Podcast: Play in new window

Subscribe: RSS

Reference: Kareemi et al Artificial intelligence-based clinical decision support in the emergency department: a scoping review. AEM April 2025.

Date: April 15, 2025

Guest Skeptic: Dr. Kirsty Challen is a Consultant in Emergency Medicine at Lancashire Teaching Hospitals.

Case: It may be April, but as you sit in your departmental meeting with your emergency physician colleagues, you all note that the winter “surge” of patients hasn’t stopped. The decision fatigue at the end of shifts is as present as ever. “Surely AI will be making some of these decisions better than us soon?” says one of your colleagues, only half joking. Another colleague chips in that the medical students at the nearby university have been warned against using ChatGPT to create differential diagnoses and you are left wondering whether AI might be “working” in the ED soon.

Background: Emergency departments can be a high-pressure environment. Clinical decisions must be made quickly and accurately, often with incomplete information. Clinical decision support (CDS) tools aim to address this challenge by offering real-time, evidence-informed recommendations that help clinicians make better diagnostic, prognostic, and therapeutic decisions.

CDS spans a wide spectrum from traditional paper-based clinical decision rules to smartphone apps (MDCalc) to more integrated systems into electronic health records (EHRs). These tools function by combining patient data with expert-driven algorithms or guidelines to inform care pathways. They can help determine disease likelihood, risk stratify patients and even guide resource utilization such as imaging or admission decisions​.

Recent years have seen a growing interest in applying artificial intelligence (AI), particularly machine learning (ML), to CDS. Unlike traditional “knowledge-based” CDS that relies on literature-based thresholds, AI-driven tools derive patterns from large datasets (“big data“) to identify associations and make predictions. These “non–knowledge-based” systems promise to augment human decision-making by uncovering insights that might be overlooked by clinicians or static rules​.

However, the majority of AI-based CDS (AI-CDS) tools remain in early development. Few have been rigorously tested in the ED, and even fewer have demonstrated improvements in patient outcomes or clinician workflow. Despite FDA clearance for some tools, evidence for real-world impact remains limited​. Emergency physicians are right to approach this technology with skeptical optimism. We will need to balance the transformative potential of AI with a critical eye toward evidence, safety, and usability.

Clinical Question: (1) What is the current landscape of AI-CDS tools for prognostic, diagnostic, and treatment decisions for individual patients in the ED? and (2) What phase of development have these AI-CDS tools achieved?

Reference: Kareemi et al Artificial intelligence-based clinical decision support in the emergency department: a scoping review. AEM April 2025.

  • Population: Studies involving AI or ML-based clinical decision support tools applied to individual patient care in the ED, published 2010 – 2023.
    • Excluded: Models that assessed a specific test (e.g. imaging) without clinical context, administrative or operational outcomes (e.g. patient census), models involving irrelevant data (e.g. vignettes or data not available following the emergency assessment), length of stay as a primary outcome, studies without full text or abstract in English.
  • Intervention: AI- or ML-based clinical decision support tools used during patient care in the ED.
  • Comparison: Not applicable for a scoping review. However, the review identified whether studies involved any comparison with usual care, clinician judgment, or non-AI tools.
  • Outcomes: The review didn’t focus on a single outcome but instead categorized studies by their targeted clinical decision task—diagnosis, prognosis, disposition, treatment, etc. Outcomes were only included if they were relevant to emergency clinicians’ decision-making, such as predicting ICU admission, mortality, or need for intervention​.
  • Type of Study: Scoping review.

Dr. Hashim Kareemi

This is an SGEM HOP and we are pleased to have the first author on the show. Dr. Hashim Kareemi is an emergency physician and researcher at Vancouver General Hospital who is exploring ways to improve the development and implementation of artificial intelligence models in emergency medical care.

Authors’ Conclusions: “We found a large number of studies involving a variety of clinical applications, patient populations, and artificial intelligence models. Despite an increased rate of publication in recent years, few studies have advanced from preclinical development to later phases of clinical evaluation and implementation.”

Quality Checklist for Scoping Reviews:

  1. Was the main question being addressed clearly stated? Yes
  2. Was the search for studies detailed and exhaustive? Yes 
  3. Were the criteria used to select articles for inclusion appropriate? Yes
  4. Were the included studies sufficiently valid for the type of question asked? Yes
  5. Were the results similar from study to study? Yes
  6. Were there any financial conflicts of interest? This project was funded by the Canadian Institutes of Health Research. Several authors have financial interests declared in AI companies or research.

Results: The found 5,168 records were identified electronically, and 605 studies were included in the final analysis. Publication rates have increased significantly since 2019. Many studies (40%) came from North America, and under 1% from Africa.

Key Result: Despite a rapidly increasing volume of studies across a breadth of clinical applications, few describe advanced phases of testing or implementation.

  • Outcomes:
    • The majority of data came from retrospective studies (79%)
    • The most common outcome category was for prognosis (45%)
    • There were only a few high-quality trials with only the randomized trial protocols and one quasi-experimental study. There were no published RCTs in live ED settings.
    • The majority of studies were in the preclinical (“in silico“) phase; <3% reached clinical implementation or post-market surveillance​

Listen to the SGEM podcast to hear Hashim’s response to our five nerdy questions.

  1. A Different Kind of Nerdy: There is a lot of technical detail in this field and in your data. Could you start by helping us with some definitions – is there a difference between AI and machine learning? You also mention supervised versus unsupervised machine learning.
  2. Article Screening: Your initial search found over 5,000 records, and you screened the full text of 721. This is impressive. Were you not tempted to use AI at any point to help with this?
  3. Anglocentricity: You excluded papers which didn’t have either the full text or an abstract in English – from figure 1 this was 6 papers in the end (so under 1%). Despite this, most of the literature came from North America, Asia and Europe. Arguably, AI could be of more potential benefit in parts of the world like Africa, where healthcare resources are more stretched, but these aren’t represented. Did you get any feel from your review of whether this might happen or why it isn’t?
  4. Outcomes: You found that the largest group of tools (270 of 606) used the AI-CDS to inform prognosis. As a clinician this makes me wonder “so what”? Knowing the expected clinical course can be useful, but I’m not sure I need AI to tell me that a frail 102-year-old with kidney failure is unlikely to do well, or that I have limited scope to change that! Did you find that there was generally any exploration of what the AI-CDS tools added from the patient’s point of view?
  5. Development Phases: You highlight that most of the literature is from the early phases of AI-CDS development – ie preclinical and offline validation, with few tools undergoing large-scale safety and effectiveness trials or post-marketing surveillance. How do you think regulators, administrators and the EM clinical community should use this information?

Comment on Authors’ Conclusion Compared to SGEM Conclusion: We agree with the authors’ conclusions.

SGEM BOTTOM LINE: Artificial intelligence–based clinical decision support tools in the ED show promise, but we need rigorous evaluation before routine implementation​.

Dr. Kirsty Challen

Case Resolution: You agree with your colleagues that this is a rapidly expanding field but that your jobs are safe for another few years.

Clinical Application: You ask the Chief Clinical Information Officer [NB this is the UK term, not sure of a North American equivalent?] to attend the next staff meeting to discuss the potential benefits and harms of AI-CDS in the ED.

Keener Kontest: The last episode’s winner was our good friend Steven Stelts from New Zealand. He knew sudden onset shortness of breath or dyspnoea is the most common presenting symptom for pulmonary embolism.

Listen to the SGEM podcast for this week’s question. If you know, then send an email to thesgem@gmail.com with “keener” in the subject line. The first correct answer will receive a shoutout on the next episode.

SGEMHOP: Now it is your turn SGEMers. What do you think of this episode on AI for CDS? Post your comments on social media using #SGEMHOP.  What questions do you have for Hashim and his team, ask them on the SGEM blog. The best social media feedback will be published in AEM.

REMEMBER TO BE SKEPTICAL OF ANYTHING YOU LEARN, EVEN IF YOU HEARD IT ON THE SKEPTICS’ GUIDE TO EMERGENCY MEDICINE.

Other SGEM Episodes on AI

  • SGEM Xtra: Rock, Robot Rock – Ai For Clinical Research
  • SGEM#459: Domo Arigato Misuta Roboto – Using AI To Assess The Quality Of The Medical Literature
  • SGEM#460: Why Do I Feel Like, Somebody’s Watching Me – Chartwatch To Predict Clinical Deterioration