Patterns to Predictions

By Katherine Shaver

Sabyasachi Sen and other physicians who treat diabetes try to detect abnormal metabolic changes as early as possible in the hopes that patients can change their diet and lifestyle to avoid progressing to full-blown Type 2 diabetes.

Traditionally, doctors advise patients with elevated blood glucose levels—a condition known as prediabetes—to cut back on carbs and sugar and step up the exercise. 

Even so, Sen said, people with prediabetes have a 25 percent chance of developing Type 2 diabetes within three to five years—a diagnosis that typically requires significant lifestyle changes and medication while significantly raising their risk for a heart attack, stroke and other serious complications.

The problem: Doctors don’t know which of their prediabetes patients will end up with a diabetes diagnosis, or which drug or other approach would work best to prevent it in that group.

“Doctors are asking what a patient’s chances are of progressing from prediabetes to diabetes,” said Sen, a GW professor of medicine and endocrinology and chief of endocrinology at the Washington VA Medical Center. “What medication would they respond to? Is it time to start them on medication now, or can they wait for another couple of months or years to see if changes in diet and exercise work?”

Artificial intelligence software developed by Sen’s GW colleagues could soon remove that guesswork. The AI tool, called PredictMod, analyzes reams of data culled from published medical studies about different diseases, drugs and other interventions. It then finds patterns in patient outcomes by interpreting study participants’ health data. That data can include people’s race and ethnicity, blood glucose levels, heart rates, diets, activity and stress levels, gut bacteria and other health conditions. 

PredictMod is designed to be intuitive and user-friendly while safeguarding patient privacy and strictly complying with state, federal and institutional regulations. The idea is to allow doctors and other clinicians without expertise in health data science to securely upload a patient’s information, such as bloodwork results or DNA sequencing, in a protected environment that keeps it confidential and ensures compliance with regulations. It then quickly predicts the most effective drug or other treatment based on the outcomes of similar patients. 

In addition to prediabetes and diabetes, models on the PredictMod website—which is free and accessible to anyone—can predict the most effective drugs and other therapies for epilepsy and ovarian cancer.

It’s the future of medicine, in which clinicians harness the power of AI to predict disease progression and provide highly personalized, targeted care.

“The potential is huge–it’s almost rethinking the way we do science,” said PredictMod developer Raja Mazumder, professor of biochemistry and molecular medicine in the GW School of Medicine and Health Sciences. He’s also co-director of GW’s McCormick Genomic & Proteomic Center.

In the case of prediabetes, Sen, Mazumder and other researchers recently found that PredictMod was better at detecting abnormal blood sugar fluctuations than an HbA1c test. That common blood test shows a patient’s average blood glucose levels over the previous two to three months.

In their study, which is undergoing peer review, the researchers used PredictMod to interpret data from an earlier study of more than 1,000 people who wore continuous glucose monitors, typically on the back of their upper arms. The monitors measured their blood glucose every five minutes, around the clock. 

PredictMod found that 57 percent of the study participants who had reported themselves as “healthy,” most likely based on their HbA1c level, actually showed wide swings in their glucose levels after meals, similar to people with prediabetes. Diagnosing the problem earlier, researchers said, would allow for earlier—and more effective—interventions.

Sen said researchers are further developing the model to predict which people with prediabetes will go on to develop full-blown diabetes and determine which medications will work best for them and when to begin treatment.

With one in three Americans having prediabetes, Sen said, the U.S. health care system could save billions of dollars by getting the most effective treatments to the most at-risk patients as early as possible.

"We're almost in the midst of a prediabetes and obesity epidemic,” Sen said. “A metabolic disease prediction model is almost critical at this point.”

Mazumder and his team developed PredictMod in 2020 for a competition hosted by the U.S. Department of Veterans Affairs to develop AI tools to improve veterans’ health and well-being. PredictMod took fifth place out of 44 teams for its prediabetes work.

PredictMod software also was used in a 2022 study that analyzed data from the gut microbiomes of children with drug-resistant epilepsy to predict which ones would have fewer seizures on a ketogenic diet, which is high-fat, moderate-protein and very low-carb.

The AI tool is particularly helpful in medical research, Mazumder said, because the compelling insights it can glean from small datasets in pilot studies can be used to justify larger studies.

Lori Krammer, a GW research associate working to refine PredictMod, said she and its other developers are sensitive to privacy concerns about patient health data. She noted that PredictMod relies on publicly available health datasets and those culled from published research.

The website also does not store any patient information uploaded to its models, Krammer said. However, the software does allow clinicians and researchers to refine the models and create new ones for different diseases.

“It’s meant to be a very collaborative space,” Krammer said.

She and Mazumder have set their sights on the next challenge: Getting more health data to make the models increasingly sophisticated–and useful.

Ideally, they said, the models will incorporate data now off limits in millions of patients’ electronic health records kept by doctors and hospitals across the country. Mazumder is co-principal investigator on a national project called FEAST to develop the infrastructure needed to make health data AI-ready without compromising patient privacy.

When that happens, Mazumder said, personalized medicine will reach a whole new level.

“There will be an explosion of these models,” he said, “where everybody will be able to have a model with patients like them to help guide their treatment.”

By John DiConsiglio

It’s July 2025 and the Federal Reserve is facing a moment of crisis. As its Federal Open Market Committee (FOMC) meeting convenes, inflation remains stubborn, the labor numbers show signs of strain and global financial markets are parsing clues for insights on fiscal policy. Looming over the already-charged summit is intense political pressure from President Donald Trump to slash interest rates.

At the center of the meeting is beleaguered Fed Chair Jerome Powell, guiding the agenda amid speculation that his ongoing feud with Trump will lead to his dismissal. Federal Reserve Governor Christopher Waller, a traditional interest rate hawk and a rumored candidate for Powell’s job, seems to have softened his stance on Trump’s demand for cuts. Fellow Board of Governors member Michelle W. Bowman is also sending dove-ish rate signals, a notable shift from her past hold-the-line position.

The stakes are high. The atmosphere is tense.

But the meeting isn’t real.

Welcome to the SIM Fed, an artificial intelligence simulation of a pivotal Federal Reserve meeting. Co-created by George Washington University Professor of Economics and International Affairs Tara M. Sinclair, “FOMC In Silico,” as she dubs it, is a virtual laboratory that gives economists, Fed watchers and market trackers a window into one of the most consequential decision-making processes in the world. 

The framework—built as a multi-agent platform powered by an LLM and guided by text prompt commands that Sinclair likens to the “Oregon Trail” computer game—features AI-generated “personas” of actual Fed members. Grounded in real-time data, it models policy deliberations, disagreements and decisions that simulate the drama inside the Fed’s D.C. headquarters—but within a computer database.

“The idea is to use LLMs and publicly available information about these very public figures to create personas of FOMC members,” said Sinclair, who created FOMC In Silico with Stanford Research Scientist Sophia Kazinnik. 

Indeed, the SIM Fed members are constructed from an archive of information: speeches, voting records, biographies, policy positions—even their communication styles. Sinclair and  Kazinnik then add economic data into the mix—from unemployment rates to GDP growth to inflation figures—and stir the pot by sprinkling in high-stakes scenarios like boiling-point political pressure. 

“There’s no way I can get Chair Powell and the rest of the FOMC to come over to GW for a day and let me run them through different scenario exercises,” said Sinclair, who chairs the Economics Department at the Columbian College of Arts and Sciences. “But we can create personas with flexible personalities that give us a close simulation of the real person.”

Comparing FOMC In Silico to a flight simulator for monetary policy, Sinclair launched a test run hours before the actual July FOMC meeting. After building the AI persona profiles, she and Kazinnik handed them the same economic data sheets they’d find in their D.C. Eccles Building boardroom. (Days later, they updated the simulation with newly-released labor market figures.) They then introduced the political pressure effect, instructing the personas to reflect the real-world tensions over interest rate cuts.

The AI committee nearly mirrored the real-life policymakers’ decisions. Both the AI and actual FOMC kept interest rates unchanged in a range of 4.25% to 4.5%. The AI version landed at a 4.42% midpoint.

Perhaps more impressively, the AI simulation recorded rare dissent among Fed members. In reality, both Waller and Bowman voted for lower rates—the first dual dissent for rate cuts since 1993. Applying the political influences, Sinclair explained, swayed arguments and fragmented AI personas in the same direction as their real-life counterparts.

“This simulation shows that the Federal Reserve is only partially insulated from politics,” Sinclair and Kazinnik noted in a working paper. “Outside scrutiny can shape internal decision-making, even in an institution guided by formal rules.”

The researchers have continued to run the simulation for each subsequent FOMC meeting—they convene every six to eight weeks—and have also applied the model to past sessions dating back to 2000. Its historical record has been stellar, predicting the policy rate within 25 basis points in 93% of past meetings.

In its current format, Sinclair stresses that the model is a simulator, not a forecaster. For now the goal is to grasp how the committee sets policy, not necessarily predict its decisions in advance. “It helps us better understand how all of these mechanisms work, what [the FOMC] is thinking and what kinds of influences are at play,” she said.

As they finalize their research paper, Sinclair and Kazinnik eventually plan to make the tool widely available online. In addition to economists and financial market professionals, she envisions a host of experts adapting the framework to their own fields—like political scientists simulating U.S. Supreme Court deliberations. “We’ve built an entire laboratory and it can be used for many different types of experiments,” Sinclair said.

For her purposes, Sinclair hopes the methodology will move closer to forecasting—perhaps replicating the Fed's quarterly economic projections. Meanwhile, as the traditionally cautious economics field warms to technology innovations, Sinclair sees FOMC In Silico as a bridge between market fervor and AI ascendance. “The tool meets the moment,” she said.

Likewise, her own CCAS department features an AI Economics program within its Center for Economic Research. And in class discussions, Sinclair said the Fed simulator has sparked enthusiasm among students.

“It’s impressive how fast [economists] have adopted these tools and the impact they’ve had on education,” she said. “Now we don’t just teach students how to code—we teach them how to Claude code.”

By Katherine Shaver

To measure the health effects of artificial sweeteners found in diet sodas and other drinks and food, Allison Sylvetsky must analyze trillions of microbes living in the human gut.

Sylvetsky, a former associate professor at the George Washington University’s Milken Institute School of Public Health, compares the DNA of microorganisms contained in stool samples from people who consume the sweeteners against those who don’t. That allows her to understand ways that sweeteners can influence insulin sensitivity, inflammation and other health conditions. 

Such complex analysis requires collaboration with bioinformatics experts and can take up to a year or more, adding cost and time to already tight budgets and lengthy studies. 

“If you can speed up the analysis part and have findings you can communicate, you can advance the field faster,” said Sylvetsky, now a professor and chair of the Department of Nutrition at the University of Rhode Island.

That’s why Sylvetsky is excited about the work of her former GW public health colleagues, Professors Ali Rahnavard and Keith A. Crandall, who are using artificial intelligence to make the analysis of DNA sequences more accessible, quicker, less expensive and more accurate.

Rahnavard, associate professor in the Department of Biostatistics and Bioinformatics, and Crandall, director of GW's Computational Biology Institute, have developed a family of genomic language models (gLMs) that “read” DNA code like a sentence. The software recognizes patterns in DNA’s “chemical letters” that have biological meaning, much like large language models, such as ChatGPT, interpret text based on patterns in words.

Their AI tool, called microCafe, comes as scientists and tech firms seek to harness the rapidly evolving power of AI to interpret vast amounts of data produced by modern DNA sequencing machines. That work is in high demand as the costs of DNA sequencing have fallen dramatically. The process determines the order of the chemical building blocks of DNA, which can reveal mutations and other variations in an organism’s genetic instructions.

Their AI tool is designed to allow researchers and hospital lab staff to do in a few clicks on their laptops the kind of specialized analysis long done by bioinformatics experts using costly, high-powered computers that laboriously probe enormous DNA databases.

“It's going to open a completely new avenue for research,” Rahnavard said. “We have massive amounts of [DNA] data. The problem is we need to get insights from it.”

Beyond deciphering the gut microbiome, scientists are using DNA sequencing to detect and analyze infectious diseases, differentiate types of cancers, track viruses and personalize medical treatments, while also revealing genetic risk for chronic diseases, such as heart disease and Alzheimer’s. Beyond human health, it can be used to protect crops and animals, as well as study and monitor environmental changes linked to climate and pollution.

While DNA sequencing has opened the flood gates to vast amounts of genomic information, the terabytes of raw data pose their own problem — how to make sense of it all. AI tools like gLMs could be the key to unlocking discoveries hidden in the data.

Because a gLM learns as it goes, Rahnavard and Crandall said, it can interpret genomes it hasn’t seen before, which helps scientists discover new biological markers of disease or environmental change.

The technology grew out of their 2020 research into DNA changes in the virus that causes COVID-19. The research, funded by the National Science Foundation and National Institutes of Health, came as medical experts and public health officials sought to understand why some people infected during the pandemic had no symptoms or a few sniffles, even as others died or needed to be hospitalized. To help answer that question, they developed machine learning approaches to integrate and analyze large volumes of diverse data coming from labs, hospitals, and healthcare systems around the world.

At the same time, computer scientists were developing large language models that could generate new content by recognizing patterns of words and sentences in text. Rahnavard, Crandall and others started considering how a similar technology could be applied to “reading” DNA.

“The challenge was how we could treat DNA as a language,” Rahnavard said, “because the technology was there.”

Rahnavard and Crandall said their genomic language models, which are smaller and more efficient than other gLMs, provide more accurate insights because they’re trained on more targeted DNA data to answer more specific questions. For example, to combat the growing problem of bacterial infections resistant to common antibiotics, their team has trained a model to identify drug-resistant microbe variants.

Rahnavard and Crandall are working with GW’s Technology Commercialization Office to patent their AI technology and sell microCafe through their company, seqSight. Potential customers include companies that provide DNA sequencing machines, as well as universities, hospital labs and biotech companies that perform DNA sequencing, they said.

They recently received a big boost from a NVIDIA Academic Grant awarding them time on NVIDIA’s sophisticated computing infrastructure to train their AI models and from an NSF I-Corps award to accelerate the company’s development.

They’re also continuing to develop the next generation of scientists to keep moving the technology forward, including graduate and undergraduate students in the new Health Data Science programs at the Milken Institute School of Public Health and D.C.-area high school students in the Path2Max program.

“It’s really exciting to involve the students at all levels, from high school through Ph.D. and even postdocs, in this exciting research environment that we have developed in health data science,” Crandall said. “There’s huge potential to both build novel tools for major research and ensure the next generation of health data scientists can leverage the latest in AI technology.”

   Photography by William Atkins