identify participants, predict wound infection, predict mortality, predict patients for clinical trials, and read types of imaging.(1) Most studies however did not control against human performance, and lacked high-quality evidence and control groups.
A few examples of these studies are below:
Beyond clinical care, AI was also found to be useful in predicting hospital wait-times and optimizing staffing hours.(6,7) Two other studies also used AI to rapidly identify patients for research recruitment in the ED.(8,9) As AI is used more frequently in policy planning, we can hope to see its benefits extend into the ED as well.
Considerations for AI Implementation
There are a few reasons why AI might be able to outperform human clinicians. For one, AI has the ability to process multiple variables simultaneously across large data sets. Humans often struggle when balancing numerous data points to predict outcomes, and our decision-making is often subject to biases and heuristics. In addition, AIs can take advantage of large data sets for stronger pattern recognition, which is particularly relevant in fields such as radiology. For example, an AI can process databases made of hundreds of thousands of radiographs and their reports, resulting in an algorithm that can accurately diagnose new radiographs based on pattern recognition. That’s one reason why there is a large focus on AI with radiology, such as the detection of fractures.(5)
However, there are a few things to keep in mind. For one, we need stronger evidence before widespread implementation. It isn’t enough to show that AI can accurately predict on its own – instead, it is important to show that it is consistently better than a human. In addition, AI needs to be tested in pragmatic real world settings that mimic the rapid, unpredictable environment of an ED – there isn’t much use for an AI that only works in a tightly controlled, study setting. Human errors and lack of familiarity can contribute to a lack of uptake or poor implementation.
Another concern for AI is its “black box” nature (i.e., its lack of transparency), in which it is difficult for an observer to determine what goes “into” an AI. We have seen in other fields how the opacity of technology has inadvertently discriminated against certain patients, or coded for incorrect outcomes. For example, a widely used American hospital algorithm was noted to have unintentional racial bias when allocating healthcare resources for Black patients.(10) Transparency and audits should be recommended, not only to ensure the correctness of the algorithm, but also to build public and provider trust.
So if you come across a new article on AI, what are things to consider when reading AI research?
Is there a control group?
Does the intervention only work in a narrow, tightly defined population?
Did the authors study implementation or uptake?
Is there transparency regarding which variables are included and how the algorithm is developed?
Has the intervention been validated in a different population than the population it was developed in?
Is there a conflict of interest between the evaluator and the intervention?
Recap
Ultimately, AI may indeed change the practice of emergency medicine through its simulation of human cognitive ability. Many of us are looking forward to the improvements in diagnosis and prediction. However, it is important to cut through the “hype” of AI and read research with a critical eye. We highly suggest reading our scoping review to learn more about the snapshot of AI research relevant to the ED.(1)
After doing some research into artificial intelligence and its application in medicine, you have a better understanding of its role, but also its limitations. You are optimistic that artificial intelligence will be an improvement to the field of medicine, but are also going to be sure to evaluate new and upcoming research critically with the new insight that you have gained.
That's all for this post! We hope you have a better understanding of artificial intelligence, and particularly its application to emergency medicine. If you have any topics you would like us to feature next, please reach out to the HiQuiPs team!
Senior Editor: Ahmed Taher
Copyedited by: Daniel Dongjoo Lee
References
Kirubarajan A, Taher A, Khan S, Masood S. Artificial intelligence in emergency medicine: A scoping review. J Am Coll Emerg Physicians Open. 2020;1(6):1691-1702. doi:10.1002/emp2.12277
Clarke J, Hayward C, Santora T, Wagner D, Webber B. Computer-generated trauma management plans: comparison with actual care. World J Surg. 2002;26(5):536-538. doi:10.1007/s00268-001-0263-5
Jenny M, Hertwig R, Ackermann S, et al. Are Mortality and Acute Morbidity in Patients Presenting With Nonspecific Complaints Predictable Using Routine Variables? Acad Emerg Med. 2015;22(10):1155-1163. doi:10.1111/acem.12755
Lammers R, Hudson D, Seaman M. Prediction of traumatic wound infection with a neural network-derived decision model. Am J Emerg Med. 2003;21(1):1-7. doi:10.1053/ajem.2003.50026
Lindsey R, Daluiski A, Chopra S, et al. Deep neural network improves fracture detection by clinicians. Proc Natl Acad Sci U S A. 2018;115(45):11591-11596. doi:10.1073/pnas.1806905115
Zlotnik A, Gallardo-Antolín A, Cuchí A, Pérez P, Montero M. Emergency Department Visit Forecasting and Dynamic Nursing Staff Allocation Using Machine Learning Techniques With Readily Available Open-Source Software. Comput Inform Nurs. 2015;33(8):368-377. doi:10.1097/CIN.0000000000000173
Azadeh A, Yazdanparast R, Abdolhossein Z, Keramati A. An intelligent algorithm for optimizing emergency department job and patient satisfaction. Int J Health Care Qual Assur. 2018;31(5):374-390. doi:10.1108/IJHCQA-06-2016-0086
Ni Y, Bermudez M, Kennebeck S, Liddy-Hicks S, Dexheimer J. A Real-Time Automated Patient Screening System for Clinical Trials Eligibility in an Emergency Department: Design and Evaluation. JMIR Med Inform. 2019;7(3):e14185. doi:10.2196/14185
Ni Y, Beck A, Taylor R, et al. Will they participate? Predicting patients’ response to clinical trial invitations in a pediatric emergency department. J Am Med Inform Assoc. 2016;23(4):671-680. doi:10.1093/jamia/ocv216
Obermeyer Z, Powers B, Vogeli C, Mullainathan S. Dissecting racial bias in an algorithm used to manage the health of populations. Science. 2019;366(6464):447-453. doi:10.1126/science.aax2342