A data scientist and a biomedical engineer walk into a networking event.
No, this isn’t the lead-in to a cheesy joke – it’s the beginning of an exciting collaboration!
TRANSFORM HF researchers Dr. Daniel Franklin and Dr. Chris McIntosh were awarded a $100,000 Catalyst Grant from the Data Sciences Institute and Tanenbaum Institute for Science in Sport earlier this year for their project titled Development of Convolutional Neural Network for Motion Artifact Mitigation in Wearable PPG Devices.
Franklin and McIntosh first met at a TRANSFORM HF event in 2022. They were seated at the same roundtable for a networking activity centered around digital health. There they began a discussion about their respective work in device development and data science that evolved into a fruitful collaboration. “If we hadn’t met at a TRANSFORM HF event, there would be no Catalyst Grant,” says McIntosh.
Wearable devices are becoming more prominent in remote health monitoring. Employing an optical technique known as photoplethysmography, or PPG, wearables use light to non-invasively detect changes in blood flow and oxygenation to provide heart rate and pulse oximetry readings. However, motion can greatly reduce the accuracy and usability of data. In many cases, accelerometers embedded in wearables are used to detect excessive movement and dispose of the associated ‘contaminated’ data.
You can imagine that for individuals working physical jobs or for patients on a bumpy ambulance ride, much of the collected data would be deemed contaminated. But what if we could account for motion to ensure uninterrupted data is high quality and usable?
Led by graduate trainees Matthew Kenyon Lee and Yuan (Will) Gao, Franklin and McIntosh’s research seeks to do just that.
Matthew Kenyon Lee
Lee started as a Masters student in the Franklin lab in June 2023 after completing an undergraduate degree in biomedical engineering at the University of Waterloo. He gained an interest in applying data science to medical device development for cardiac monitoring through several co-ops at tech companies. Lee’s role in this project includes collecting data and incorporating new metrics to train and test models.
Yuan (Will) Gao
Gao is a PhD student in the Department of Medical Biophysics under Dr. Chris McIntosh. His research involves applying machine learning to provide interpretability and clinical utility to wearable data. Gao is particularly interested in bringing machine learning innovation into clinical practice in heart failure. Gao’s role involves developing deep learning models and assisting in performance assessment.
Cutting Through the Noise
Lee compares the work to active noise cancellation in headphones: “Noise cancelling headphones improve the sound you’re receiving by blocking out ambient noise. We’re doing something similar, but the ‘noise’ we’re looking to cancel is motion.” By accounting for motion, wearables could provide cleaner data for more reliable insights into the cardiac status of a user.
The work first involves the development of a multimodal sensor. Combining force and multiwavelength optical measurements, Franklin and Lee’s sensor will capture motion between the sensor and actual point of contact with the body – a critical limitation of current wearable devices. By employing a full set of wavelengths (unlike the typical two or three wavelengths in most wearable devices), this novel sensor could accurately capture other helpful metrics, like blood pressure.
Then comes the data science: McIntosh and Gao will build a deep learning model and train it to reconstruct the PPG waveform to account for detected motion and produce clean data. The work will ultimately lead to more advanced remove health care devices and algorithms.
The Next Frontier
“If AI is the present, wearables with AI are the next Frontier,” states McIntosh. “When we thought about AI a year ago, we thought about static inputs, especially in medicine. But wearables with AI give us insight into real-time function and a day-to-day picture of how things are changing.”
While wearables play a key role in obtaining information, AI is needed to sift through the inordinate amount of data collected. McIntosh and Gao reflect on an earlier research project, which involved investigating accelerometer data. “For 50 patients over 90 days, we generated five terabytes of raw data,” shares Gao. “Five terabytes! That’s a crazy amount of data.” In an already overburdened healthcare system, clinicians require appropriate and interpretable data to make timely treatment decisions.
Working together, wearables and AI could also address existing health equity concerns associated with pulse oximetry. A study was published earlier this year assessing the performance of 11 fingertip pulse oximeters in healthy human participants with diverse skin pigmentation. Nine of these devices performed worse on darker skin tones. This could translate to at-risk patients being overlooked for care based solely on the pigmentation of their skin. TRANSFORM HF has previously supported the integration of novel sensors with advanced machine learning algorithms to develop more accurate and equitable devices – work that must be continued.
TRANSFORM HF congratulates the team on receiving funding for this exciting and important project. If you’re looking for collaborators or inspiration, please join us at our next network event – you never know who you might end up sitting beside!
Sign up for our monthly newsletter to get the latest from the network.
