In what they describe as the first study to detail a deep-learning model for emergency-call classification, Spanish researchers have developed a series of neural networks that could one day be used to more efficiently label and categorize emergency calls, potentially leading to faster response times for people suffering through severe medical emergencies.
If implemented in a real-world scenario, the system, which was detailed in a May 13 Artificial Intelligence in Medicine paper, may be able to provide much-needed support to emergency operators, who often must run through a series of time-consuming checklists to determine the appropriate level of response to a particular emergency. With the help of the new system, operators could instead shift their attention to patients themselves, extracting relevant information for the neural network to automatically process and sort. The network judges the urgency of an emergency by assessing the severity of a patient's symptoms as well as the time frame within which an issue needs to be addressed.
The innovation is being developed by a team of researchers from Universitat Politècnica de València and the government of the autonomous Spanish community of València. The researchers tested their system with data from València's emergency medical dispatch center, including over 1.2 million emergency calls that took place between 2009 and 2012, but have yet to implement the neural network in a real-time emergency situation.
Antonio Félix de Castro, a medical doctor who has in recent years turned his attention to reforming Spain's emergency response system, is a co-author of the study. His previous work includes testing a new electronic health-record system on ambulances that would allow first responders easy access to patients' medical histories. Prior to that, Félix responded to emergencies himself as part of an ambulance crew. He observed how the emergency system would sometimes become overrun with calls, which tended to peak once in the morning and once in the evening.
"In every part of the world, it works in the same way," Félix told The Academic Times. "Sometimes in our dispatch center, we are managing maybe 100 or 200 cases at one moment." The fast-paced environment can sometimes lead operators, usually through no fault of their own, to inadvertently place patients' calls in the wrong category.
The mislabeling of emergencies as nonurgent events can critically delay care for those who need it most. But just as troubling, Félix said, is the opposite scenario. When a minor call is incorrectly labeled as an emergency, it means that people in life-threatening situations may miss out on immediate care, given the limited number of doctors, emergency technicians and ambulances available at a given time.
"If I commit a mistake, and I say that [an emergency] is more severe than it is, it's not a problem for the patient," Félix said. "But it's a problem for the system, because you are using a resource that will not be available for another case."
The precision of emergency-call classification also impacts the public's view of emergency health care: If ambulances outfitted with doctors respond to every call with great urgency — even those for minor medical events — it may indirectly teach citizens that they should expect a highly trained professional to respond to them in moments where an emergency technician would be equally qualified for the job. Meanwhile, a lack of a response in the midst of a real emergency may erode public trust in the emergency response system, making people skeptical that it is capable of properly treating patients.
The new model implements four separate neural networks: a contextual network, which considers the demographic and circumstantial factors relating to an emergency call; a clinical network, which analyzes patients' symptoms and prior health issues; a text-based network, which interprets dispatchers' free-text written observations; and a final network that integrates and weighs each of those inputs to come up with a final overview of a medical emergency.
In macro-F1 measures — which account for accuracy, recall and precision — the system was 17.5% more effective than the existing triage protocol at identifying an appropriate response window for emergencies. It was also 12.5% more effective at determining whether an emergency was life-threatening and 5.1% better at recommending either emergency care providers or nonemergency, primary care personnel.
But the process is complicated by the unpredictable mixture of symptoms that can accompany disorders and medical issues with vastly different outcomes. A gallstone, for instance, can cause incredible pain and may easily be misinterpreted as a life-threatening medical event, although its real potential for danger is very low. And even intense anxiety or panic may be easily misdiagnosed as something far more severe, such as a heart attack. When the system is adapted to a real-world context, it could also include a safety protocol for unclear cases such as those. In instances in which the algorithm cannot decide a clear-cut category for a particular call, it will pass information along to a medical coordinator who can review the case in more detail.
The government in Spain has borrowed a French model of emergency medical care, according to Félix, in which a group of fully equipped doctors and nurses respond to the most urgent emergencies. A larger group of emergency technicians, similar to paramedics in the U.S., responds to less urgent calls. Whereas in the United States or some other countries the goal may be to transport someone to a hospital in a short time for immediate care, Spanish doctors tend to treat patients with equipment from inside the ambulance itself. For instance, medical personnel have access to advanced devices used to treat heart attacks, limiting the need to rush patients to a medical facility.
In València, emergency operators are expected to collect all the data needed to classify an emergency in around two minutes. This can prove to be a real challenge, especially when the person on the other end of a call is going through a traumatic medical event. The researchers hope their model will streamline the process.
"We want to let the system recognize the [caller's] voice and transcribe it immediately, without any clicks of the operator," Félix said. "The operator just needs to be trained to perform a good interview" — with the algorithm working behind the scenes to classify the call.
But although that kind of seamless system is the ultimate vision, putting the model into the hands of emergency operators will likely be a slow, methodical process. "We're working on the first steps of implementation of artificial intelligence in medicine," Félix said. "The next step that we need to do is implement this into the workflow of the call-takers."
The study "Deep ensemble multitask classification of emergency medical call incidents combining multimodal data improves emergency medical dispatch," published May 13 in Artificial Intelligence in Medicine, was authored by Pablo Ferri, Carlos Sáez, Javier Juan-Albarracín, Vicent Blanes-Selva and Juan M.García-Gómez, Universitat Politècnica de València; and Antonio Félix de Castro and Purificación Sánchez-Cuesta, Generalitat Valenciana.