A new AI software can quickly and accurately determine the manufacturer and model of a cardiac rhythm device from an x-ray, possibly speeding up treatment when the devices fail.
More than 1 million people are implanted with pacemakers, loop recorders or defibrillators worldwide each year, but patients can deteriorate quickly during device failure, lead author James P. Howard, MB BChir, and colleagues wrote in JACC: Clinical Electrophysiology. When the patient is unable to share the model of the device or clinicians don’t have access to records from the implanting hospital, cardiologists often compare an x-ray image to a flow chart to try to determine what manufacturer made the device. Only then can the device’s programming be adjusted accordingly.
But Howard et al. found a convolutional neural network (CNN) was significantly more accurate than cardiologists in identifying images of 1,676 devices, including 45 different models from five manufacturers.
The CNN singled out the manufacturer from the radiograph with 99.6 percent accuracy and was 96.4 percent accurate in identifying the model group. Five cardiologists—including two electrophysiologists—ranged from 62.3 percent to 88.9 percent accuracy in identifying the manufacturer, and model group identification wasn’t possible with their flow-chart algorithm. The electrophysiologists were the most accurate, but the next-best physician achieved 72 percent accuracy in selecting the manufacturer.
Howard et al. said their model could be clinically useful, particularly once further studies assess and validate its real-world accuracy. The system is currently available online as an educational tool, which physicians can interact with and upload images to.
“Pacemaker programmers are portable but bulky, and only the manufacturer, the specific programmer, would be able to communicate with the patient’s device,” Howard and co-authors wrote. “Knowing which programmer to bring saves valuable clinical time. Not only may this facilitate rapid interrogation of a device in an emergency, but also the provision of emergency treatment, such as the delivery of anti-tachycardia pacing in a patient presenting with ventricular tachycardia.”
The researchers said there was only one case in which the CNN didn’t accurately identify a device’s manufacturer. There were more instances in which the specific model wasn’t singled out, but the CNN provided the top three predictions for each image—and it included the correct model in the top three choices 99.6 percent of the time.
Howard et al. said using the algorithm simply requires uploading the x-ray image of the device to a computer equipped with the software, and a prediction will be spit back within seconds. The CNN only can identify the 45 devices for which it has been trained, but the authors encouraged other clinicians to contribute images of other devices to the online portal. It takes 25 examples of a new device to train the neural network, they wrote.
“Our approach may speed up the diagnosis and treatment of patients with cardiac rhythm devices, but this paper also demonstrates how neural networks are increasingly being deployed to process large quantities of medical data throughout the health care system, and how future patient care will likely rely increasingly on computer-aided decision making,” Howard and colleagues wrote.
Daniel joined TriMed’s Chicago editorial team in 2017 as a Cardiovascular Business writer. He previously worked as a writer for daily newspapers in North Dakota and Indiana.