Think Tank
(Stelsone/Getty)
Black drivers are pulled over more often than their White counterparts but stopped less frequently after sunset, when a “veil of darkness” masks their race. So found NYU’s Ravi Shroff and his colleagues at the Stanford Open Policing Project in a yearslong analysis of close to a hundred million stops from 21 state patrol agencies and 35 municipal police departments. Unprecedented in scope, the study presented massive challenges. Policing in America is “hyper local,” says Shroff—assistant professor of applied statistics at Steinhardt and of urban informatics at the Center for Urban Science and Progress—with some 20,000 different departments operating at the municipal, county, state or federal level and “very little standardization” in what is recorded, as well as how and when. “Part of our effort,” he says, “was to assemble these data sets so they could be analyzed as a coherent whole.” Requests for information went to all 50 states, their state patrol agencies, and about 150 cities, says Shroff, and responses ranged from “ ‘we don’t collect that data’ to ‘maybe we do, but we’re gonna have to charge you one hundred thousand dollars,’ and we weren’t able to bargain them down.” The upshot: “At a high level, the team found bias against minorities.”
Apples and Apples
A simple comparison of early-evening stops on a still-bright summer night against those at the same time on an already-darkened winter eve might not account for other variables, such as differences in driving patterns by race between seasons. So the investigators looked at stops at 7:00 p.m. just days apart—before and after daylight saving time.
The Threshold Test
An officer’s suspicion threshold for deciding to search a stopped vehicle for contraband (e.g., illegal drugs or firearms) might be, say, 10 percent. If they think there’s a 20 percent chance the driver is carrying something, that’s higher than their threshold of 10 percent and they’ll conduct a search. The question is: are different thresholds being used to make decisions about searching drivers of different races? By examining the rate at which stopped drivers were searched and the likelihood that they turned up contraband, says Shroff, the researchers found that “the bar for searching was lower for Black drivers than for White drivers.”
Policy Effects
To illustrate what else they could do with the data set, the team studied the impact of one large-scale statewide policy change on police behavior. “Searches declined pretty dramatically for all race groups in Colorado and Washington state following marijuana legalization,” says Shroff, “but there was still a gap between the search rates for White, Black, and Hispanic individuals.”
—Dulcy Israel
Health on the Line
A routine medical exam almost always involves a listen to the lungs. Abnormal breathing patterns, for instance, can point to pneumonia, asthma, or some other pulmonary condition. But in-person visits are a challenge these days, so NYU researchers are paving the way for remote screening of COVID-19 to help healthcare providers reach a wider population. Kyunghyun Cho, associate professor at the Courant Institute of Mathematical Sciences and the Center for Data Science, and doctoral candidate William Falcon are investigating the efficacy of using the phone as a kind of stethoscope with their Breathe for Science platform. Participants dial a number and either cough three times or take three breaths into the mic. Along with an online survey, their recorded breathing patterns are then stored anonymously in a database. Cho hopes that Breathe for Science will function both as a diagnostic tool by telemedicine doctors to supplement what they already know about the patient, and as a wider surveillance system for the community of disease. “If you see that a certain area has anomalies in terms of the breathing patterns,” Cho explains, “then we might be suspicious that there is some kind of disease or outbreak in a certain region.”
—Dulcy Israel
(metamorworks/Getty)
When he was a teaching assistant at another university, Bruno Abrahao—now assistant professor of information systems and business analytics at NYU Shanghai—was approached by students who were unable to concentrate and feeling depressed. Alarmed, Abrahao connected them with professionals at the school’s mental health clinic, but the students weren’t necessarily getting better after treatment. “The medication made them more apathetic and removed their passion for the subjects they were studying,” he says. “We have this treatment to get them back on track, but it’s not effective—or making them worse.”
Much of the problem stems from the fact that psychiatric drugs land on the market with insufficient testing, after brief, six-week-long clinical trials conducted on just a few hundred people. But Abrahao noticed that many share their experiences with psychiatric medications online, citing headaches, sleeping too much, and weight gain. He wondered if there was a way to help gather more data using his computational skills: “The idea was, why can’t we collect all this information about this huge population for a long period of time to learn more about the medications and, in a way, extend the clinical trials?”
Abrahao teamed with researchers from Harvard Medical School, Microsoft Research, and Georgia Tech to create an AI-based model that essentially mimics a large-scale clinical trial by harnessing publicly available information that identifies depressive language and tracks changes in mood and behaviors through users’ self-reporting across hundreds of thousands of tweets.
The researchers found that Twitter posts by those who took drugs called SSRIs—which include many of the most popular antidepressants—showed long-term worsening of symptoms including anxiety, depression, and thoughts of suicide. By contrast, the posts of users taking an older group of drugs called TCAs showed more improvement in depressive symptoms over the two-year study. Abrahao notes that the results reflect what his Harvard collaborator, psychiatry instructor John Torous, has observed in his clinical experience. In fact, psychiatrists and patients have expressed doubts about the effectiveness of SSRIs for years, so the new AI study—published in the Proceedings of the Association for the Advancement of Artificial Intelligence—substantiates the severity and scale of their negative effects.
The team hopes that AI can help practitioners better understand the long-term implications of psychiatric medications. Still, it’s fascinating and almost dichotomous to use a machine approach on such a fundamentally human issue—and it reminds Abrahao of something a psychoanalyst once told him: “A good psychoanalyst is like a musician who develops their ear to hear the structure of the music. But instead of notes, tempo, harmony, melody, we need to train our ears to be able to identify behavioral patterns as the client speaks about the most varied topics.” He notes that AI, through the data collected from Twitter, is able to provide a framework of linguistic signals to detect patterns of behavior and changes that could correlate to the user’s emotional symptoms. “In a way,” Abrahao muses, “AI is on track to developing that ‘musician’s ear.’ ”
—Paula Akpan
(anusorn nakdee/Getty)
Nigel Bunnett is tackling a near-universal element of the human condition—and with it, one of America’s most vexing social problems. “Pain is necessary for survival, because it allows you to be aware of injury or avoid dangerous situations,” explains Bunnett, professor and chair of the Department of Molecular Pathobiology at the College of Dentistry. Yet injury or disease can transform acute pain into a condition that is chronic and debilitating. It’s a common experience: one in four people suffers from chronic pain in their lives. But existing treatments aren’t especially effective, or have unacceptable side effects and high risk for addiction. “And that’s well-illustrated by opioids,” he says.
But Bunnett and an international team—including scientists at Columbia University, Monash University, and the University of Santiago, Chile—may have found a safer, more effective alternative to opioids by using nanoparticles to deliver pain meds into specific compartments of nerve cells. The researchers study a family of proteins called G protein-coupled receptors, which transmit information, such as pain signals, at the surface of nerve cells. These receptors are the target of a third of all clinically used drugs, including opioids, all of which work by blocking receptors at the surface of the cells, jamming the signal. Yet recent discoveries have shown that these surface receptors are, literally and figuratively, surface phenomena: pain signals actually move inward to compartments inside the cell called endosomes. “And there, they continue to function,” Bunnett says. “That causes nerve cells to generate long-lasting signals, which we believe underlie chronic pain.”
The discovery raised a new challenge: can pain medication be delivered directly to the endosomes—the heart of the system—and bypass the surface? “That’s why we turned to nanoparticles,” says Bunnett, whose findings were published in Nature Nanotechnology. Nanoparticles have been used to deliver cancer-treating drugs. They have never been used to treat pain, but were attractive to Bunnett’s team for several reasons. First, endosomes are acidic. The researchers thought to encase the painkilling drug in a pH-sensitive nanoparticle that would dissolve and release the medication once inside. Second, nanoparticles can be engineered to be “preferentially taken up,” he says, by cells involved in pain-signaling, producing better targeting. Third, a whole cocktail of drugs can be packaged inside a nanoparticle, targeting all the different receptors involved in pain-signaling. And there are many of them, says Bunnett, performing similar, redundant functions “because pain is such an important process for survival.”
A nanoparticle approach has many advantages. Direct targeting allows for more effective treatment of pain at lower drug doses. These can include non-opioid drugs that are non-addictive and don’t carry the same harmful side effects, but weren’t as effective when the surface cell receptors were their target. “The process we’ve developed is essentially like giving a drug infusion into the endosome of the cell,” Bunnett says. “By delivering a previously ineffective drug to the right compartment within the cell, it became highly effective as a pain treatment.”
They’re now testing the nanoparticle delivery method for oral cancer pain that impairs crucial everyday tasks like talking and eating, and is fairly resistant to opioid treatments; “one of the most painful forms of cancer,” Bunnett notes. Preclinical trials with animals have been promising. “The main challenge in all drug development is actually showing that they’re safe and effective in people,” he says, but in this case because they are using drugs that have already been FDA approved, “it might be sooner than developing a brand-new drug.”
—Audrea Lim