Artificial Intelligence

MIT engineers have discovered a new way of generating electricity using tiny carbon particles that can create a current simply by interacting with liquid surrounding them.

The liquid, an organic solvent, draws electrons out of the particles, generating a current that could be used to drive chemical reactions or to power micro- or nanoscale robots, the researchers say.

“This mechanism is new, and this way of generating energy is completely new,” says Michael Strano, the Carbon P. Dubbs Professor of Chemical Engineering at MIT. “This technology is intriguing because all you have to do is flow a solvent through a bed of these particles. This allows you to do electrochemistry, but with no wires.”

In a new study describing this phenomenon, the researchers showed that they could use this electric current to drive a reaction known as alcohol oxidation — an organic chemical reaction that is important in the chemical industry.

Strano is the senior author of the paper, which appears today in Nature Communications. The lead authors of the study are MIT graduate student Albert Tianxiang Liu and former MIT researcher Yuichiro Kunai. Other authors include former graduate student Anton Cottrill, postdocs Amir Kaplan and Hyunah Kim, graduate student Ge Zhang, and recent MIT graduates Rafid Mollah and Yannick Eatmon.

Unique properties

The new discovery grew out of Strano’s research on carbon nanotubes — hollow tubes made of a lattice of carbon atoms, which have unique electrical properties. In 2010, Strano demonstrated, for the first time, that carbon nanotubes can generate “thermopower waves.” When a carbon nanotube is coated with layer of fuel, moving pulses of heat, or thermopower waves, travel along the tube, creating an electrical current.

That work led Strano and his students to uncover a related feature of carbon nanotubes. They found that when part of a nanotube is coated with a Teflon-like polymer, it creates an asymmetry that makes it possible for electrons to flow from the coated to the uncoated part of the tube, generating an electrical current. Those electrons can be drawn out by submerging the particles in a solvent that is hungry for electrons.

To harness this special capability, the researchers created electricity-generating particles by grinding up carbon nanotubes and forming them into a sheet of paper-like material. One side of each sheet was coated with a Teflon-like polymer, and the researchers then cut out small particles, which can be any shape or size. For this study, they made particles that were 250 microns by 250 microns.

When these particles are submerged in an organic solvent such as acetonitrile, the solvent adheres to the uncoated surface of the particles and begins pulling electrons out of them.

“The solvent takes electrons away, and the system tries to equilibrate by moving electrons,” Strano says. “There’s no sophisticated battery chemistry inside. It’s just a particle and you put it into solvent and it starts generating an electric field.”

“This research cleverly shows how to extract the ubiquitous (and often unnoticed) electric energy stored in an electronic material for on-site electrochemical synthesis,” says Jun Yao, an assistant professor of electrical and computer engineering at the University of Massachusetts at Amherst, who was not involved in the study. “The beauty is that it points to a generic methodology that can be readily expanded to the use of different materials and applications in different synthetic systems.”

Particle power

The current version of the particles can generate about 0.7 volts of electricity per particle. In this study, the researchers also showed that they can form arrays of hundreds of particles in a small test tube. This “packed bed” reactor generates enough energy to power a chemical reaction called an alcohol oxidation, in which an alcohol is converted to an aldehyde or a ketone. Usually, this reaction is not performed using electrochemistry because it would require too much external current.

“Because the packed bed reactor is compact, it has more flexibility in terms of applications than a large electrochemical reactor,” Zhang says. “The particles can be made very small, and they don’t require any external wires in order to drive the electrochemical reaction.”

In future work, Strano hopes to use this kind of energy generation to build polymers using only carbon dioxide as a starting material. In a related project, he has already created polymers that can regenerate themselves using carbon dioxide as a building material, in a process powered by solar energy. This work is inspired by carbon fixation, the set of chemical reactions that plants use to build sugars from carbon dioxide, using energy from the sun.

In the longer term, this approach could also be used to power micro- or nanoscale robots. Strano’s lab has already begun building robots at that scale, which could one day be used as diagnostic or environmental sensors. The idea of being able to scavenge energy from the environment to power these kinds of robots is appealing, he says.

“It means you don’t have to put the energy storage on board,” he says. “What we like about this mechanism is that you can take the energy, at least in part, from the environment.”

The research was funded by the U.S. Department of Energy and a seed grant from the MIT Energy Initiative. More

In his research and in other parts of life, Ankur Moitra likes to journey off the beaten path. His explorer mentality has brought him to at least one edge of the unknown — where he seeks to determine how machine learning, used in increasingly diverse and numerous applications, actually works.

“Machine learning is eating up the world around us,” says Moitra, a theoretical computer scientist and associate professor in MIT’s Department of Mathematics, “and it works so well that it is easy to forget that we don’t know why it works.”

Moitra says he is attempting to put machine learning “on a rigorous foundation,” analyzing the methods that are currently used to put it into practice. He is also, he says, “trying to design fundamentally new algorithms that can expand our toolkit. As a byproduct, algorithms we understand rigorously can also aspire to be ones that are more robust, interpretable, and fair.”

Moitra was raised to be an independent thinker. Growing up in Niskayuna, New York, he was surrounded by a family of computer scientists. His parents encouraged him, however, to explore his many other interests.

“I decided pretty early on that computer science was definitely not cool,” he says. “But the joke was on me. Eventually I came to discover computer science and mathematics on my own and fell in love with them.”

Moitra received his bachelor’s degree in electrical and computer engineering from Cornell University in 2007. He earned his master’s and PhD from MIT in computer science, in 2009 and 2011, and joined the MIT faculty in 2013. Moitra received tenure in 2019, and is currently a principal investigator in MIT’s Computer Science and Artificial Intelligence Laboratory and a core member of the Statistics and Data Science Center.

Throughout Moitra’s education, his independence only grew. He discovered that not only did he want to come up with his own answers involving algorithms and their connections with such areas as machine learning, statistics and operations research, he wanted to be the one formulating the questions.

“I realized that I do my best research when I make up my own questions,” Moitra says, “and that’s a perfect fit for theoretical machine learning where we often don’t know where to begin.”

Moitra says that in his approach to research, “every trick you can dream up is fair game. It doesn’t matter how ugly or complicated your proof gets.”

His intellectual adventurousness has drawn the admiration of colleagues and mentors along the way. When Moitra won a David and Lucile Packard Fellowship in 2016, Professor Tomasz Mrowka said, “He is the dream colleague: He is deeply intellectually curious,” and referred to his “fundamental contributions to his discipline.”

In his teaching, Moitra encourages his students to venture out of “safe areas where other researchers have laid the groundwork and asked the right questions that you’re now hoping to answer.”

On the other hand, he tempers this free-ranging approach while teaching or giving talks: “I think about how simple I can make something, and whether there are some real-world examples that help drive it home.”

This teaching approach lands well. In 2018, Moitra won a School of Science teaching prize for his graduate-level course 18.408 (Algorithmic Aspects of Machine Learning). His nominators called him an “inspirational, caring and captivating” teacher.

Moitra says MIT is an excellent environment for him.

“Everyone is brimming with energy, and excited to make the world a better place,” he says. “It’s infectious.”

Between his teaching responsibilities and spending time with his wife and two children — with a little time out for playing or watching sports — Moitra’s schedule is full. Late at night, when he’s on his own, is when he does his best thinking, he says.

“Once in a while I get so obsessed with a problem and feel like it’s so close to being solved that I just can’t sleep,” he says. “I stay up for hours pacing around the house. As a professor, your inbox is always flooded and your schedule is always jam-packed with meetings. But at night, no one needs you and everyone is sound asleep, and I can think deeply without any distractions.”

It’s in those hours that Moitra can “venture into uncharted territory” and wander freely, sometimes making discoveries that become pivots to new areas of research.

“There really are basic, fundamental questions out there that are exciting and that no one has dared to ask before,” he says. “And when you discover something new like that, it’s a special kind of joy when other people start to join you on your expedition.” More

The year 2030 serves as the resolution to the United Nation’s Agenda for Sustainable Development. The agenda, adopted in 2015 by all UN member states including the United States, mobilizes global efforts to protect the planet, end poverty, foster peace, and safeguard the rights of all people. Nine years out from the target date, the sustainable development goals of the agenda still remain ambitious, and as relevant as ever.

MIT Lincoln Laboratory has been growing its efforts to provide technology solutions in support of such goals. “We need to discuss innovative ways that advanced technology can address some of these most pressing humanitarian, climate, and health challenges,” says Jon Pitts, who leads Lincoln Laboratory’s Humanitarian Assistance and Disaster Relief Systems Group.

To help foster these discussions, Pitts and Mischa Shattuck, who serves as the senior humanitarian advisor at Lincoln Laboratory, recently launched a new lecture series, called the Future of Humanitarian Technology.

In the inaugural session on April 28, Lincoln Laboratory researchers presented three topics inherently linked to each other — those of climate change, disaster response, and global health. The webinar was free and open to the public.

Play video

The Future of Humanitarian Technology: MIT Lincoln Laboratory hosted a seminar exploring climate change, disaster response, and global health technology and how these areas might look ten years from now.

Accelerating sustainable technology

Deb Campbell, a senior staff member in the HADR Systems Group, started the session with a discussion of how to accelerate the national and global response to climate change.

“Because the timeline is so short and challenges so complex, it is essential to make good, evidence-based decisions on how to get to where we need to go,” she said. “We call this approach systems analysis and architecture, and by taking this approach we can create a national climate change resilience roadmap.”

This roadmap implements more of what we already know how to do, for example utilizing wind and solar energy, and identifies gaps where research and development are needed to reach specific goals. One example is the transition to a fully zero-emission vehicle (ZEV) fleet in the United States in the coming decades; California has already directed that all of the state’s new car sales be ZEV by 2035. Systems analysis indicates that achieving this “fleet turnover” will require improved electric grid infrastructure, more charging stations, batteries with higher capacity and faster charging, and greener fuels as the transition is made from combustion engines.

Campbell also stressed the importance of using regional proving grounds to accelerate the transition of new technologies across the country and globe. These proving grounds refer to areas where climate-related prototypes can be evaluated under the pressures of real-world conditions. For example, the Northeast has older, stressed energy infrastructure that needs upgrading to meet future demand, and is the most natural place to begin implementing and testing new systems. The Southwest, which faces water shortages, can test technologies for even more efficient use of water resources and ways to harvest water from air. Today, Campbell and her team are conducting a study to investigate a regional proving ground concept in Massachusetts.

“We will need to continuously asses technology development and drive investments to meet these aggressive timelines,” Campbell added.

Improving disaster response

The United States experiences more natural disasters than any other country in the world and has spent $800 billion in last 10 years on recovery, which on average takes seven years.

“At the core of disaster support is information,” said Chad Council, also a researcher in the HADR Systems Group. “Knowing where impacts are and the severity of those impact drives decisions on the quantity and type of support. This can lay the ground work for a successful recovery … We know that the current approach is too slow and costly for years to come.”

By 2030, Council contends that the government could save lives and reduce costs by leveraging a national remote sensing platform for disaster response. It would use an open architecture that integrates advanced sensor data, field data, modeling, and analytics driven by artificial intelligence to deliver critical information in a standard way to emergency managers across the country. This platform could allow for highly accurate virtual site inspections, wide area search-and-rescue, determination of road damage at city-wide scales, and debris quantifications.

“To be clear, there’s no one-size-fits-all sensor platform. Some systems are good for a large-scale disaster, but for a small disaster, it might be faster for local transportation department to fly a small drone to image damage,” Council said. “The key is if this national platform is developed to produce the same data as local governments are used to, then this platform will be familiar and trustworthy when that level of disaster response is needed.”

Over the next two years, the team plans to continue to work with the Federal Emergency Management Agency, the U.S. National Guard, national laboratories, and academia on this open architecture. In parallel, a prototype remote sensing asset will be shared across state and local governments to gain enthusiasm and trust. According to Council, a national remote sensing strategy for disaster response could be employed by the end of 2029.

Predicting disease outbreaks

Kajal Claypool, a senior staff member in the Biological and Chemical Technologies Group, concluded with a discussion on using artificial intelligence to predict and mitigate the spread of disease.

She asks us to fast-forward nine years, and imagine we have convergence of three global health disasters: a new variant of Covid-30 spreading across globe, vector-borne diseases spreading in central and south America, and the first carrier with Ebola has flown into Atlanta. “Well, what if we were able to bring together data from existing surveillance systems, social media, environmental conditions, weather, political unrest, and migration, and use AI analytics to predict an outbreak down to a geolocation, and that first carrier never gets on the airplane?” she asked. “None of these are a far stretch.”

Artificial intelligence has been used to tackle some of these ideas, but the solutions are one-offs and siloed, Claypool said. One of the greatest impediments to using AI tools to solve global health challenges is harmonizing data, the process of bringing together data of varying semantics and file formats and transforming it into one cohesive dataset.

“We believe the right solution is to build a federated, open, and secure data platform where data can be shared across stakeholders and nations without loss of control at the nation, state, or stakeholder level,” Claypool said. “These siloes must be broken down and capabilities available for low- and middle-income nations.”

Over next few years, the laboratory team aims to develop this global health AI platform, building it one disease and one region as a time. The proof of concept will start with malaria, which kills 1.2 million people annually. While there are a number of interventions available today to fight malaria outbreaks, including vaccines, Claypool said that the prediction of hot spots and the decision support needed to intervene is essential. The next major milestone would be to provide data-driven diagnostics and interventions across the globe for other disease conditions.

“It’s an ambitious but achievable vision. It needs the right partnerships, trust, and vision to make this a reality, and reduce transmission of disease and save lives globally,” she said.

Addressing humanitarian challenges is a growing R&D focus at Lincoln Laboratory. Last fall, the organization established a new research division, Biotechnology and Human Systems, to further explore global issues around climate change, health, and humanitarian assistance. 

“Our goal is to build collaboration and communication with a broader community around all of these topics. They are all terribly important and complex and require significant global effort to make a difference,” Pitts says.

The next event in this series will take place in September. More

MIT researchers have created the first fiber with digital capabilities, able to sense, store, analyze, and infer activity after being sewn into a shirt.

Yoel Fink, who is a professor of material sciences and engineering, a Research Laboratory of Electronics principal investigator, and the senior author on the study, says digital fibers expand the possibilities for fabrics to uncover the context of hidden patterns in the human body that could be used for physical performance monitoring, medical inference, and early disease detection.

Or, you might someday store your wedding music in the gown you wore on the big day — more on that later.

Fink and his colleagues describe the features of the digital fiber today in Nature Communications. Until now, electronic fibers have been analog — carrying a continuous electrical signal — rather than digital, where discrete bits of information can be encoded and processed in 0s and 1s.

“This work presents the first realization of a fabric with the ability to store and process data digitally, adding a new information content dimension to textiles and allowing fabrics to be programmed literally,” Fink says.

MIT PhD student Gabriel Loke and MIT postdoc Tural Khudiyev are the lead authors on the paper. Other co-authors MIT postdoc Wei Yan; MIT undergraduates Brian Wang, Stephanie Fu, Ioannis Chatziveroglou, Syamantak Payra, Yorai Shaoul, Johnny Fung, and Itamar Chinn; John Joannopoulos, the Francis Wright Davis Chair Professor of Physics and director of the Institute for Soldier Nanotechnologies at MIT; Harrisburg University of Science and Technology master’s student Pin-Wen Chou; and Rhode Island School of Design Associate Professor Anna Gitelson-Kahn. The fabric work was facilitated by Professor Anais Missakian, who holds the Pevaroff-Cohn Family Endowed Chair in Textiles at RISD.

Memory and more

The new fiber was created by placing hundreds of square silicon microscale digital chips into a preform that was then used to create a polymer fiber. By precisely controlling the polymer flow, the researchers were able to create a fiber with continuous electrical connection between the chips over a length of tens of meters.

The fiber itself is thin and flexible and can be passed through a needle, sewn into fabrics, and washed at least 10 times without breaking down. According to Loke, “When you put it into a shirt, you can’t feel it at all. You wouldn’t know it was there.”

Making a digital fiber “opens up different areas of opportunities and actually solves some of the problems of functional fibers,” he says.

For instance, it offers a way to control individual elements within a fiber, from one point at the fiber’s end. “You can think of our fiber as a corridor, and the elements are like rooms, and they each have their own unique digital room numbers,” Loke explains. The research team devised a digital addressing method that allows them to “switch on” the functionality of one element without turning on all the elements.

A digital fiber can also store a lot of information in memory. The researchers were able to write, store, and read information on the fiber, including a 767-kilobit full-color short movie file and a 0.48 megabyte music file. The files can be stored for two months without power.

When they were dreaming up “crazy ideas” for the fiber, Loke says, they thought about applications like a wedding gown that would store digital wedding music within the weave of its fabric, or even writing the story of the fiber’s creation into its components.

Fink notes that the research at MIT was in close collaboration with the textile department at RISD led by Missakian.  Gitelson-Kahn incorporated the digital fibers into a knitted garment sleeve, thus paving the way to creating the first digital garment.

On-body artificial intelligence

The fiber also takes a few steps forward into artificial intelligence by including, within the fiber memory, a neural network of 1,650 connections. After sewing it around the armpit of a shirt, the researchers used the fiber to collect 270 minutes of surface body temperature data from a person wearing the shirt, and analyze how these data corresponded to different physical activities. Trained on these data, the fiber was able to determine with 96 percent accuracy what activity the person wearing it was engaged in.

Adding an AI component to the fiber further increases its possibilities, the researchers say. Fabrics with digital components can collect a lot of information across the body over time, and these “lush data” are perfect for machine learning algorithms, Loke says.

“This type of fabric could give quantity and quality open-source data for extracting out new body patterns that we did not know about before,” he says.

With this analytic power, the fibers someday could sense and alert people in real-time to health changes like a respiratory decline or an irregular heartbeat, or deliver muscle activation or heart rate data to athletes during training.

The fiber is controlled by a small external device, so the next step will be to design a new chip as a microcontroller that can be connected within the fiber itself.

“When we can do that, we can call it a fiber computer,” Loke says.

This research was supported by the U.S. Army Institute of Soldier Nanotechnologies, National Science Foundation, the U.S. Army Research Office, the MIT Sea Grant, and the Defense Threat Reduction Agency. More

Improved computing power and an exponential increase in data have helped fuel the rapid rise of artificial intelligence. But as AI systems become more sophisticated, they’ll need even more computational power to address their needs, which traditional computing hardware most likely won’t be able to keep up with. To solve the problem, MIT spinout Lightelligence is developing the next generation of computing hardware.

The Lightelligence solution makes use of the silicon fabrication platform used for traditional semiconductor chips, but in a novel way. Rather than building chips that use electricity to carry out computations, Lightelligence develops components powered by light that are low energy and fast, and they might just be the hardware we need to power the AI revolution. Compared to traditional architectures, the optical chips made by Lightelligence offer orders of magnitude improvement in terms of high speed, low latency, and low power consumption.

In order to perform arithmetic operations, electronic chips need to combine tens, sometimes hundreds, of logic gates. To perform this process requires the electronic chip transistors to switch off and on for multiple clock periods. Every time a logic gate transistor switches, it generates heat and consumes power.

Not so with the chips produced by Lightelligence. In the optical domain, arithmetic computations are done with physics instead of with logic gate transistors that require multiple clocks. More clocks means a slower time to get a result. “We precisely control how the photons interact with each other inside the chip,” says Yichen Shen PhD ’16, co-founder and CEO of Lightelligence. “It’s just light propagating through the chip, photons interfering with each other. The nature of the interference does the mathematics that we want it to do.”

This process of interference generates very little heat, which means Shen’s optical computing chips enable much lower power consumption than their electron-powered counterparts. Shen points out that we’ve made use of fiber optics for long-distance communication for decades. “Think of the optical fibers spread across the bottom of the Pacific Ocean, and the light propagating through thousands of kilometers without losing much power. Lightelligence is bringing this concept for long-distance communication to on-chip compute.”

With most forecasters projecting an end to Moore’s Law sometime in 2025, Shen believes his optic-driven solution is poised to address many of the computational challenges of the future. “We’re changing the fundamental way computing is done, and I think we’re doing it at the right time in history,” says Shen. “We believe optics is going to be the next computing platform, at least for linear operations like AI.”

To be clear, Shen does not envision optics replacing the entire electronic computing industry. Rather, Lightelligence aims to accelerate certain linear algebra operations to perform quick, power-efficient tasks like those found in artificial neural networks.

Much of AI compute happens in the cloud at data centers like the ones supporting Amazon or Microsoft. Because AI algorithms are computationally intensive, AI compute takes up a large percentage of data center capacity. Picture tens of thousands of servers, running continuously, burning millions of dollars worth of electricity. Now imagine replacing some of those conventional servers with Lightelligence servers that burn much less power at a fraction of the cost. “Our optical chips would greatly reduce the cost of data centers, or, put another way, greatly increase the computational capability of those data centers for AI applications,” says Shen.  

And what about self-driving vehicles? They rely on cameras and AI computation to make quick decisions. But a conventional digital electronic chip doesn’t “think” quickly enough to make the decisions necessary at high speeds. Faster computational imaging leads to faster decision-making. “Our chip completes these decision-making tasks at a fraction of the time of regular chips, which would enable the AI system within the car to make much quicker decisions and more precise decisions, enabling safer driving,” says Shen.

Lightelligence boasts an all-MIT founding team, supported by 40 technical experts, including machine learning pioneers, leading photonic researchers, and semiconductor industry veterans intent on revolutionizing computing technology. Shen did his PhD work in the Department of Physics with professors Marin Soljajic and John Joannoupolos, where he developed an interest in the intersection of photonics and AI. “I realized that computation is a key enabler of modern artificial intelligence, and faster computing hardware would be needed to complement the growth of faster, smarter AI algorithms,” he says.

Lightelligence was founded in 2017 when Shen teamed up with Soljajic and two other MIT alumni. Fellow co-founder Huaiyu Meng SM ’14, PhD ’18 received his doctorate in electrical engineering and now serves as Lightelligence’s vice president of photonics. Rounding out the founding team is Spencer Powers MBA ’16. Powers, who received his MBA from MIT Sloan School of Management, is also a Lightelligence board member with extensive experience in the startup world.

Shen and his team are not alone in this new field of optical computing, but they do have key advantages over their competitors. First off, they invented the technology at the Institute. Lightelligence is also the first company to have built a complete system of optical computing hardware, which it accomplished in April 2019. Shen is self-assured in the innovation potential of Lightelligence and what it could mean for the future, regardless of the competition. “There are new stories of teams working in this space, but we’re not only the first, we’re the fastest in terms of execution. I stand by that,” he says.

But there’s another reason Shen’s not worried about the competition. He likens this stage in the evolution of the technology to the era when transistors were replacing vacuum tubes. Several transistor companies were making the leap, but they weren’t competing with each other so much as they were innovating to compete with the incumbent industry. “Having more competitors doing optical computing is good for us at this stage,” says Shen. “It makes for a louder voice, a bigger community to expand and enhance the whole ecosystem for optical computing.”

By 2021, Shen anticipates that Lightelligence will have de-risked 80-90 percent of the technical challenges necessary for optical computing to be a viable commercial product. In the meantime, Lightelligence is making the most of its status as the newest member of the MIT Startup Exchange accelerator, STEX25, building deep relationships with tier-one customers on several niche applications where there is a pressing need for high-performance hardware, such as data centers and manufacturers. More

Health care is at a junction, a point where artificial intelligence tools are being introduced to all areas of the space. This introduction comes with great expectations: AI has the potential to greatly improve existing technologies, sharpen personalized medicines, and, with an influx of big data, benefit historically underserved populations.

But in order to do those things, the health care community must ensure that AI tools are trustworthy, and that they don’t end up perpetuating biases that exist in the current system. Researchers at the MIT Abdul Latif Jameel Clinic for Machine Learning in Health (Jameel Clinic), an initiative to support AI research in health care, call for creating a robust infrastructure that can aid scientists and clinicians in pursuing this mission.

Fair and equitable AI for health care

The Jameel Clinic recently hosted the AI for Health Care Equity Conference to assess current state-of-the-art work in this space, including new machine learning techniques that support fairness, personalization, and inclusiveness; identify key areas of impact in health care delivery; and discuss regulatory and policy implications.

Nearly 1,400 people virtually attended the conference to hear from thought leaders in academia, industry, and government who are working to improve health care equity and further understand the technical challenges in this space and paths forward.

During the event, Regina Barzilay, the School of Engineering Distinguished Professor of AI and Health and the AI faculty lead for Jameel Clinic, and Bilal Mateen, clinical technology lead at the Wellcome Trust, announced the Wellcome Fund grant conferred to Jameel Clinic to create a community platform supporting equitable AI tools in health care.

The project’s ultimate goal is not to solve an academic question or reach a specific research benchmark, but to actually improve the lives of patients worldwide. Researchers at Jameel Clinic insist that AI tools should not be designed with a single population in mind, but instead be crafted to be reiterative and inclusive, to serve any community or subpopulation. To do this, a given AI tool needs to be studied and validated across many populations, usually in multiple cities and countries. Also on the project wish list is to create open access for the scientific community at large, while honoring patient privacy, to democratize the effort.

“What became increasingly evident to us as a funder is that the nature of science has fundamentally changed over the last few years, and is substantially more computational by design than it ever was previously,” says Mateen.

The clinical perspective

This call to action is a response to health care in 2020. At the conference, Collin Stultz, a professor of electrical engineering and computer science and a cardiologist at Massachusetts General Hospital, spoke on how health care providers typically prescribe treatments and why these treatments are often incorrect.

In simplistic terms, a doctor collects information on their patient, then uses that information to create a treatment plan. “The decisions providers make can improve the quality of patients’ lives or make them live longer, but this does not happen in a vacuum,” says Stultz.

Instead, he says that a complex web of forces can influence how a patient receives treatment. These forces go from being hyper-specific to universal, ranging from factors unique to an individual patient, to bias from a provider, such as knowledge gleaned from flawed clinical trials, to broad structural problems, like uneven access to care.

Datasets and algorithms

A central question of the conference revolved around how race is represented in datasets, since it’s a variable that can be fluid, self-reported, and defined in non-specific terms.

“The inequities we’re trying to address are large, striking, and persistent,” says Sharrelle Barber, an assistant professor of epidemiology and biostatistics at Drexel University. “We have to think about what that variable really is. Really, it’s a marker of structural racism,” says Barber. “It’s not biological, it’s not genetic. We’ve been saying that over and over again.”

Some aspects of health are purely determined by biology, such as hereditary conditions like cystic fibrosis, but the majority of conditions are not straightforward. According to Massachusetts General Hospital oncologist T. Salewa Oseni, when it comes to patient health and outcomes, research tends to assume biological factors have outsized influence, but socioeconomic factors should be considered just as seriously.

Even as machine learning researchers detect preexisting biases in the health care system, they must also address weaknesses in algorithms themselves, as highlighted by a series of speakers at the conference. They must grapple with important questions that arise in all stages of development, from the initial framing of what the technology is trying to solve to overseeing deployment in the real world.

Irene Chen, a PhD student at MIT studying machine learning, examines all steps of the development pipeline through the lens of ethics. As a first-year doctoral student, Chen was alarmed to find an “out-of-the-box” algorithm, which happened to project patient mortality, churning out significantly different predictions based on race. This kind of algorithm can have real impacts, too; it guides how hospitals allocate resources to patients.

Chen set about understanding why this algorithm produced such uneven results. In later work, she defined three specific sources of bias that could be detangled from any model. The first is “bias,” but in a statistical sense — maybe the model is not a good fit for the research question. The second is variance, which is controlled by sample size. The last source is noise, which has nothing to do with tweaking the model or increasing the sample size. Instead, it indicates that something has happened during the data collection process, a step way before model development. Many systemic inequities, such as limited health insurance or a historic mistrust of medicine in certain groups, get “rolled up” into noise.

“Once you identify which component it is, you can propose a fix,” says Chen.

Marzyeh Ghassemi, an assistant professor at the University of Toronto and an incoming professor at MIT, has studied the trade-off between anonymizing highly personal health data and ensuring that all patients are fairly represented. In cases like differential privacy, a machine-learning tool that guarantees the same level of privacy for every data point, individuals who are too “unique” in their cohort started to lose predictive influence in the model. In health data, where trials often underrepresent certain populations, “minorities are the ones that look unique,” says Ghassemi.

“We need to create more data, it needs to be diverse data,” she says. “These robust, private, fair, high-quality algorithms we’re trying to train require large-scale data sets for research use.”

Beyond Jameel Clinic, other organizations are recognizing the power of harnessing diverse data to create more equitable health care. Anthony Philippakis, chief data officer at the Broad Institute of MIT and Harvard, presented on the All of Us research program, an unprecedented project from the National Institutes of Health that aims to bridge the gap for historically under-recognized populations by collecting observational and longitudinal health data on over 1 million Americans. The database is meant to uncover how diseases present across different sub-populations.

One of the largest questions of the conference, and of AI in general, revolves around policy. Kadija Ferryman, a cultural anthropologist and bioethicist at New York University, points out that AI regulation is in its infancy, which can be a good thing. “There’s a lot of opportunities for policy to be created with these ideas around fairness and justice, as opposed to having policies that have been developed, and then working to try to undo some of the policy regulations,” says Ferryman.

Even before policy comes into play, there are certain best practices for developers to keep in mind. Najat Khan, chief data science officer at Janssen R&D, encourages researchers to be “extremely systematic” when choosing datasets. Even large, common datasets contain inherent bias.

Even more fundamental is opening the door to a diverse group of future researchers.

“We have to ensure that we are developing folks, investing in them, and having them work on really important problems that they care about,” says Khan. “You’ll see a fundamental shift in the talent that we have.”

The AI for Health Care Equity Conference was co-organized by MIT’s Jameel Clinic; Department of Electrical Engineering and Computer Science; Institute for Data, Systems, and Society; Institute for Medical Engineering and Science; and the MIT Schwarzman College of Computing. More

Disinformation campaigns are not new — think of wartime propaganda used to sway public opinion against an enemy. What is new, however, is the use of the internet and social media to spread these campaigns. The spread of disinformation via social media has the power to change elections, strengthen conspiracy theories, and sow discord.

Steven Smith, a staff member from MIT Lincoln Laboratory’s Artificial Intelligence Software Architectures and Algorithms Group, is part of a team that set out to better understand these campaigns by launching the Reconnaissance of Influence Operations (RIO) program. Their goal was to create a system that would automatically detect disinformation narratives as well as those individuals who are spreading the narratives within social media networks. Earlier this year, the team published a paper on their work in the Proceedings of the National Academy of Sciences and they received an R&D 100 award last fall.

The project originated in 2014 when Smith and colleagues were studying how malicious groups could exploit social media. They noticed increased and unusual activity in social media data from accounts that had the appearance of pushing pro-Russian narratives.

“We were kind of scratching our heads,” Smith says of the data. So the team applied for internal funding through the laboratory’s Technology Office and launched the program in order to study whether similar techniques would be used in the 2017 French elections.

In the 30 days leading up to the election, the RIO team collected real-time social media data to search for and analyze the spread of disinformation. In total, they compiled 28 million Twitter posts from 1 million accounts. Then, using the RIO system, they were able to detect disinformation accounts with 96 percent precision.

What makes the RIO system unique is that it combines multiple analytics techniques in order to create a comprehensive view of where and how the disinformation narratives are spreading.

“If you are trying to answer the question of who is influential on a social network, traditionally, people look at activity counts,” says Edward Kao, who is another member of the research team. On Twitter, for example, analysts would consider the number of tweets and retweets. “What we found is that in many cases this is not sufficient. It doesn’t actually tell you the impact of the accounts on the social network.”

As part of Kao’s PhD work in the laboratory’s Lincoln Scholars program, a tuition fellowship program, he developed a statistical approach — now used in RIO — to help determine not only whether a social media account is spreading disinformation but also how much the account causes the network as a whole to change and amplify the message.

Erika Mackin, another research team member, also applied a new machine learning approach that helps RIO to classify these accounts by looking into data related to behaviors such as whether the account interacts with foreign media and what languages it uses. This approach allows RIO to detect hostile accounts that are active in diverse campaigns, ranging from the 2017 French presidential elections to the spread of Covid-19 disinformation.

Another unique aspect of RIO is that it can detect and quantify the impact of accounts operated by both bots and humans, whereas most automated systems in use today detect bots only. RIO also has the ability to help those using the system to forecast how different countermeasures might halt the spread of a particular disinformation campaign.

The team envisions RIO being used by both government and industry as well as beyond social media and in the realm of traditional media such as newspapers and television. Currently, they are working with West Point student Joseph Schlessinger, who is also a graduate student at MIT and a military fellow at Lincoln Laboratory, to understand how narratives spread across European media outlets. A new follow-on program is also underway to dive into the cognitive aspects of influence operations and how individual attitudes and behaviors are affected by disinformation.

“Defending against disinformation is not only a matter of national security, but also about protecting democracy,” says Kao. More

Anesthestic drugs act on the brain, but most anesthesiologists rely on heart rate, respiratory rate, and movement to infer whether surgery patients remain unconscious to the desired degree. In a new study, a research team based at MIT and Massachusetts General Hospital shows that a straightforward artificial intelligence approach, attuned to the kind of anesthetic being used, can yield algorithms that assess unconsciousness in patients based on brain activity with high accuracy and reliability.

“One of the things that is foremost in the minds of anesthesiologists is ‘Do I have somebody who is lying in front of me who may be conscious and I don’t realize it?’ Being able to reliably maintain unconsciousness in a patient during surgery is fundamental to what we do,” says senior author Emery N. Brown, the Edward Hood Taplin Professor in The Picower Institute for Learning and Memory and the Institute for Medical Engineering and Science at MIT, and an anesthesiologist at MGH. “This is an important step forward.”

More than providing a good readout of unconsciousness, Brown adds, the new algorithms offer the potential to allow anesthesiologists to maintain it at the desired level while using less drug than they might administer when depending on less direct, accurate, and reliable indicators. That can improve patient’s post-operative outcomes, such as delirium.

“We may always have to be a little bit ‘overboard,’” says Brown, who is also a professor at Harvard Medical School. “But can we do it with sufficient accuracy so that we are not dosing people more than is needed?”

Used to drive an infusion pump, for instance, algorithms could help anesthesiologists precisely throttle drug delivery to optimize a patient’s state and the doses they are receiving.

Artificial intelligence, real-world testing

To develop the technology to do so, postdocs John Abel and Marcus Badgeley led the study, published in PLOS ONE, in which they trained machine learning algorithms on a remarkable dataset the lab gathered back in 2013. In that study, 10 healthy volunteers in their 20s underwent anesthesia with the commonly used drug propofol. As the dose was methodically raised using computer-controlled delivery, the volunteers were asked to respond to a simple request until they couldn’t anymore. Then when they were brought back to consciousness as the dose was later lessened, they became able to respond again. All the while, neural rhythms reflecting their brain activity were recorded with electroencephalogram (EEG) electrodes, providing a direct, real-time link between measured brain activity and exhibited unconsciousness.

In the new work, Abel, Badgeley, and the team trained versions of their AI algorithms, based on different underlying statistical methods, on more than 33,000 2-second-long snippets of EEG recordings from seven of the volunteers. This way the algorithms could “learn” the difference between EEG readings predictive of consciousness and unconsciousness under propofol. Then the researchers tested the algorithms in three ways.

First, they checked whether their three most promising algorithms accurately predicted unconsciousness when applied to EEG activity recorded from the other three volunteers of the 2013 study. They did.

Then they used the algorithms to analyze EEG recorded from 27 real surgery patients who received propofol for general anesthesia. Even though the algorithms were now being applied to data gathered from a “noisier” real-world surgical setting where the rhythms were also being measured with different equipment, the algorithms still distinguished unconsciousness with higher accuracy than other studies have shown. The authors even highlight one case in which the algorithms were able to detect a patient’s decreasing level of unconsciousness several minutes before the actual attending anesthesiologist did, meaning that if it had been in use during the surgery itself, it could have provided an accurate and helpful early warning.

As a third test, the team applied the algorithms to EEG recordings from 17 surgery patients who were anesthetized with sevoflurane. Though sevoflurane is different from propofol and is inhaled rather than infused, it works in a similar manner, by binding to the same GABA-A receptors on the same key types of brain cells. The team’s algorithms again performed with high, though somewhat-reduced accuracy, suggesting that their ability to classify unconsciousness carried over reliably to another anesthetic drug that works in a similar way.

The ability to predict unconsciousness across different drugs with the same mechanism of action is key, the authors said. One of the main flaws with current EEG-based systems for monitoring consciousness, they said, is that they don’t distinguish among drug classes, even though different categories of anesthesia drugs work in very different ways, producing distinct EEG patterns. They also don’t adequately account for known age differences in brain response to anesthesia. These limitations on their accuracy have also limited their clinical use.

In the new study, while the algorithms trained on 20-somethings applied well to cohorts of surgery patients whose average age skewed significantly older and varied more widely, the authors acknowledge that they want to train algorithms distinctly for use with children or seniors. They can also train new algorithms to apply specifically for other kinds of drugs with different mechanisms of action. Altogether, a suite of well-trained and attuned algorithms could provide high accuracy that accounts for patient age and the drug in use.

Abel says the team’s approach of framing the problem as a matter of predicting consciousness via EEG for a specific class of drugs made the machine learning approach very simple to implement and extend.

“This is a proof of concept showing that now we can go and say let’s look at an older population or let’s look at a different kind of drug,” he says. “Doing this is simple if you set it up the right way.”

The resulting algorithms aren’t even computationally demanding. The authors noted that for a given 2 seconds of EEG data, the algorithms could make an accurate prediction of consciousness in less than a tenth of a second running on just a standard MacBook Pro computer.

The lab is already building on the findings to refine the algorithms further, Brown says. He says he also wants to expand testing to hundreds more cases to further confirm their performance, and also to determine whether wider distinctions may begin to emerge among the different underlying statistical models the team employed.

In addition to Brown, Abel and Badgeley, the paper’s other authors are Benyamin Meschede-Krasa, Gabriel Schamberg, Indie Garwood, Kimaya Lecamwasam, Sourish Chakravarty, David Zhou, Matthew Keating, and Patrick Purdon.

Funding for the study came from the National Institutes of Health, The JPB Foundation, A Guggenheim Fellowship for Applied Mathematics, and Massachusetts General Hospital. More