The app accesses data from the Swedish Cardiopulmonary Resuscitation Register on tens of thousands of patient cases. The University of Gothenburg researchers have used an advanced form of machine learning to teach clinical prediction models to recognize various factors that have affected previous outcomes. The algorithms take into account numerous factors relating, for example, to the cardiac arrest, treatment provided, previous ill-health, medication, and socioeconomic status
.
It will be a few years before official recommendations for cardiac arrest are likely to include AI-based decision support, but doctors are free to use these prediction models and other new, evidence-based methods. The research group working on decision support for cardiac arrest is headed by Araz Rawshani, a researcher at the University’s Sahlgrenska Academy and resident physician in cardiology at Sahlgrenska University Hospital.
“Both I and several of my colleagues who treat emergency patients with cardiac arrest have already started using the prediction models as part of our process for deciding on the level of care. The answer from these tools often means we get confirmation of views we’ve already arrived at. Still, it helps us not to subject patients to painful treatment that is very unlikely to be of benefit to the patient, while saving care resources,” Rawshani says.
To date, the research group has published two decision support tools. One clinical prediction model, known as SCARS-1, is presented in The Lancet’s eBioMedicine journal. This model indicates whether a new patient case resembles other, previous cases where, 30 days after their cardiac arrest, patients had survived or died.
The model’s accuracy is unusually high. Based on the ten most significant factors alone, the model has a sensitivity of 95 percent and a specificity of 89 percent. The “AUC-ROC value” (ROC being the receiver operating characteristic curve for the model and AUC the area under the ROC curve) for this model is 0.97. The highest possible AUC-ROC value is 1.0 and the threshold for a clinically relevant model is 0.7.
This decision support was developed by Fredrik Hessulf, a doctoral student at Sahlgrenska Academy, University of Gothenburg, and anesthesiologist at Sahlgrenska University Hospital/Mölndal.
“This decision support is one of several pieces in a big puzzle: the doctor’s overall assessment of a patient. We have many different factors to consider in deciding whether to go ahead with cardiopulmonary resuscitation. It’s a highly demanding treatment that we should give only to patients who will benefit from it and be able, after their hospital stay, to lead a life of value to themselves,” Hessulf says.
This form of support is based on 393 factors affecting patients’ chances of surviving their cardiac arrest for 30 days after the event. The model’s high accuracy may be explained by the huge number of patient cases (roughly 55,000) on which the algorithm is based and the fact that ten of the nearly 400 factors have been found to impact heavily on survival. By far the most important factor was whether the heart regained a viable cardiac rhythm again after the patient’s admission to the emergency department.
The second decision support tool published has been presented in the journal Resuscitation. This tool is based on data from patients who survived their out-of-hospital cardiac arrest until they were discharged from hospital. The predictive models are based on 886 factors in 5098 patient cases from the Swedish Cardiopulmonary Resuscitation Register. This tool is partly aimed at helping doctors identify which patients are at risk of another cardiac arrest or death within a year of discharge from hospital following their cardiac arrest. It also aims to highlight which factors are important for long-term survival after cardiac arrest — an aspect of the subject area that has not been well studied.
“The accuracy of this tool is reasonably good. It can predict with about 70 percent reliability whether the patient will die, or will have had another cardiac arrest, within a year. Like Fredrik’s tool, this one has the advantage that just a few factors can predict outcome almost as well as the model with several hundred variables,” says Gustaf Hellsén, the research doctor who developed this decision support tool.
“We hope,” he continues, “to succeed in developing this prediction model, so as to enhance its precision. Today, it can already serve as support for doctors in identifying factors with an important bearing on survival among cardiac arrest patients who are to be discharged from hospital.”
MEDICA-tradefair.com; Source: University of Gothenburg
