All Thayer News

Every Breath You Take

Dec 01, 2021   |   by Kathryn Lapierre   |   Dartmouth Engineer

Professor Kofi Odame and PhD candidate Maria Nyamukuru use “smart” monitoring and machine-learning algorithms to effectively respiratory diseases.

Maria Nyamukuru
Maria Nyamukuru Th'24 (Photo by Lars Blackmore)

Difficulty breathing, congestion, coughing, and wheezing—these are all-too-familiar symptoms in our current COVID-19 world. But long before the pandemic, respiratory illnesses have plagued millions across the globe, with chronic conditions such as chronic obstructive pulmonary disease (COPD) being among the deadliest—and most expensive—to treat.

“The problem with COPD is monitoring patients,” explains Maria Nyamukuru, PhD candidate in Professor Kofi Odame’s Analog Lab at Dartmouth. The group focuses on designing highly efficient analog and mixed signal integrated circuits for sensor interfaces in wearable biomedical devices.

“The most conventional way that COPD is monitored involves patients going into the hospital once or twice a year, depending on how severe their disease is, in order to get a spirometry test,” says Odame. “They go into a lab and blow into this machine. The machine measures how forcefully they can blow out, which measures their lung capacity, pointing to how healthy their lungs are.”

Repeated hospital visits can be time-consuming, inconvenient, and expensive, and experts are exploring the possibility of shifting how they can better monitor lung disease at home. This practice has already been in place for years for other conditions, including everything from pregnancy to diabetes to food sensitivity.

“In-home monitoring is a relatively new development,” says Odame. “And two big reasons for it not being trusted by doctors is because, first, patients forget to do these tests; and second, even when they do remember, it’s not done with the best level of accuracy.”

This level of in-home testing can provide questionable results, Odame continues, which delivers a collection of bad data and leads to poor patient care and diagnosis. So Odame and Nyamukuru are seeking a solution that collects accurate data at home from patients through a more streamlined and passive process.

The pair is trying to leverage a ubiquitous modern device: the smart watch.

“The spirometry test requires a lot of human intervention to use it. We want to address the issues with the handheld spirometer by monitoring COPD using a smartwatch,” says Nyamukuru.

Much of the population carries on their wrists everything they need to monitor their lung health—with a few advanced technical moderations, of course. “Smart watches, which are becoming more and more common, already have a way to measure your heart rate,” says Odame. “One way of measuring heart rate is to use an electrocardiograph, or ECG, signal. The other way is to use photoplethysmography, or PPG, signals.”

Most smart watches can already measure heart rate, and measuring lung health is related. “There is a relationship between the heart signal that’s measured—how fast your heart is beating and how hard—and your breathing pattern,” he says. “We’re trying to extract information about breathing from the heart signal and then take the information about breathing and make these broader inferences about lung health. We’re trying to correlate that to what we would be measuring if the patients went into the hospital and did the gold standard conventional spirometry test.”

In addition to the issues with monitoring COPD, patients also struggle with managing the disease and often succumb to exacerbations, which are acute events that worsen the disease. “It often ends with patients having to be rushed to the ER and sometimes hospitalized,” says Odame. “That’s not a pleasant experience for the patient or their family, and it’s an expensive experience for the healthcare system. It turns out that often a few weeks or several days before such an acute event, there is this measurable drop in lung function and lung health. One potential use case would be this device identifying such a drop and alerting whomever, maybe the patients, maybe their care-givers, maybe the healthcare givers, and try to prevent that exacerbation that is eminent.

The algorithm they are delivering may do just that.

“A cool thing about this algorithm is that it takes advantage of the existing technology in the space. We’ll just use these ECG signals to develop our algorithm,” says Nyamukuru.

It all sounds simple when you boil it down: Take an existing tool and adapt it to measure lung health. Problem solved. But, naturally, there is more to it.

“Developing the machine-learning algorithm and making sure it can fit and run on a smart watch device accurately is my main task,” says Nyamukuru. “It’s a resource-constrained device with limited memory and computation resources.”

The Power of Neural Networks

Nyamukuru is currently leveraging the power of multitask neural networks to extract both respiratory rate and fractional inspiratory time (FIT)—or what percentage of a breath is characterized as “inspiring,” or breathing in—from ECG signals.

“This is important because it’s telling of lung and airway obstruction,” she says. “For example, people with obstructed airways often have a lower FIT, usually less than 20 percent, because it takes longer to exhale. Respiratory rate and FIT are measurements that we are looking at to infer COPD severity. Extracting these metrics from ECG signals would be great because that means we can take advantage of ubiquitous wear-ables with ECG sensing to monitor lung airways in real time.”

“Most of the challenge is in developing the machine-learning algorithm that can extract some of these respiratory metrics from ECG signals. It has been really difficult to find an existing database that has clean enough data that’s also collected in a way that’s usable,” Nyamukuru adds.

The team worked with a pulmonologist at the Dartmouth-Hitchcock Medical Center (DHMC) using a hospital-grade spirometer to collect respiratory signals and COPD metrics.

“I don’t know if it’s because of the size of the school or just the inherent culture, but this is a collaborative environment,” says Odame. “I am not a biomedical engineer. I’m not a machine-learning scientist. My expertise is in circuit design, specifically analog integrated circuit design. That’s a very specialized, narrow field. Coming to Thayer, and to Dartmouth in general, I was struck by just how accommodating the environment is to crossing these boundaries between traditional fields. The strong interaction between Thayer and DHMC was a huge attraction.”

Nyamukuru came for similar reasons, noting that Dartmouth’s collaborative approach and the freedom given to students to explore interests beyond traditional engineering made it an attractive place to learn and apply her knowledge in the real world.

“Before I came to Thayer I was working in industry, mostly focusing on embedded systems, and I was exposed to a project that used machine learning that was really interesting. When I was applying for my PhD, I looked at machine learning, and hardware and embedded systems. Most of the programs seemed to be very constrained, except for Professor Odame’s lab, which is why I was initially drawn to it.”

The flexibility to work on projects such as these allowed Nyamukuru to explore multiple interests, and she relishes the chance to work side by side with researchers and doctors at DHMC as they collect data.

The COVID-19 pandemic temporarily put the work on hold. “We had to stop collecting data from the hospital, which has been a huge challenge,” says Nyamukuru.

While data collection will resume in the coming months, the team is adapting the best they can. “We’re trying to figure out ways to collect our own data down the road and make use of what we have right now and learn from it.”

—Kathryn Lapierre is editor of Dartmouth Engineer

For contacts and other media information visit our Media Resources page.