Artificial Intelligence in the Cardiac Intensive Care Unit: Current and Future Application in Mechanical Circulatory Support

Sarah Grebennikov, DO
Department of Internal Medicine, Riverside Methodist Hospital, Columbus, OH

Abstract

Artificial intelligence (AI) is becoming increasingly integrated into clinical practice, with promise in high-acuity environments such as the cardiac intensive care unit (CICU). Critically ill patients requiring mechanical circulatory support (MCS) present unique challenges that demand continuous monitoring, timely decision-making, and interpretation of complex physiologic data. AI-enabled tools have the potential to assist in prognostication, device management, and optimization of weaning strategies.¹ This review explores the current and emerging applications of AI in the CICU, with a focus on patients requiring MCS, and identifies associated barriers to implementation, including data quality, ethical concerns, and the need for clinical validation.

Introduction

The rise of artificial intelligence (AI) across industries has been exponential in recent years, and healthcare is no exception. While AI offers the potential to revolutionize patient care, its integration also brings ethical, practical, and clinical challenges. Within cardiovascular medicine, AI has already been applied to risk prediction, diagnostics, and therapeutic decision-making. Its utility is especially promising in the care of critically ill patients. The sickest population often require MCS in the CICU.

Cardiogenic shock, a state of end-organ hypoperfusion due to inadequate cardiac output, is commonly seen as a complication of acute myocardial infarction.² Management of CICU patients is uniquely challenging due to the dynamic and often rapidly deteriorating nature of their clinical course. In patients that require MCS, including devices such as venoarterial extracorporeal membrane oxygenation (VA-ECMO), Impella, and left ventricular assist devices (LVAD), there is a need for continuous, high-fidelity monitoring, and timely interventions. This high-acuity environment presents an ideal opportunity for AI to support clinical decision-making and potentially improve outcomes. AI-enabled tools have the potential to assist in prognostication, device management, and optimization of weaning strategies.² This review explores current and emerging applications of AI in the CICU, with a focus on patients requiring MCS, and discusses associated barriers to implementation, including data quality and ethical concerns.

Overview of AI in Critical Care

AI exists on a spectrum that ranges from basic statistical models, such as logistic regression, to more advanced machine learning (ML) and deep learning (DL) techniques. ML models take information and learn from large, structured datasets and update predictions as new data is introduced. DL, a subset of ML, utilizes neural networks designed to mimic the way the human brain processes complex data inputs. These more complex systems are particularly suited to manage dynamic clinical data, such as labs, waveforms, imaging, and device parameters.¹

1. Prognosis of Patients on Mechanical Circulatory Support
   Accurate prognostication is critical in guiding the management of patients on MCS. Traditionally, clinicians have relied on scoring tools, such as the SAVE (Survival After Veno-Arterial ECMO) score, to estimate survival likelihood in patients on VA-ECMO.³ However, these tools are limited by static variables and are not reflective of dynamic and evolving clinical status.
   
   Recent machine learning approaches have demonstrated superior performance by integrating real-time hemodynamic and laboratory data. This includes lactic acid clearance, trends in renal function, and hemodynamic parameters which cumulatively help to dynamically estimate survival. For example, the ECMO Predictive Algorithm (ECMO PAL), a deep learning model trained on over 18,000 patients in the international Extracorporeal Life Support Organization (ELSO) registry, significantly outperformed the SAVE score in predicting in-hospital mortality.⁴
   
   Another ML model has also shown promise in predicting right ventricular failure (RVF) after LVAD implantation. A video-based deep learning model, developed from a post hoc analysis of the ENDURANCE trial, was able to stratify risk of developing RVF.⁵ Further ML investigation led to the identification of phenotypes of LVAD patients which identify those at highest risk of adverse events.⁶ These tools can serve to adjunct clinical judgment and improve physician-led decision-making regarding escalation, de-escalation, or goals-of-care.
   
   
2. Device Monitoring and Complications
   Recent innovations have explored AI-assisted monitoring and adjustment of MCS devices. Abiomed’s Impella SmartAssist Technology, for example, integrates real-time hemodynamic monitoring and assists with device placement and repositioning.⁷ Although this device is FDA-approved and currently in use within the United States, there are no direct comparative studies with non-AI-supported devices.
   
   Another burden in the CICU is the continuous adjustment and optimization of MCS devices. Future applications may include fully automated systems that dynamically adjust device parameters based on perfusion targets, such as mean arterial pressure or mixed venous oxygen saturation.⁸ AI-based suction detection algorithms have already been tested in Impella and LVAD systems, allowing for more precise control and earlier identification of complications, with future potential to automatically respond to suction alarm.⁹ These systems have the potential to offload some of the manual titration demands from clinicians while maintaining precise physiologic control.
   
   In addition to device optimization, AI may play a role in early complication detection. One study demonstrated that a smartphone-based application (InDetector) utilizing AI could identify LVAD driveline infections with 93.7% accuracy, 100% sensitivity, and 85.7% specificity.¹⁰ Such tools may assist in early intervention and reduce infectious morbidity in this high-risk population.
   
   
3. AI-Guided Weaning from MCS
   Determining appropriate timing of weaning from MCS is a pivotal decision in a patient’s clinical course. Premature device explantation can lead to hemodynamic collapse, while prolonged support increases the risk of device-related complications. Traditional weaning assessments involve synthesis of multiple data points, including echocardiographic findings, arterial waveform trends, and multidisciplinary discussion. To assist with determining whether weaning is appropriate, a ML model has been evaluated in those patients with left ventricular assist devices (LVAD), detecting and predicting myocardial recovery, which then helps with decision making for LVAD explantation.¹¹ This would then, in turn, decrease morbidity associated with early and delayed weaning.
   
   
4. Barriers to Implementation
   Despite the considerable potential of AI, there are barriers that remain, preventing widespread adoption and implementation of AI into healthcare. One of the major concerns is data quality, as AI algorithms are highly dependent on the integrity of their training datasets. Incomplete, biased, or institution-specific data can lead to erroneous predictions.¹² Another limitation is the difficulty of conducting randomized controlled trials (RCTs) to validate AI algorithms in real-world clinical settings. The variation of electronic health record systems across institutions also makes standardization and generalizability of AI models challenging.¹³ Furthermore, privacy concerns related to the use of protected health information require strict de-identification protocols, and clinicians as well as governing bodies should be mindful of establishing regulations to ensure safety, accountability, and transparency in AI applications.

Conclusion

AI has the potential to transform care in the CICU, particularly for patients requiring MCS. By continuing to improve prognostication, automating device management, and optimizing weaning strategies, AI can support clinicians in delivering more personalized and effective care. Importantly, AI should be viewed as a tool that supports clinical decision making instead of replacing it. As technological advancements continue and validation studies expand, the future of AI in the CICU holds promise for improving outcomes, provided implementation remains ethically grounded, data-informed, and clinician-led.

 References

1. Bajwa J, Munir S, Nori A, et al. Artificial intelligence in healthcare: transforming the practice of medicine. Future Healthc J. 2021;8(2):e188-e194.
2. Waksman R, Kapur NK. Cardiogenic shock: definitions and diagnosis. Circulation. 2020;141(18):1434-1448.
3. Schmidt M, Burrell A, Roberts L, et al. Predicting survival after VA-ECMO for refractory cardiogenic shock: the SAVE-score. Eur Heart J. 2015;36(33):2246-2256.
4. Seitz KP, Abramson TM, Arora S, et al. Machine learning-based prediction of mortality in VA-ECMO: the ECMO PAL model. Intensive Care Med. 2023;49(4):509-520.
5. Shad R, Celi LA, Mark RG, et al. Predicting post-operative right ventricular failure using video-based deep learning. Nat Commun. 2021;12:5192.
6. Kilic A, Goyal A, Miller JK, et al. Machine learning approaches to analyzing adverse events following durable LVAD implantation. Ann Thorac Surg. 2021;112(3):770-777.
7. Consolo F, Attisani M, Panoulas V, et al. The intelligent Impella: future perspectives of artificial intelligence in the setting of Impella support. ESC Heart Fail. Published online 2024. doi:10.1002/ehf2.14865
8. Saugel B, Thiele RH, Teboul JL. Tracking cardiac output with AI-enabled waveform analysis. Br J Anaesth. 2020;125(2):252-254.
9. Maw M, Kapur S, Bishop M, et al. Development of a suction detection algorithm from patient pump data. Artif Organs. 2017;41(7):678-686.
10. Skaria R, Saha A, Fuery M, et al. Artificial intelligence improves left ventricular assist device driveline infection detection. J Am Coll Surg. 2018;227(4 Suppl):e85.
11. Topkara VK, Sayer G, Kurlansky P, et al. Machine learning-based prediction of myocardial recovery in patients with left ventricular assist device support. Circ Heart Fail. 2022;15(1):e008711.
12. Miller PE, Birnbaum G, Iliescu C, et al. Predictive abilities of machine learning techniques may be limited by dataset characteristics: insights from the UNOS database. J Card Fail. 2019;25(6):479-483.
13. Kelly CJ, Karthikesalingam A, Suleyman M, et al. Key challenges for delivering clinical impact with artificial intelligence. BMC Med. 2019;17:195.