Artificial Intelligence: Refining STEMI Interventions

STEMI Interventions in 2019

Door-to-balloon time (D2B)-mandated ST-elevation myocardial infarction (STEMI) interventions improve short- and long-term outcomes in acute myocardial infarction (MI).¹ Improvements in the STEMI process, paired with meticulous reductions in D2B times, have greatly reduced STEMI mortality and as a result, cardiovascular disease is no longer our biggest killer. Improvement in the STEMI process requires compulsive teamwork from paramedics, emergency room personnel, and the cardiovascular laboratory.¹

Added to this group of stakeholders must be the patient, who requires continuous education regarding early recognition of symptoms and the local availability of D2B facilities. These 24/7, on-call STEMI squads, working across the breadth of our vast nation, have collaborated to save the lives of millions of patients. Their dedicated efforts are contributing to enormous improvements in cardiovascular health, and in saving billions of dollars resulting from early death and disability. As a society, we must remain vigilant and avoid becoming complacent in these efforts. Ongoing process improvements are critical, as is the collection of data and its thoughtful interpretation. Continuous medical education of the entire team is also vital, and it requires monitoring at the level of individual hospitals.

The STEMI process must seamlessly merge into the STEMI procedure. An efficient STEMI process will facilitate the STEMI procedure. An efficient STEMI process brings the STEMI patient to the cardiovascular laboratory more quickly, where it is almost always easier to intervene in a fresh lesion. System delays cause clinical and hemodynamic deterioration of the patient. Such situations are pandemic in developing countries, where patients can proceed to cardiogenic shock; they face very high mortality resulting from system delays as well as delays by the patient in recognizing symptoms and seeking urgent care. Another major deficit in developing countries is a lack of organized ambulance systems. Patients often end up reaching a facility late and often it has poor resources for the treatment of acute MI. Thrombolysis is relatively ineffective in these situations, due to delayed presentation and, although thrombolytics are often administered, their efficacy is greatly reduced. From a pathophysiological standpoint, the STEMI lesion in patients presenting where the systems are inefficient will also more often demonstrate a dense, organized thrombus that is more challenging to treat and whose management is suboptimal.

We are fortunate that in the United States, system-wide delays are an exception, and once the patient has called 911, our excellent systems take over and, in the majority of cases, the outcomes are very good. We have coined the statement, “the fate of a patient with a STEMI depends on your zip code” to illustrate how much variability exists, in particular, between developed and developing countries.

In order to facilitate newer innovations to further improve STEMI performance in the U.S., emergency department (ED) bypass should be incorporated into the STEMI process as much as possible. However, it needs to be done with careful triage, specifically for identification of the acute MI patient that benefits from evaluation in the ED. When performed effectively, intelligent ED bypass can be as effective in reducing D2B time as prehospital STEMI management.

As existing STEMI processes continue to be improved, artificial intelligence (AI) can be incorporated into the various components of the STEMI process and procedure. In this manner, we can exploit a unique aspect of STEMI: its diagnosis can be made with a single, accurate electrocardiogram (EKG), a simple step that can automatically trigger the STEMI process.^2,3

What Does AI Mean in the Era of the Apple Watch?

A brief understanding of AI may be useful at this stage. We are now ushering a new era of discovery, an era that emphasizes the creation of intelligent systems that seamlessly merge all previous discoveries: simplifying them, cutting out the small menial tasks, and thus increasing our already outstanding productivity. We are in an era that blurs the lines between the physical, digital, and biological spheres. Determination, curiosity, a thirst for knowledge: whatever one may call it, we are at a dawn of the fourth industrial revolution.

Figure 1 illustrates some of the suggested definitions of AI. Although these various terminologies describe some different and unique aspects of AI, our favorite glossary is the one used by Amazon: “AI is a broad and complex branch of computer science whose ultimate finality is the creation of computerized systems that can function intelligently and independently of humans, mimicking as much as possible the complexity of human thought process.” Integrated into AI is the concept of Machine Learning, in which teaching is synonymous with feeding large quantities of data to an intelligent system, whether images, results, patterns, or even trends. Aided by an impressive storage capability, the intelligent systems gradually deconstruct the data they receive to its most basic components; allowing the system to identify patterns within different, previously unrelated fields in a shorter period of time.^4-6

Of course, AI is not only confined to in-hospital settings. The popularization of AI is already happening: AI is going mainstream, as evidenced by the recent release of the Apple Watch Series 4, and the latest news of Amazon Comprehend Medical, described by Amazon as “a natural language processing service that makes it easy to use machine learning to extract relevant medical information from unstructured text.” Consider the numerous portals created for the fast access to medical records, those same portals that accurately remind each patient of their follow-ups, to refill and take their medications, even going so far as to remind them of their vaccines. These are baby steps of transition our industry is facing. Is that the limit? No. As flawless as a transition may be, it still entails a price. It creates reasonable doubts and fears, materialized after years of constant contact with popular science fiction that has, in a detrimental, albeit unintentional manner, cemented a subconscious aversion towards progress; a “hell scenario”, if you will. The reality is far from such fears. Though it may be true that, like any other business, healthcare may see some degree of automatization, it does not spell the end. AI is meant to be a tool. A tool that not only propels human evolution, but evolves alongside it. This mutual evolution will have its bumps along the way, but there is nothing that limitless human adaptability and creativity cannot overcome. A full realization of the impact of incorporating AI is the absolute key to success.^7-9

As with any new advancement in technology, some problems may surface: the requirement of acquiring yet another set of skill to manipulate AI systems accordingly is a must, some loss of human contact between physician and patient, shifting paradigms of certain specialties, regulatory and legislation reimbursement issues — all will be present in the short run. Yet the benefits of AI are manifold: we can assign AI systems specific roles that will improve the critical thinking of physicians and reduce the number of medical errors committed. AI can allow us to focus on the human aspect of the industry. AI can be designed to fulfill niche roles that will help with the prompt identification of treatable diseases at an earlier stage, allowing for adequate treatment options. Consider AI an augmentation of previously honed skills, optimizing healthcare practice.

As one example, Figure 2 depicts the progress of EKG, tracking its course from the successful development of Einthoven’s Electrocardiograph in 1901 to the incorporation of the Apple Watch. As one comprehends this remarkable journey, it is easy to appreciate how AI is the next logical step in STEMI detection.

As we noted above, AI tools for diagnosis are being incorporated to wearable devices. The most prominent example is the Apple Watch 4, backed by the American Heart Association and with FDA clearance. This wearable device can detect atrial fibrillation.¹⁰ Via photoplethysmography, the device traces changes in blood flow and pulse. This information is translated into a tachogram that is processed via an AI-powered algorithm. If an irregular pulse is detected, the algorithm can detect atrial fibrillation with a positive predictive value of 84%. The user is notified to seek medical attention via an application installed in their iPhone. Once atrial fibrillation is noted, the app immediately offers physician consultation. This is a clear example of how technology can bring us closer to the patient and improve screening, diagnosis, and at-home patient monitoring. Another very useful application built into the Apple Watch 4 is its ability to sense when the user falls, bringing onscreen the option to call 911 in case the patient needs help. the Apple Watch 4 can be helpful for the 2.7-6.1 million patients with atrial fibrillation¹¹ and could become a must-have item for the elderly in particular, since an estimated 3 million elderly patients are seen every year at the emergency departments.¹²

Applying AI Broadly to STEMI Interventions

STEMI interventions are unique in that the entire process is enormously time-dependent and the outcomes are linearly correlated with the system efficiency. AI can be incorporated into the various specific components of both the STEMI process and the procedure. It can foster large efficiency gains by simplifying each of the various components, be that at the individual patient, paramedic, ED, or the cath lab. Figures 3 and 4 illustrate the far-reaching manner in which the process of STEMI intervention can benefit through incorporation of AI. As we demonstrate various ways to incorporate AI into the STEMI pathway, two important aspects need to be discussed that will help to evaluate the impact of AI in STEMI interventions. First, AI augmentations can occur within the existing pathways that remain essentially the same. As a result, the individual role of the various stakeholders remains precisely the same. However, each team performs better, and with improved communication and efficiency. Second, we hypothesize that the AI enhancements lead to cost reductions at the individual stage of STEMI. The gains from AI-directed STEMI pathways are, as such, twofold: through the lowering of D2B times and by reducing cost.

Specific STEMI Enhancements

Figures 5 and 6 demonstrate niche applications of AI to the specific components of the STEMI process. In Figure 5, the standard components of the STEMI process are depicted. To these individual components, from procuring the EKG to achieving ED bypass, each step is made more efficient by incorporation of AI. As an example, cloud computing, GPS navigation, and machine learning algorithms strengthen the communication and the pre-hospital process, significantly increasing ED bypass. Single-page activation, an essential component of the STEMI pathway, becomes real-time with AI, as is the communication between the entire team.

The often complicated and redundant process required for a STEMI confirmation is altogether eliminated. Figures 5 and 6 deconstruct this mundane step and how it is eliminated via AI. Currently, in our existing STEMI systems, the paramedics contact the ED physician, who then tracks down the interventional cardiologist. Along the way, the EKG findings are first communicated along a cumbersome chain, first by the paramedic to the ED physician. This is done either via an impractical phone conversation, by digital transmission of the EKG, or by relying on the other “modern-day” technology of using a fax. We incorrectly assume that these communications occur accurately and efficiently, whereas the reality is far different. This process is flawed from an inherent deficiency: the communication between paramedics and the ED relies on the knowledge of the paramedic and their capability for making an accurate EKG diagnosis.

Let us dive deep into the specifics of verifying a STEMI diagnosis. Two models are presently practiced in the United States. The first model features better-trained paramedics relaying the EKG findings to the ED physician, a scenario perfected in Ottawa, Canada, by Dr. Michel LeMay. With this pathway, the well-trained advanced paramedic relays the EKG findings and diagnosis to the ED physician. In the United States, it has been difficult to achieve a wide dissemination of such highly trained, advanced paramedics, and in its most truthful version in the U.S., paramedics are reading off the computer-generated EKG diagnosis. In the second model, where the paramedics either use a fax or the Internet, this task diverts from the more urgent responsibility of monitoring a critical patient who may be unstable and who may need to be cardioverted, intubated, or both. The paramedic’s focus and duties are therefore divided, as they also need to quickly review the EKG with the ED physician and initiate a single-page activation to alert the STEMI team.

With cloud computing and innovative techniques of EKG acquisition, AI enables a complete deconstruct of potential stumbling blocks. Figure 6 depicts the modifications that will occur with the application of AI and how the numerous steps discussed above are altogether eliminated. We have adapted the existing STEMI systems to these individual steps, but it is easy to comprehend how these components can be redundant and time-consuming. In addition, each component is prone to miscommunication and resulting medical errors.^13-16

When rushed decisions are being made simultaneously by different teams, the inherent flaws in our present STEMI systems make the entire system cumbersome, expensive, and fraught with possibilities of errors.

STEMI chaos is maximal, however, at the hospital, where there is even more miscommunication and delay in a definitive diagnosis of STEMI. In the present models, the ED physician communicates with the interventional cardiologist on call. This particular process has enormous efficiency variations in individual hospitals, as response times for the interventional cardiologists vary greatly. As it is, the on-call, non-hospital-based interventional cardiologist must disengage immediately from his or her professional or personal activities to attend to the STEMI patient. Although systems have become more competent, there are still lag times in this particular methodology. AI improves the communicative step between the paramedics and the ED physician. The AI-boosted pathway in Figure 6 depicts the larger efficiency gain potentially occurring during communications between the ED physician and the interventional cardiologist.

AI enables cloud computing to completely eliminate the superfluous steps between the paramedic —> ED physician —> interventional cardiologist conduits. In fact, it may put an end to all these conversations. With cloud computing and machine learning, the entire STEMI diagnosis, activation, and alert process occurs almost at the point of care, even before the transportation of the patient into the ambulance. The transfer from a non-STEMI to a STEMI facility can also occur more efficiently and with fewer medical errors.

Various aspects of the STEMI procedure can be enhanced with machine learning-based algorithms and with the use of robotics that can make the procedure simpler and more predictable. Critical STEMI procedural decisions, such as selecting the access site, determining the need for thrombectomy, stent sizing, augmenting TIMI flow and myocardial perfusion grades, left ventricular support, and management of complications, are already being researched through AI-guided techniques of machine learning-based algorithms. We strongly believe that some or all of these innovations will permeate the existing STEMI procedure. Although less relevant for STEMI, the application of appropriate use criteria (AUC) is a straightforward application of machine learning. Finally, in-hospital management, administrative, and discharge systems will also be restructured with AI, with particular benefits accruing from the reduction of medical errors, waste, and fraud.^13-15

Moving Forward

Similar to the development that occurred with the initial D2B processes, there will be early adopters of AI technology who will pick and choose individual steps based upon their particular needs and deficiencies. Ambulance services may augment their systems to become “Uber AI” or they may customize some of the existing steps. There may or may not be the need for 510(k) abbreviated FDA approvals of the AI-based STEMI systems. However, even if an approval is required, it will probably be considerably faster as a result of a renewed vigor at the FDA for approving digital technologies. Quality assurance, however, will be required in a most deliberate manner. We also believe that larger systems, like the American Heart Association’s Mission:Lifeline and European Society of Cardiology’s Stent for Life initiative, will evolve to support AI systems. In doing so, they may be able to aggregate major data, demonstrate the validity of the new systems with the use of AI, and evaluate their cost-effectiveness. The accumulated data will automatically improve supervised and unsupervised machine learning algorithms and continually enhance the STEMI processes.¹⁶ AI gains are, of course, demonstrable in other areas of cardiology. In fact, use of AI is flourishing worldwide (Figure 7), as various applications of AI systems demonstrate the potential to improve management of cardiovascular disease. 

The authors can be contacted via Sameer Mehta, MD, at sameer.lumenglobal@gmail.com.

1. McNamara RL, Wang Y, Herrin J, et al. Effect of door-to-balloon time on mortality in patients with ST segment elevation myocardial infarction. J Am Coll Cardiol. 2006; 47:2180-2186.
2. Johansson BW. A history of the electrocardiogram. Dan Medicinhist Arbog. 2001; 163-176.
3. Meissner A, Trappe HJ, de Boer MJ, et al. The value of the ECG for decision making at first medical contact in the patient with acute chest pain. Netherlands Heart J. 2010; 18: 301-306.
4. Bush J. How AI is taking the scut work out of health care. Harvard Business Review. 2018. Available online at https://hbr.org/2018/03/how-ai-is-taking-the-scut-work-out-of-health-care. Accessed April 12, 2019.
5. Acharya UR, Fujita H, Oh SL, et al. Application of deep convolutional neural network for automated detection of myocardial infarction using ECG signals. Information Sciences. 2017 Nov;415-416: 190-198.
6. Liu W, Zhang M, Zhang Y, et al. Real-time multilead convolutional neural network for myocardial infarction detection. IEEE J Biomed Health Inform. 2018 Sep; 22(5): 1434-1444.
7. Banks J. The human touch: practical and ethical implications of putting AI and robotics to work for patients. IEEE Pulse. 2018 May-Jun; 9(3): 15-18.
8. Khader S, Johnson KW, Glicksberg BS, et al. Machine learning in cardiovascular medicine: are we there yet? Heart. 2018; 104: 1156-1164.
9. Nsoesie EO. Evaluating artificial intelligence applications in clinical settings. JAMA Network Open. 2018; 1(5): e182658.
10. Kipp WJ, Torres J, Glicksberg BJ, et al. Artificial intelligence in cardiology. J Am Coll Cardiol. 2018 June; 71(23): 2668-2679.
11. Garvey JL, Zegre-Hemsey J, Gregg R, Studnek JR. Electrocardiographic diagnosis of ST segment elevation myocardial infarction: an evaluation of three automated interpretation algorithms. J Electrocardiol. 2016; 49: 728-732.
12. Krittanawong C, Zhang H, Wang Z, et al. Artificial intelligence in precision cardiovascular medicine. J Am Coll Cardiol. 2017; 69: 2657-2664.
13. Kalis B, Collier M, Fu R. 10 promising AI applications in health care. Harvard Business Review. 2018 May. Available online at https://hbr.org/2018/05/10-promising-ai-applications-in-health-care. Accessed April 12, 2019.