Creating a photoplethysmogram (PPG) sensor—shining light into the skin to measure blood volume changes—is as easy as assembling a few cheap parts from Amazon or downloading an app for a smartphone. Dr. Ali Etemad and PhD student Pritam Sarkar are using machine learning to let anyone with a PPG sensor generate continuous electrocardiogram (ECG) readings.

Dr. Etemad, an Assistant Professor in the Department of Electrical and Computer Engineering, is researching human-centric applications for machine learning and deep learning in wearable devices. “All the wearable devices are looking to provide ECG from an easy-to-wear form factor like wrists,” he said.

Electrocardiogram sensors in wearable tech are currently only available in expensive smart watches worth upwards of $600 or $700. The user can also only get a reading when they touch the watch face with their other hand to complete the electrical circuit. Dr. Etemad and Sarkar’s PPG workaround would let even the cheapest fitness wearables give the user continuous ECG readings.

A paper they authored on the applications of using PPG sensors to generate synthetic ECG readings was accepted at the prestigious Association for the Advancement of Artificial Intelligence conference in February 2021.

“It’s a big validation that what we’ve done is valuable from an AI and an application perspective,” Dr. Etemad said.

Ali Etemad and Pritam Sarkar
PhD candidate Pritam Sarkar (at left) and Dr. Ali Etemad meet at Ingenuity Labs Research Institute.

Their synthetic reading model was created using the same technology as deepfakes: Generative Adversarial Networks, which are referred to as GANs. Dr. Etemad and Sarkar taught the GAN model to take PPG readings and spit out synthetic ECG, the more accurate and valuable measurement of cardio activity.

“This is a very interesting example of a technology that has the potential to be used for bad purposes,” Dr. Etemad said. “We’re putting it to really good use.”

Now the team is working on extending their research, doubling down on testing the performance of their model by trying to detect cardiovascular diseases from the generated ECG.

“We already know that the ECG that our model creates is very accurate and can even provide more accurate measurements of heart rate compared to the original PPG,” Sarkar said. “But now we want to analyze whether our synthetic ECG can be used to identify a number of heart conditions. For that, we need large amounts of PPG data that have been flagged with heart conditions.”

They are looking to partner with Kingston General Hospital to recruit patients and expand their work into generating accurate ECGs for people with heart conditions. Dr. Etemad is cautiously optimistic that their model could even be used in the health sector to this end. “Even if the fidelity is not as high as real ECG,” he said, “We could use it to fill in the gaps and boost our knowledge of how someone’s cardiovascular system is operating in situations where direct ECG monitoring would be difficult.”

Remote cardiovascular monitoring is usually conducted with the Holter monitoring system, a stocky and uncomfortable machine that requires nodes to be connected to the wearer’s chest at all times.

As the team works on filing a patent for their technology, interest in their work is growing with a number of different companies already inquiring after their PPG-to-ECG model. “Queen’s has been super helpful in connecting us with potential interested partners,” Dr. Etemad said.

At present, their model is freely available to the public.

 “Accessible, affordable, and continuous ECG monitoring would be the ideal outcome here,” Dr. Etemad said.