Investigating Use Cases for AI, IT Tools in Emergency Medical Services

DEVELOPING MACHINE LEARNING TO FLAG CARDIAC EVENTS

The University of Pittsburgh Medical Center (UPMC) EMS department is taking a different approach to optimizing its workflow and improving patient outcomes by building a machine-learning (ML) tool that uses electrocardiograms (ECG/EKG) to classify cardiac events.

Typically, EMS personnel and other medical staff use classification systems like the History, ECG, Age, Risk Factors and Troponin (HEART) score to risk stratify patients with chest pain, explained Christian Martin-Gill, MD, chief of the EMS division at UPMC. The HEART score, in addition to experienced clinician interpretation of ECGs, is considered the gold standard measure for evaluating potential cardiac events.

However, accurately stratifying and identifying patients experiencing a serious cardiac event, like a heart attack, can be challenging. Unclear ECGs are a major hurdle care teams must deal with when triaging chest pain patients and these can lead to care delays that may negatively impact patient outcomes.

Martin-Gill indicated that in addition to this issue, limitations in commonly used risk scores also present potential shortcomings in chest pain triage.

“If you think about clinical risk scores like the HEART score, it is largely based on patient history factors and then a couple of clinical factors, like a blood cardiac enzyme, and then our general interpretation of the EKG, but compounded with some past medical history or other clinical risk factors like age,” he stated. “And any of those types of risk factor scores are generally based on a handful of data points. If you think about people that develop those, they might evaluate a dozen or two dozen factors that might put somebody at risk of one diagnosis versus the other. Then, we develop these risk scores based on that handful of features that predispose somebody or are associated with that diagnosis.”

ML enables the analysis of hundreds of features on an ECG all at once. Martin-Gill underscored that these algorithms could process a vast amount of raw data for each patient based on ECG readings, providing a more comprehensive view of the patient's heart health.

The ML tool developed by Martin-Gill and his team, and externally validated in health systems outside of UPMC, can examine almost 700 features found within ECGs, which can help EMS teams identify conditions like cardiac ischemia or blockages in the blood vessels. The tool is designed to support human interpretation of ECGs, as algorithms can analyze and interpret a larger number of data features, including those that may not be observable with the naked eye.

“We do think of this as a tool not to replace a physician or a paramedic's interpretation of the 12-lead [ECG], but one of the next steps that we're doing is to develop a dashboard where we can put the information that the algorithm is picking up on,” Martin-Gill said.

He further noted that the tool is meant to help direct users to more closely examine parts of the ECG that are abnormal or important but subtle and, therefore, easy to miss.

In the coming years, the tool will be leveraged in partnership with the City of Pittsburgh Bureau of Emergency Medical Services.

Now that the ML tool has been developed and externally validated, Martin-Gill indicated that the team is interested in how the information it generates can best be used. In the continuum of care for chest patient triage, three groups of healthcare professionals need to be considered: paramedics in the field, emergency physicians or EMS medical command physicians providing online medical direction, and cardiologists responsible for advanced interventions.

As part of the partnership, Martin-Gill’s team is working on focus groups with these staff to inform a dashboard to facilitate improved decision-making and care coordination between EMS and in-hospital clinicians.

“With that dashboard, what you can envision is that a paramedic would do a 12-lead EKG and would send it off to a medical communication center, a facility where an emergency medicine or EMS physician is there that the paramedic is going to talk to on the radio or the phone,” he explained.

“[The ECG], in the process of being transmitted, will be interpreted by our machine-learning model and will be processed through this dashboard,” he continued. “That, then, is displayed to the physician on the other end of the line, who can look at the individual features [flagged by the ML] in addition to interpreting the 12-lead on their own.”

From there, clinicians can guide EMS to determine which hospital to transport the patient to or what medical interventions may be needed based on the likelihood of a cardiac event. These insights could also impact what in-hospital interventions the cardiologist would want to pursue for the patient.

Martin-Gill emphasized that tools like the ML algorithm and the dashboard have significant potential to improve EMS patient triage.

“We already do a great job at identifying patients when they present with chest pain or other atypical symptoms, in terms of who might have an event or heart attack. However, we still have lots of opportunities to identify folks that are not having the most obvious events, especially when it comes to EKG interpretation,” he said. “These algorithms really are the next step for us to be able to interpret those in an improved way and be able to identify more of these patients so that we can risk stratify them better and get earlier care for the patients that truly need it, and potentially also to identify those that are really at low risk for an event and prevent them from having to have very extensive testing or observation periods.”

LEVERAGING IT SUPPORT FOR FASTER EMS RESPONSE

When utilizing mobile incident alert and management software, first responders face a host of challenges, according to Thomas Craighton, coordinator of the Hardin County, Iowa Emergency Management Agency.

He explained that while these technologies are crucial in assisting EMS professionals, they run through cellular networks, which can be problematic.

“Our first issue is always that the responders decide they don't carry a pager, and they're relying on this technology,” he explained. “But because we know cellular technology has failure points and has choke points, that becomes a huge issue.”

However, cellular networks are the best option to back up EMS paging systems, as they provide additional information that can help first responders, like location. These networks, in combination with the end-to-end emergency response system IamResponding, allow EMS professionals to receive the location of an incident and map it on their phones.

Thus, first responders can navigate to the scene of an incident more efficiently, whether they are coming from their home or the station, as many of these individuals are volunteers, Craighton noted.

Another challenge lies in coordination and communication, in part because many emergency services organizations have a high volume of volunteer-based staff. When people are responding to an incident alert from their homes, determining which and how many personnel should go to the scene of the incident can be difficult.

An end-to-end emergency response system enables EMS team members to indicate their response status, which allows other staff to quickly assess how many responders are available and who is coming to the scene.

Craighton said that at some stations, first responders can more specifically denote their status, with options for ‘going to the station,’ ‘going to scene, please bring my gear,’ or ‘not responding.’

Once dispatchers put out the page to notify staff of an incident, these responses can be leveraged to understand how many team members are responding and what they need. If enough staff do not respond, the communications staff can then page other departments in real time.

Despite the benefits of leveraging a solution like IamResponding — which Craighton explained is an extremely common tool used by first responder organizations — keeping the system up, running, and stable creates additional hurdles for EMS teams, especially those that are small or mostly made up of volunteers.

Many of these organizations lack the time, staff, resources, and expertise to troubleshoot problems with the emergency response system they rely on. Hardin County has worked to combat this by partnering with IT support company MiPi Support, which provides assistance and training in setting up these systems for individual departments, making it easier for them to use the technology effectively.

Craighton stated that the partnership goes beyond typical IT support and has helped his organization significantly in addressing the challenges related to using mobile incident alert software technologies.

The company specializes in IT support for first responder organizations, meaning that they have a comprehensive understanding not only of these emergency response systems but also of how EMS organizations operate and what their needs are, he added.

The company also serves as a liaison between the county and IamResponding, helping determine whether the software supports certain features or functionalities. If a desired feature isn’t supported, the company can share that with IamResponding to help EMS organizations provide feedback more effectively.

In addition to providing support, the company offers software and small computers to help improve dispatch and provide ongoing monitoring of the emergency response system, Craighton explained.

Using the company’s software, Hardin County can bridge the dispatch communication gap created by relying on radios and pagers. The tool routes dispatches directly to first responders’ phones, which helps better alert personnel if they don’t have their pager or are out of range.

The county also leverages the company’s monitoring software to help flag and address potential issues before they lead to system downtime. If any issues are detected, the county’s MiPi representative reaches out to flag and help fix them, which Craighton stated significantly reduces downtime. This can help improve EMS efficiency and response times in an emergency.

Craighton highlighted that this IT support has demonstrated substantial value for his organization, noting that it has led to results he hadn’t seen at other points in his career.

“I've only been in this county for going on five years now, but in my previous county, we did not have that,” he said. “Then, I had to know and figure out issues with IamResponding on my own. Now, it's so much easier to just call [our representative], send him an email, whatever. And within an hour, or, at the most, half a day, things are fixed and back online and ready to go.”

Martin-Gill and Craighton's experiences underscore that use cases for AI and IT tools in EMS are diverse, but the COVID-19 pandemic highlighted critical gaps in emergency medicine that these solutions alone cannot fix. Prior to the widespread adoption of these tools by health systems and EMS organizations, stakeholders will need to identify and address ethical and legal hurdles while adhering to industry best practices.