Dunn Receives NSF CAREER Award to Improve Biosignal Algorithms

3/20/24 Pratt School of Engineering

Competitive five-year grant will help Jessilyn Dunn address long-standing issues in how biomedical data for wearable devices is collected and studied

Jessilyn Dunn
Dunn Receives NSF CAREER Award to Improve Biosignal Algorithms

Jessilyn Dunn, assistant professor of biomedical engineering at Duke University, has been awarded a National Science Foundation Faculty Early Career Development (CAREER) Award.

The competitive five-year, $597,774 grant for outstanding young faculty will support Dunn’s research as she explores new methods to improve the AI-based biosignal algorithms that support health and clinical decision-making from biomedical data.

Dunn runs the Duke BIG IDEAs Lab (Biomedical Informatics Group: Integrating Data Engineering and Analytics), where she explores how wearable devices, like smartwatches, fitness trackers and other medical monitors, can identify digital biomarkers for diseases including diabetes, heart disease, COVID-19 and even the flu.

Drawing from her expertise in both data science and biomedical engineering, Dunn also develops new methods to assess and improve the accuracy of wearable technologies. With the award, Dunn will build on this work to address two problems that have long plagued the accuracy of biosignal algorithms: bias and drift.

Data for training such algorithms is expensive to collect and analyze, so researchers often work with data from easy-to-access sources. Unfortunately, this data usually comes from populations that are most often healthy, wealthy and white. This limitation often leads to bias, where the algorithms do not work equally well for all people under all circumstances.

Jessilyn Dunn of Duke University

The question becomes, ‘Is this data actually comparable? If you develop an algorithm based on data collected from one technology, is it going to work as well for the data coming from a different technology? These are questions that we need to know so we can set guidelines about when to retrain a model or when more data is needed to improve the model.

Jessilyn Dunn Assistant Professor of Biomedical Engineering

“Previous work shows a relationship between skin tone and device performance, for example, in pulse oximetry measurements,” said Dunn. “A lot of wearable technologies use the propagation of a light signal through the skin to take measurements of biosignals like heart rate and blood oxygenation—detecting those signals can be more challenging on darker skin tones. This could be mitigated by using a stronger light source or a different wavelength. However, this issue is often unaccounted for, leading to poor quality data used to train algorithms for certain subgroups of people like those with darker skin.”  

As the capabilities of wearable devices continue to grow and change, the sensors used to track biometric data like heart rates or blood pressure must also constantly improve.

But, this constant evolution of sensor technologies leads to a complication known as drift, which occurs when algorithms are trained using data from sensors that collect data differently over time. For example, blood pressure could be measured by either the auscultatory method (i.e., a traditional upper arm cuff) or using optical signals similar to pulse oximetry technology.

Dunn is hopeful that developing a framework to identify sources of bias and drift in these biosignal algorithms will make wearable tools more helpful to larger populations inside and outside hospital systems.

To help ensure their broader success, Dunn and her team will work with Recycle Health, a device recycling program that can help distribute wearable fitness devices to populations that wouldn’t normally purchase or have access to one. Her team will also launch a new collaboration with Women in Data Science to create new educational and professional development opportunities for women interested in data science, AI and health.

“One in three Americans uses wearable devices, but the data we use to train these tools isn’t reflective of that number,” said Dunn. “This work will allow us to improve not only the effectiveness of wearables in early disease detection and clinical monitoring but also increase trust for using these low-cost, ubiquitous tools in health care settings.”