INTRODUCTION
In emergency medicine (EM) critical care, patients may arrive clinically unstable, with limited information located in multiple medical records, and require critical interventions before their clinical course becomes clear. Decision-making is time-sensitive and patients' conditions may continually change in response to treatment.
Interest in artificial intelligence (AI) in critical care has grown with advances in machine learning, the availability of longitudinal data, and the encouraging performance of predictive models in retrospective intensive care unit (ICU) studies [
1,
2]. Much of this work has been conducted in the traditional ICU settings and has provided foundational insights. However, emergency critical care presents important challenges for many existing AI methods and needs systems to be evaluated and adapted for this environment. Understanding why AI implementation in EM critical care differs from that in other areas of medicine is important for effective AI integration in this setting.
CHALLENGES
AI models for early deterioration prediction often use continuous physiologic and laboratory data to predict clinical decline [
2,
3]. However, structured modelling of dynamic changes from acute interventions such as airway management, vasopressors initiation, fluid resuscitation, or procedural treatments, which occur frequently in emergency critical care, remains relatively limited in many current approaches.
Early predictions may need to change as the patient's condition continues to change over time [
4,
5]. A mortality or deterioration risk assessment performed prior to airway management may not accurately predict the patient's physiological status after intubation. Similarly, hemodynamic predictions made prior to initiating vasoactive support may not accurately represent the patient's cardiovascular physiology after treatment.
Therefore, AI tools designed for the EM critical care environment should incorporate time awareness and continuously update outputs as the patient's physiology and treatment continue to change. Instead of providing a single forecast, AI is more beneficial when tracking physiological trends over time, identifying changes in the patient’s clinical status, and signaling to the clinician when previous assessments may no longer apply [
4,
5].
Emergency critical care patients are often given working diagnoses early in their course to guide initial management. However, these diagnoses are frequently revised as the patient responds to treatment and when additional data and clinical patterns develop.
AI tools in the EM critical care environment should focus less on classifying diagnoses and more on recognizing and describing evolving patient states. This could aid clinicians in assessing how the patient's physiology is responding to therapy and how new risks are developing. Systems that can analyze and display multiple vital signs, laboratory trends, and ventilatory parameters over time may be better suited for caring for critically ill patients.
Uncertainty is also part of emergency critical care. Clinicians must develop plans of action with incomplete data and weigh the risk of potential complications against the risk of the patient's condition deteriorating. Some AI systems may report results that appear confident despite underlying uncertainty [
6]. When AI system outputs appear to be authoritative, clinicians may be reluctant to challenge them [
7,
8].
APPLICATIONS OF AI IN EM CRITICAL CARE
Despite the limitations, AI systems may hold value in the emergency critical care environment when aligned with clinical workflows and the dynamic physiology of the patient.
Continuous longitudinal physiologic monitoring
AI can help clinicians with analyzing information across multiple notes, prior records, and documented goals of care. It could also help monitor physiological trends over time rather than relying on isolated data points. Machine learning models are well-suited for identifying patterns in patients' continuous vital signs and combining data from multiple monitoring sources to identify early warning signs of patient decompensation [
1,
3,
9–
12]. In some cases, these systems generate alerts based on patterns of physiologic change that may not be readily apparent to clinicians, yet still accurately predict clinical decline [
9,
10]. The effective implementation of AI could help refocus monitoring environments to prioritize clinically significant alarms, reduce the frequency of false alarms, and provide earlier recognition of potentially life-threatening instability.
Instead of producing predictions independent of care, AI systems should be designed to consider the interventions already initiated. Accounting for treatment context (e.g., recent intubation or fluid administration) can increase the applicability of AI-generated output and decrease the likelihood of misleading output [
4,
5,
13].
Transition of care handoffs are important in emergency critical care and a recognized area of communication risk [
14–
16]. AI-generated summaries that show active issues, physiological changes, code status, and pending tasks can help with continuity of care. The key to successfully integrating AI-generated summaries into clinical workflows is to streamline the information, making it concise and focused on what is needed to support immediate clinical decisions.
INTEGRATING INTO THE ENVIRONMENT
For AI systems to be successfully integrated into the EM critical care environment, workflow integration, clinician trust, and education are important. AI tools that add extra steps to the workflow, require separate interfaces, or deliver excessive information will increase the cognitive burden on the clinician and defeat the purpose of implementing AI to reduce workload [
17,
18]. Clinicians' trust in AI systems will depend on the transparency of the AI system, continuous monitoring of its performance, and its responsiveness to feedback as clinical practices develop [
2,
17,
18]. Education on AI should be viewed as a clinical skill and include training in interpreting AI-generated output, understanding AI system limitations, and maintaining the ability to make independent judgments [
7,
18].
Data drift is the gradual difference between the data distributions a model was trained on and the data it is exposed to when applied in practice. In EM critical care settings, data drift may occur due to variations in patient populations, differing local clinical protocols, variability in documentation practices, and use of differing medical equipment among hospitals [
19,
20]. When models trained on data collected within a development environment begin to fail when they are applied at another clinical site or when clinical practices change over time, models that were effective in one environment may perform poorly in other environments [
21,
22]. Addressing data drift requires monitoring a model’s performance after deployment continuously, defining triggers for reevaluation or retraining, and engaging with local clinical champions who can provide a perspective on performance issues.
FUTURE DIRECTIONS
AI in medicine is moving towards the development of foundation models and multimodal systems that are capable of processing heterogeneous data streams. This evolution is also applicable to EM critical care. The ability of multimodal AI systems to concurrently process electrocardiogram data, imaging studies, vital sign data, laboratory results, and clinical documentation provides the capability for the construction of a more comprehensive and dynamic view of the critically ill patient [
23,
24]. Instead of relying upon a single data modality to identify patterns, multimodal AI systems can identify relationships between data modalities that are individually ambiguous, but collectively informative. The integration of multimodal AI systems into EM critical care departments will depend on the degree to which these systems have been validated in this specific clinical environment [
25,
26].
As AI systems continue to develop in complexity, so too does the need for transparency and explainability. High-stakes clinical environments require careful scrutiny of black-box models that offer recommendations without showing how the outputs were generated and why the clinician should act on them [
27,
28]. Explainable AI (XAI) methods, such as revealing which individual data elements contributed the most to the model’s output, can help clinicians understand how the model reached its recommendations [
27–
30]. Prioritizing the development of XAI methodologies is crucial as AI tools are developed and improved for EM critical care departments.
There are ethical and legal considerations that may influence the future development of AI. One ethical concern is algorithmic fairness, which aims to prevent AI systems from reinforcing existing inequities in healthcare. When AI systems are trained on historical data, they may reflect disparities. As a result, they may produce systematically different recommendations across racial, socioeconomic, and demographic subgroups [
31,
32]. Ensuring that AI systems operate fairly across a diverse range of patient populations will require explicit consideration of equity during development, validation, and deployment [
31,
32]. Another issue is determining accountability when an AI system is involved in an adverse outcome. For example, who is responsible for harm caused by an AI system [
33]? Other considerations include data privacy, obtaining informed consent for AI-assisted care, and the possibility of reidentification of deidentified training data. Therefore, multidisciplinary governance will be essential to the design and governance of AI systems.
CONCLUSION
The EM critical care setting has the potential to demonstrate both how AI can be used successfully in a medical context as well as the limits of AI in medicine. In order for AI systems to support the dynamic and intervention-rich environment of EM critical care, they must be able to adapt to changing conditions, aware of the time constraints involved in providing acute care, and compatible with the bedside workflow that is typical of the EM critical care setting. Without careful validation and governance, they risk adding to, rather than reducing, the cognitive burden of clinicians.
NOTES
-
Conflicts of interest
The authors have no conflicts of interest to declare.
-
Funding
The authors received no financial support for this study.
-
Data availability
Data sharing is not applicable as no new data were created or analyzed in this study.
REFERENCES
- 1. Pinsky MR, Bedoya A, Bihorac A, et al. Use of artificial intelligence in critical care: opportunities and obstacles. Crit Care 2024;28:113.
- 2. Oh W, Veshtaj M, Sawant A, et al. ORAKLE: Optimal Risk prediction for mAke30 in patients with sepsis associated AKI using deep LEarning. Crit Care 2025;29:212.
- 3. Hyland SL, Faltys M, Hüser M, et al. Early prediction of circulatory failure in the intensive care unit using machine learning. Nat Med 2020;26:364-73.
- 4. Acosta JN, Falcone GJ, Rajpurkar P, Topol EJ. Multimodal biomedical AI. Nat Med 2022;28:1773-84.
- 5. Tomašev N, Glorot X, Rae JW, et al. A clinically applicable approach to continuous prediction of future acute kidney injury. Nature 2019;572:116-9.
- 6. Lee P, Bubeck S, Petro J. Benefits, limits, and risks of GPT-4 as an AI chatbot for medicine. N Engl J Med 2023;388:1233-9.
- 7. Sutton RT, Pincock D, Baumgart DC, Sadowski DC, Fedorak RN, Kroeker KI. An overview of clinical decision support systems: benefits, risks, and strategies for success. NPJ Digit Med 2020;3:17.
- 8. Khera R, Simon MA, Ross JS. Automation bias and assistive AI: risk of harm from AI-driven clinical decision support. JAMA 2023;330:2255-7.
- 9. Chang H, Park JE, Lee D, et al. Development and validation of a transformer model-based early warning score for real-time prediction of adverse outcomes in the emergency department. Sci Rep 2025;15:23021.
- 10. Salehinejad H, Meehan AM, Rahman PA, Core MA, Borah BJ, Caraballo PJ. Novel machine learning model to improve performance of an early warning system in hospitalized patients: a retrospective multisite cross-validation study. EClinicalMedicine 2023;66:102312.
- 11. Nemati S, Holder A, Razmi F, Stanley MD, Clifford GD, Buchman TG. An interpretable machine learning model for accurate prediction of sepsis in the ICU. Crit Care Med 2018;46:547-53.
- 12. Zhang Y, Xu W, Yang P, Zhang A. Machine learning for the prediction of sepsis-related death: a systematic review and meta-analysis. BMC Med Inform Decis Mak 2023;23:283.
- 13. Buell KG, Spicer AB, Casey JD, et al. Individualized treatment effects of oxygen targets in mechanically ventilated critically ill adults. JAMA 2024;331:1195-204.
- 14. Starmer AJ, Spector ND, Srivastava R, et al. Changes in medical errors after implementation of a handoff program. N Engl J Med 2014;371:1803-12.
- 15. Hartman V, Zhang X, Poddar R, et al. Developing and evaluating large language model-generated emergency medicine handoff notes. JAMA Netw Open 2024;7:e2448723.
- 16. Song JW, Park J, Kim JH, You SC. Large language model assistant for emergency department discharge documentation. JAMA Netw Open 2025;8:e2538427.
- 17. Lukac PJ, Turner W, Vangala S, et al. Ambient AI scribes in clinical practice: a randomized trial. NEJM AI 2025;2:10.1056/aioa2501000.
- 18. Amann J, Blasimme A, Vayena E, Frey D, Madai VI. Explainability for artificial intelligence in healthcare: a multidisciplinary perspective. BMC Med Inform Decis Mak 2020;20:310.
- 19. Finlayson SG, Subbaswamy A, Singh K, et al. The clinician and dataset shift in artificial intelligence. N Engl J Med 2021;385:283-6.
- 20. Wong A, Otles E, Donnelly JP, et al. External validation of a widely implemented proprietary sepsis prediction model in hospitalized patients. JAMA Intern Med 2021;181:1065-70.
- 21. Kore A, Abbasi Bavil E, Subasri V, et al. Empirical data drift detection experiments on real-world medical imaging data. Nat Commun 2024;15:1887.
- 22. Subasri V, Krishnan A, Kore A, et al. Detecting and remediating harmful data shifts for the responsible deployment of clinical AI models. JAMA Netw Open 2025;8:e2513685.
- 23. Cecconi M, Greco M, Shickel B, et al. Implementing artificial intelligence in critical care medicine: a consensus of 22. Crit Care 2025;29:290.
- 24. Azarfar G, Naimimohasses S, Rambhatla S, et al. Responsible adoption of multimodal artificial intelligence in health care: promises and challenges. Lancet Digit Health 2025;7:100917.
- 25. Schouten D, Nicoletti G, Dille B, et al. Navigating the landscape of multimodal AI in medicine: a scoping review on technical challenges and clinical applications. Med Image Anal 2025;105:103621.
- 26. Kotula CA, Martin J, Carey KA, et al. Comparison of multimodal deep learning approaches for predicting clinical deterioration in ward patients: observational cohort study. J Med Internet Res 2025;27:e75340.
- 27. Abgrall G, Holder AL, Chelly Dagdia Z, Zeitouni K, Monnet X. Should AI models be explainable to clinicians? Crit Care 2024;28:301.
- 28. Okada Y, Ning Y, Ong ME. Explainable artificial intelligence in emergency medicine: an overview. Clin Exp Emerg Med 2023;10:354-62.
- 29. Rosenbacke R, Melhus Å, McKee M, Stuckler D. How explainable artificial intelligence can increase or decrease clinicians' trust in AI applications in health care: systematic review. JMIR AI 2024;3:e53207.
- 30. Abbas Q, Jeong W, Lee SW. Explainable AI in clinical decision support systems: a meta-analysis of methods, applications, and usability challenges. Healthcare (Basel) 2025;13:2154.
- 31. Hasanzadeh F, Josephson CB, Waters G, Adedinsewo D, Azizi Z, White JA. Bias recognition and mitigation strategies in artificial intelligence healthcare applications. NPJ Digit Med 2025;8:154.
- 32. Amiot F, Potier B. Artificial intelligence (AI) and emergency medicine: balancing opportunities and challenges. JMIR Med Inform 2025;13:e70903.
- 33. Mello MM, Guha N. Understanding liability risk from using health care artificial intelligence tools. N Engl J Med 2024;390:271-8.
