Graduate Student Innovation: Using AI, Health Data to Improve Stroke and Diabetes Care
To pursue his Ph.D., Louis Gomez selected a Stevens computer science lab. Now he develops AI-driven methods to inform better medical decisions.
Each year approximately 800,000 Americans experience a stroke. Some lapse into comas or other states of minimal consciousness, during which treatment decisions must be made quickly by emergency room and intensive care unit (ICU) clinicians.
More than 1 million new cases of Type 2 diabetes are also diagnosed in the United States annually. Caught in time, it can often be managed with exercise, lifestyle changes and accurate blood-sugar monitoring and prediction.
Louis Gomez, a Stevens doctoral student working in the Health & AI Lab (HAIL) directed by accomplished computer science professor Samantha Kleinberg, wants to help address both these medical challenges.
"There are a lot of big questions we care about in relation to treating patients in an ICU, such as which stroke patients are still conscious, which are likely vegetative, and why consciousness changes," explains Gomez. "That is a big determination that guides clinical care. But sometimes you need to begin answering a big question by examining the smaller questions.
"How are we assessing consciousness? Which sources of ICU data are most readily available and continuously monitored to help with that assessment? How can we analyze and understand the data in more automatic ways to help make decisions? And how can we infer meals from glucose data to help people with diabetes to actively control their glucose levels?
"That's some of what I am looking at right now."
Leveraging faculty expertise to aid critical care
Gomez first traveled to the U.S. from his native Nigeria at the age of 16 and entered the electrical engineering program at Wichita State University.
"I was always good at math, and growing up it was just understood that those good in math end up as engineers," he recalls.
While an undergraduate in Kansas, Gomez learned that his undergraduate research professor had received a graduate degree from Stevens' biomedical engineering program.
That was enough to pique his curiosity. Noting that Kleinberg conducted leading research in healthcare analytics, he reached out to the professor, discussed their mutual areas of interest in machine learning and healthcare analytics, and applied and was accepted to the university's Ph.D. program with Kleinberg as his advisor.
"While Louis majored in electrical engineering, I saw his clear passion for computing and health," says Kleinberg. "He was so motivated and so eager to take advantage of all opportunities that I felt certain he’d have no problem changing directions.
"It has been exciting watching him grow and seeing projects already coming to fruition."
At Stevens, Gomez' primary work involves the analysis of neurological intensive care unit (NICU) data.
"When I joined HAIL, Dr. Kleinberg gave me this project to work on, arranged a data collaboration with a Columbia University hospital and gave me computing power — then directed me to figure the rest of the problem out," he explains. "It is my project, building upon her ideas."
The data, collected from approximately 90 stroke patients in the hospital (which does not identify individual patients), tracks two dozen physiological signs such as heart rate, respiration rate and intracranial pressure.
Interestingly, the Stevens team is not collecting and analyzing brain-wave EEG data for this purpose.
"That type of analysis has been done by other researchers, in prior work," notes Gomez. "Heart rate, respiration and other physiological data are always readily available in the ICU and we are trying to build a system based on this data for the best, most continual accuracy."
Current treatment protocols for stroke patients involve making careful assessments of their level of consciousness before proceeding. Most of the tests used are time-consuming, and do not allow real-time, continuous insight into each patient’s state and how it is changing.
In addition, patients are typically only assessed once per day.
"Our goal," says Gomez, "is to develop a model that can monitor and assess in a more continuous fashion, so that you can see real-time assessments and ask: is this patient's level of consciousness changing throughout the day? Why is it changing throughout the day?
"These insights can lead to early interventions as well as insight into prognosis: identifying the causes of changes in consciousness may further allow them to be predicted in advance, thus yielding better treatment decisions and preventing premature withdrawal of care."
Gomez, Kleinberg and others created an analysis system that learns to classify state of consciousness based on the collected, tracked physiological data from the 90 Columbia stroke patients and known human observations and diagnoses about their states of consciousness.
"Is there some pattern we can detect from these signals? Can we extract meaningful information from them about a state of high or low consciousness that may not be apparent to a human observer? This is what we are investigating," he explains. "Our current results show good performance classifying states of consciousness."
"Once we have this system tuned, the goal is to have clinicians use it as a tool for critical care patients."
In findings submitted, peer-reviewed and accepted to the Black in AI Workshop co-located with the prestigious 2020 Conference on Neural Information Processing Systems (NeurIPS), Gomez and Kleinberg investigated a subset of 37 stroke patients. Their artificial intelligence-driven system, using 20 points of continuously collected physiological data, performed comparably well at classifying patient state of consciousness with several other systems currently in use.
"We have gotten the same level of accuracy as studies using fMRI or EEG data, which is exciting," Gomez points out.
Eventually the Stevens team hopes to generalize its detection system to other types of ICUs, such as for those monitoring cardiac arrest or neonatal care, for example.
Hope for millions with Type 2 diabetes
Gomez is also working on a project to detect meals and improve glucose monitoring for Type 2 diabetes patients.
"A lot of people use an 'artificial pancreas' system to know, within a margin of error, what their blood glucose level is, to help forecast their future blood glucose and to deliver insulin to correct high blood glucose levels," he explains. "But these monitors have delays. If I am eating breakfast, there's usually a gap between the increase in the monitoring record and the actual increase in my blood glucose.
"If you could automatically predict a meal is happening — without having the users input this data — you could forecast an increase in blood glucose, and preemptively correct it, by giving yourself the right insulin doses whenever needed."
Indeed, Kleinberg's lab has previously published research toward development of a system that can automatically predict when people with Type 1 diabetes — whose bodies cannot produce any insulin at all, usually from an early age — are consuming meals using continuous glucose, activity and insulin dosage data.
Now Gomez hopes to extend a similar predictive system to Type 2 diabetes, a far more prevalent form of the disease in which the body does manage to manufacture some insulin — but cannot use it effectively to lower blood sugar.
"The current model we use for Type 1 diabetes has an exercise module using both heart rate to factor in intensity and insulin dosage information to predict glucose levels. We have not done this yet for Type 2 diabetes — included exercise in the model — and we also don’t have accurate insulin data, which makes this challenging. So I will be working on this and hoping to produce useful insights."
Strong student support at Stevens
Throughout the COVID-19 pandemic, Gomez has been able to continue his research largely unhindered.
"We really haven't skipped a beat since going virtual in March," he notes. "Since the material we are working with is data, it can be accessed on servers anytime remotely with the right security protocols. So, we have continued."
While Gomez hasn't been able to experience campus often this year due to COVID-19 restrictions, he did join the campus National Society of Black Engineers chapter and hopes to participate in additional social activities once the pandemic has lessened.
As he continues pursuing his doctorate, he also credits his graduate advisor for both scientific rigor and a willingness to support his Stevens experience in multiple ways.
"Dr. Kleinberg expects a lot from you, no doubt," concludes Gomez, "but where she really shines is in her support for her students. She is a great person to have on your side. As a Black student, the only one in HAIL, she has been very proactive from day one connecting me not only with support services here at Stevens but also outside the university.
"This has meant a lot to me personally and professionally, the ability to meet and interact with fellow Black researchers both on campus and also at other institutions."
Gomez also credits his fellow Ph.D. students in HAIL.
"They have also been incredibly helpful," he concludes, "providing both knowledge and guidance along my journey here at Stevens."
Learn more about computer science at Stevens: