NIH-funded researchers have created an electrocardiogram-analyzing deep neural network that showed comparable outcomes to cardiologists' interpretations in a study.
The network correctly classified a range of distinct arrhythmias, leading the researchers to talk up the prospect of it being used to better triage patients.
However, the network is yet to be validated in clinical settings and could suffer from problems including false positives unless it is tailored.
Computers have come to play a central role in the interpretation of EKGs since they were introduced into the field decades ago. Many of the more than 300 million EKGs captured globally every year are automatically analyzed. The algorithms that perform these analyses are improving but remain prone to diagnostic mistakes; if a physician fails to spot the mistake, the patient may be mismanaged.
Technological advances are creating opportunities to reduce the rate of diagnostic mistakes. With devices such as AliveCor and Apple's wrist-worn sensors and iRhythm Technologies' wireless patches capable of capturing EKGs, cardiovascular data is digitizing and proliferating at a time when the rise of deep neural networks, or machine learning, is creating new opportunities to automate analysis of health information.
Recognizing the opportunity, NIH-funded researchers from Stanford University and other organizations trained a neural network to classify 12 rhythm classes, including 10 arrhythmias, using single-lead EKG data captured by iRhythm’s Zio patch from more than 50,000 patients.
When tested on a distinct dataset covering 328 patients, the neural network performed comparably to a committee of cardiologists. The researchers then applied the neural network to external data to assess the ability to generalize its performance. The neural network performed comparably to algorithms developed to analyze the external data.
The researchers think the results suggest the algorithm could improve cardiovascular workflows by accurately prioritizing patients with the most urgent conditions. However, the neural network will need tailoring to target clinical applications before being used in practice to address potential issues including “nontrivial false-positive diagnoses.”
"The findings suggest that artificial intelligence can be used to improve the accuracy and efficiency of EKG readings," National Institutes of Health Director Francis Collins wrote in a blog Tuesday. "As impressive as this is, we are surely just at the beginning of AI applications to health and health care."
While those concerns suggest there is scope for improvements and refinements, iRhythm has already incorporated the algorithm into the process it uses to analyze data from patients who use its Zio patch.