Dina Katabi Gives The 22nd Annual Kanellakis Memorial Lecture

    Click the link that follows for more news about the Paris C. Kanellakis Memorial Lecture.

    The Paris C. Kanellakis Memorial Lecture, a tradition of more than two decades, honors a distinguished computer scientist who was an esteemed and beloved member of the Brown CS community. Paris came to Brown in 1981 and became a full professor in 1990. His research area was theoretical computer science, with emphasis on the principles of database systems, logic in computer science, the principles of distributed computing, and combinatorial optimization. He died in an airplane crash on December 20, 1995, along with his wife, Maria Teresa Otoya, and their two young children, Alexandra and Stephanos Kanellakis.

    Each year, Brown CS invites one of the field's most prominent scientists to address wide-ranging topics in honor of Paris. Last month, Dina Katabi, Inaugural Thuan and Nicole Pham Professor of Electrical Engineering and Computer Science at Massachusetts Institute of Technology, delivered the twenty-second annual Paris C. Kanellakis Memorial Lecture: "Monitoring Health and Diseases Using Radio Signals and Machine Learning". 

    In his opening remarks, Brown CS faculty member Philip N. Klein described Dina as a remarkable thinker and scientist that Paris would have been "very proud" to be represented by. Stepping to the podium, Katabi described herself as someone who doesn't research wearable technology but something that she calls invisibles: "The future of healthcare is data-driven healthcare, but where does that data come from?....The sick and the old don't use wearables."

    To answer her own question, Dina walked the audience through an extensive series of projects culled from just seven years of work. At their center is Emerald, a touchless sensor and machine learning platform for health analytics that plugs into a wall outlet and resembles a Wi-Fi router. Emerald observes a patient by sending radio signals through their home and then analyzes the results with machine learning, quantifying important data points such as stride length and gait velocity and detecting falls without requiring the user to leave home or change their routine.  

    "How can you take a technology and really connect it to a disease?" Dina asked. "What does it mean in the context of a disease?" One of her most powerful examples was analyzing sleep, which she described as a platform for understanding many health issues. Not only do invisibles make it possible, she explained, to quantify things that doctors typically can't, such as how many times a patient gets up in the night, they can do this in a natural setting, without the travel, extensive preparation, and technician-led analysis typical of a sleep study. 

    The challenges that her research group faced were many. As Dina put it, "How do you extract the human body from this mess?" Radio signals reach their receiving antenna by multiple paths, often complicated by numerous reflections; her team's neural network needed to be trained for very different modalities than the standard sets of images and videos; clinical trials required small datasets of real patients, where spurious correlations were possible. 

    Despite these obstacles, Katabi's results have compared favorably to the gold standards of the diagnostic world. As just two examples, Emerald's sleep analysis obtained 80% accuracy when compared to analysis of polysomnograms by trained technicians, and 97% accuracy when compared to measurement of breathing by a FDA-approved belt.

    The benefits, Dina explained, are very real: by using Emerald to look at certain biomarkers, pharmaceutical companies testing a new drug could potentially have findings worth reviewing in one year instead of several, allowing it to be brought to market much more quickly. By observing a patient in their own home instead of waiting for a doctor to notice the visible symptoms of Parkinson's, Emerald might be able to diagnose the disease as much as a decade earlier, before 40-80% dopamine loss has set in. 

    As she wound down her talk, Dina described herself as an optimist, interested in using machine learning to do something that present-day medicine isn't capable of on its own. "People work about AI taking jobs," she said, "but I feel it will do things we can't do today."

    Brown CS faculty member James Tompkin was one of the event's many attendees. "Remote measurement for long-term persistent health monitoring," he says, "would provide a new kind of insight to healthcare professionals, so this technology holds real promise. Dina's findings on early detection of disease via long-term monitoring – made possible by these low-cost minimal intervention monitoring methods – are exciting."

    A recording of the lecture is available here (Brown login required).

    For more information, click the link that follows to contact Brown CS Communications Manager Jesse C. Polhemus.