How can artificial intelligence advance emergency medicine? A look at the current state

Nearly two years have passed since OpenAI unveiled ChatGPT, propelling artificial intelligence’s (AI) capabilities into the public spotlight and sparking a technological race to incorporate AI into virtually every field – from computer operating systems, to film writing, to ordering food at the drive-through. Beyond the excitement, large AI models have important pitfalls to pay attention to. Most importantly, the quality of its outputs is directly linked to the quality of its training data, where there is always the risk of bias.^​1​ Chatbots such as ChatGPT will also occasionally ‘hallucinate’ – worryingly inventing false information that it was never trained on.^​2​

In this article, we will delve into four key rapidly developing applications of AI relevant to emergency medicine (EM).

1. Enhancing department operations and management,
2. Triaging and documentation,
3. AI-Based Clinical Decision Support Systems
4. Improving Efficiencies in EM Research

Looking for a further introduction to AI before diving in? Check out this previous CanadiEM HiQuiPs article here!

1) Emergency Department (ED) Operations and Management:

ED overcrowding and bed-blocking have been an ever-increasing issue in the Canadian healthcare system that results in risks for patient care, including care delays, medical errors, and increased mortality.^​3​ Predicting ED volume is a difficult task given the complex interplay of multiple factors, rendering it difficult to develop staffing plans to accommodate for surges in ED utilization. However, AI algorithms can play a promising role in improving predictions.

As an example, Fralick et al. conducted a study using data from three hospitals in Ontario, in which an AI model was developed to predict ED patient volumes employing variables including the weather, patient acuity, and historical ED volumes.^​4​ Validated with prospective data, the model achieved an accuracy of 94% in predicting patient arrival volumes 72h in advance, with daily patient numbers ranging from 107 to 273. With the ability to accurately predict ED patient volumes, hospitals can look to optimize ED resource allocation, particularly regarding staffing.

2) Triaging and Documentation:

Natural language processing (NLP) is a subset of AI that refers to how computers analyze, process, and derive meaning from human language in a useful way. As a result, language learning models (LLMs) have been developed, which are AI models that can then take this understanding of human language and produce its own. You likely have experienced a primitive version of this capability through your Google Home, Amazon Alexa, or Apple’s Siri. NLPs and LLMs may help streamline ED work in two major ways – triaging and documentation.

When it comes to triaging, NLP models analyzing both structured and free-text data have shown high accuracy in identifying and triaging high-acuity patients. Wang et al. achieved the greatest performance coherence in predicting Emergency Severity Index using their “DeepTriager” model.^​5​  Kim et al. achieved a similar accuracy to DeepTriager in assigning a triage category from an auto-transcribed triage dialogue, which was only slightly lower than the performance achieved using human-transcribed triage dialogue.^​6​

In the realm of documentation, EM physicians may increase their patient volumes with AI reducing the time required for documentation. Many healthcare technology companies have developed tools that automatically generate clinician notes through NLP by listening to the initial patient assessment.

3) AI-Based Clinical Decision Support Systems

Machine learning (ML) is a subset of AI that involves training algorithms on vast amounts of data to make predictions or decisions. In the ED, ML can analyze vast datasets of clinical patient data to help stratify risk for outcomes of interest. This predictive capability can help ED physicians in decision-making and improve patient care by helping anticipate needs and potential complications.

For example, a study by Hsu et al. examined the use of ML to predict adverse outcomes among patients presenting with a hyperglycemic crisis to the ED.^​7​ 2,666 ED patients with hyperglycemic crises were analyzed using AI models to predict sepsis, ICU admission, and mortality. Using 22 features from their medical records, the model was able to effectively predict sepsis, ICU admission, and all-cause mortality. Once integrated with the hospital information system, the tool also outperformed the standard Predicting Hyperglycemic Death score in predicting mortality.

4) Improving Efficiencies in EM Research

NLPs and LLMs have significantly influenced research in EM, enhancing efficiency and accuracy in research-based tasks. For example, in systematic reviews, several startups have begun to develop NLPs that can facilitate study screening and data extraction by swiftly analyzing vast amounts of literature to identify relevant studies based on predefined criteria. This too has been used for chart reviews, attaining accuracy of up to 96% for paper screening and 98% accuracy for data extraction.^​8​ In early trials, these tools have been able to screen and extract up to 13-26% more than manual human reviewers!^​8​

5) Challenges, Limitations, and Future Directions

Some of the key challenges in the implementation of AI in the clinical setting are patient consent regarding the use of their data, and the risk of biases within AI algorithms that can perpetuate existing systemic biases leading to inequitable patient care. If a data breach occurs or a patient’s care is negatively influenced by AI, a difficult question arises regarding how accountability falls between the AI system, its developers, and the EM physician. AI is not immune to errors, and any generation of misleading information or ‘hallucinations’ may endanger patient safety and clinical decision-making, underscoring the need for robust validation and ongoing monitoring.

Most importantly, trust needs to be developed between ED physicians and AI algorithms, which will require time and further demonstration of AI’s safety and effectiveness. The future of AI in EM relies on thoughtful implementation without compromising on data privacy and security. Collaboration across disciplines is crucial, with ED physicians, ethicists, and AI developers jointly influencing AI’s development and deployment.^​9​ Additionally, ongoing education for healthcare professionals is vital to ensure they are equipped to use AI effectively and understand its limitations. As further research addresses these challenges, the use of AI in EM will become an essential component of care delivery in the coming years.

This post was copyedited by Tim Zhang (@_timothyzhang) and was edited by Daniel Ting.

Reviewing with the Staff

This article is important for several reasons:

Highlighting AI’s Impact on Healthcare: The article underscores how AI, particularly large language models (LLMs) and machine learning (ML), are transforming emergency medicine (EM). By improving operations, triaging, documentation, and clinical decision support, AI can significantly enhance patient care and efficiency in emergency departments.
Addressing Critical Healthcare Challenges: It addresses pressing issues like ED overcrowding and bed-blocking, which are critical problems in the Canadian healthcare system. By showcasing how AI can predict ED patient volumes with high accuracy, the article suggests practical solutions to optimize resource allocation and reduce risks associated with overcrowding.
Showcasing Real-World Applications: The article provides concrete examples of AI applications, such as the study by Fralick et al. on predicting ED patient volumes and the use of NLP for triaging and documentation. These examples illustrate the tangible benefits and potential of AI in real-world healthcare settings.
Raising Awareness of AI’s Limitations: It also highlights the pitfalls of AI, such as the risk of bias in training data and the phenomenon of AI “hallucinations.” This balanced perspective is crucial for understanding both the potential and the limitations of AI in healthcare.

Overall, the article is a valuable resource for healthcare professionals, policymakers, and anyone interested in the intersection of AI and healthcare, providing insights into both the opportunities and challenges of integrating AI into emergency medicine.

Dr. Sameer Masood

Dr. Sameer Masood is an Emergency Physician at the University Health Network. In addition, he has completed his Masters in Public Health at Harvard University, with a focus on quality improvement and clinical effectiveness. His interests include health technology integration into clinical care and quality improvement in emergency medicine.

@masood_sameer

• Bio
• Latest Posts

Richard Cheng

Richard Cheng is third-year medical student at Queen’s University. He has an interest in emergency medicine, trauma, and digital health. In his free time he can be found either enjoying or cooking a good meal with his friends.

Latest posts by Richard Cheng (see all)

• How can artificial intelligence advance emergency medicine? A look at the current state - November 19, 2024
• Approach to Unhoused Patients Presenting to the Emergency Department with a History of Opioid Use Disorder - January 23, 2024

• Bio
• Latest Posts

Timothy Zhang

Tim is an EM resident at McMaster following medical school at U of T. He is interested in anything related to trauma, POCUS, technology in medicine, and MedEd. Outside of medicine, his money gets drained by exploring restaurants, interesting recipes, the financial markets, and slowly learning to golf.

Latest posts by Timothy Zhang (see all)

• How can artificial intelligence advance emergency medicine? A look at the current state - November 19, 2024