large language models

National Science Foundation

College of Engineering and Computing

Can AI help make medical records less biased? New study suggests yes—with caveats

Heather Carroll — Fri, 10 Jul 2026 16:50:49 +0000

Can AI help make medical records less biased? New study suggests yes—with caveats Heather Carroll Fri, 07/10/2026 - 12:50

Large language models can identify judgmental language in clinical notes but the settings play a major role in accuracy.

Teenu Xavier PhD, RN. Photo provided

“Addict,” “non-compliant,” “failed treatment,” and “obese person” are examples of stigmatizing language that can appear in medical records. At 911��’s College of Public Health, researchers are exploring whether artificial intelligence (AI) can help identify this kind of language in clinical notes before it impacts patient care.

Nurse scientist Teenu Xavier and colleagues found that large language models (LLMs) show promise in identifying stigmatizing language in clinical documentation, but their performance is highly dependent on their settings. Model size, temperature settings, prompting strategies, and even note type can substantially influence results.

One finding was consistent across every model tested: Providing examples of stigmatizing language improved accuracy.

“Simply selecting an LLM is not enough when used for clinical documentation,” said Xavier, an assistant professor in the School of Nursing. “Careful attention must be paid to settings and prompting before these tools can be reliably used in health care environments.”

Why does this matter?

The use of stigmatizing language in clinical documentation can reinforce bias and affect a patient’s future care. AI tools may be able to help identify this kind of language, promoting more equitable care and improving patient trust and experience.

“Pre-trained models, when optimized for identifying stigmatizing language, could help enable more timely interventions and modifications to the documentation process,” Xavier said. “Our research highlights the need for continued collaboration between health care professionals and AI developers to create tools that improve communication, reduce bias, and improve the overall patient experience.”

What are the detailed study findings?

The largest LLM (trained on large amounts of data) was the best at predicting “stigmatizing” language (94%), but the worst at predicting “not stigmatizing” correctly (47%).
The smallest LLM was the best at predicting “not stigmatizing” correctly (99.7%), but worst at predicting “stigmatizing” correctly (2%).
When researchers gave the LLM an example of stigmatizing language, accuracy improved in all models.
Emergency provider notes were most accurately (69%) categorized as “stigmatizing” or “non-stigmatizing,” and plan of care notes had the lowest accuracy (56%). Misclassifications most commonly arose in long, clinically dense notes where neutral descriptions of complex illness or adverse events were mistaken by the models for judgmental language.
Larger models worked best at lower temperature (how predictable or random the LLMs’s output is when making a classification) and smaller models improved with higher temperature, which means the LLM took more risks in interpretation.

“Detecting stigmatizing language with large language models: mind the settings” was published in JAMIA Open in April 2026. Co-authors include Jane M. Carrington from the University of Florida and Joshua Lambert from the University of Cincinnati.

Thumbnail photo by issaronow from Adobe Stock.

Key Takeaways

A 911�� study found that large language models show promise in identifying stigmatizing language in clinical documentation, but their performance is highly dependent on their settings, such as model size, temperature settings, prompting strategies, and note type.
The ability to detect and correct stigmatizing language early can reduce bias in patient care and lead to improved patient trust and better health outcomes.
Continued collaboration between health care workers and AI developers is needed to create tools that enhance communication, reduce bias, and improve the overall patient experience.

Learn about Artificial Intelligence at 911��

Topics

Machine Learning in Health Care

Health Care Services

College of Public Health

911�� researcher is using CAREER award to build brain-inspired AI that is more adaptive

Nathan Kahl — Tue, 07 Jul 2026 18:49:00 +0000

911�� researcher is using CAREER award to build brain-inspired AI that is more adaptive Nathan Kahl Tue, 07/07/2026 - 14:49

In This Story

Maryam Parsa

Though artificial intelligence (AI) is making extraordinary advances, today's AI systems still have significant limitations. They often require enormous computing power and energy (not to mention well-documented water usage), struggle to adapt when conditions change, and can make decisions in ways that are difficult for humans to interpret, validate, or trust.

Maryam Parsa. Photo provided

Maryam Parsa, assistant professor in 911��'s Department of Electrical and Computer Engineering, believes those shortcomings point toward a compelling source of inspiration: the human brain.

Parsa received a five-year, $655,000 National Science Foundation Early Career Development (CAREER) Award for developing a new framework for brain-inspired AI. The project seeks to make AI systems more efficient, adaptable, resilient, and transparent by borrowing principles from biological nervous systems.

"AI was originally inspired by the brain," Parsa said. "Concepts such as neurons, synapses, and neural networks drew inspiration from how brains process information. But over time, AI has moved farther away from those biological principles.”

Parsa believes researchers now have an opportunity to reconnect AI and neuroscience. “Neuroscience advanced tremendously over the past few decades, with only a small fraction of those discoveries finding their way into AI systems.”

As computing power grew, researchers found AI systems could make major gains by scaling neural networks and training them on massive datasets. The approach, fueled by increasingly powerful computers, is leading to significant breakthroughs, particularly with the rise of large language models (LLMs).

"They are producing remarkable devices," she said. "The progress we see in LLMs and modern AI grew out of relatively simple neural-network ideas. Over time, researchers added more layers, more data, more computational power, and more sophisticated training methods, leading to powerful systems.”

The goal is not to replace today's successful AI systems with direct copies of the brain. Instead, her research focuses on areas where current AI falls short. "We know some of the key challenges AI still faces," she said. "It can be difficult to understand how these systems reach their decisions, and they do not always transfer well when conditions, data, or environments change. That is where biological intelligence can offer useful design principles.”

When today’s AI systems encounter unfamiliar environments or changing conditions, they often need to be retrained or fine-tuned. Biological intelligence, however, is remarkably adaptable and energy efficient, offering clues to how future AI systems could learn and respond more flexibly.

Her CAREER project builds a unified framework around three characteristics of biological intelligence. The first pillar, dimensionality, examines how the brain creates meaningful representations of the world rather than simply processing raw information. The second, heterogeneity, explores the brain's remarkable diversity. Rather than relying on identical neurons performing tasks, biological brains contain many neuron types with distinct behaviors. "The brain is not copy, paste, copy, paste, copy, paste,” she said.

The third pillar, nonlinearity, draws inspiration from the retina, which performs extensive processing before visual information reaches the brain. "It's not like the raw pixels go directly to our central brain," Parsa said. "The retina digests and processes the information first."

Parsa said the broader idea is combining these principles into a more biologically inspired approach to AI. “If I know how to process my input, create a better representation of my environment, and leverage this massive heterogeneity, then I can move AI a little closer to the way biological intelligence works,” she said.

Beyond advancing AI research, the project includes an educational effort to help students see how ideas from the brain can shape the future of computing and artificial intelligence. Parsa plans to bring these concepts into graduate courses, hands-on hackathons, and STEM outreach activities for K-12 students, creating new pathways for students to explore the intersection of neuroscience, engineering, and AI.

Topics

Electrical and Computer Engineering

Electrical and Computer Engineering Faculty

National Science Foundation

Department of Cyber Security Engineering

C-TASC

Chaos theory finds its voice

Nathan Kahl — Tue, 19 May 2026 15:42:55 +0000

Chaos theory finds its voice Nathan Kahl Tue, 05/19/2026 - 11:42

Talk may be cheap, but it is still rich with data, and 911��'s Anomadarshi Barua is using artificial intelligence (AI) and a new approach to chaos theory to rebuild missing pieces of speech with incredible accuracy.

"When we talk, our voices have a frequency component from zero kilohertz (kHz) to around 8 kHz," said Barua, an assistant professor in the Department of Cyber Security Engineering. "Capturing all those frequencies normally requires devices to collect and process large amounts of data, which consumes power, storage, and computing resources."

He will present a paper he coauthored on this topic at the Association for Computational Linguistics (ACL) annual meeting later this summer. ACL is one of the world's leading conferences in natural language processing and speech technologies, and the paper was accepted in the top 15 percent of accepted submissions at a conference with a 19 percent acceptance rate overall.

Smaller devices and sensors operating with limited battery life or bandwidth have difficulty capturing an entire signal, and so record a smaller slice, attempting to reconstruct the missing frequencies later. Barua says this is known as bandwidth extension or bandwidth reconstruction.

Imagine listening to a song through a wall and trying to mentally fill in the muffled higher notes. Barua's research teaches AI systems to do something similar, but with far greater mathematical precision. The team's major breakthrough came from incorporating a concept not often associated with speech: chaos theory.

"Our speech is actually a chaotic signal," Barua said. "'Chaos' does not mean completely random."

Instead, he describes speech as "deterministically random," meaning sounds and phonemes are strongly connected to one another in predictable ways. "You can actually determine what could be the next phoneme if you have the sufficient information from the previous phoneme," he said.

To capture those hidden relationships, the researchers developed what Barua calls a "chaotic discriminator" inside a machine-learning framework known as a generative adversarial network, or GAN. In simple terms, one part of the AI system generates reconstructed speech while another checks whether the recreated speech preserves the natural chaotic patterns found in human voices.

The approach significantly improved reconstruction quality compared to previous methods. "We are getting more improved results compared to the previous baseline," Barua said, "because of the incorporation of the chaotic properties."

By integrating chaos-informed modeling, the team dramatically reduced the size of the AI system needed for reconstruction. "We actually reduce the size of the discriminator by 14 times," Barua said. Smaller models require less memory and computing power, making them more practical for real-world devices.

For Barua, whose earlier work focused primarily on cybersecurity before expanding into speech and natural language processing, the ACL conference's acceptance marks an important milestone. "This is my first paper in natural language processing," he said.

While the current paper focuses on speech, Barua sees much broader possibilities ahead. The same reconstruction techniques could eventually be applied to electrical signals, sonar, lidar, and other sensing technologies. "This concept is not only limited to speech," he said. "We are trying to open up a larger branch in different signal modality."

And that message comes through loud and clear.

In This Story

Anomadarshi Barua

Explore Cybersecurity at 911��

Powering Artificial Intelligence

Topics

Cybersecurity

natural language processing

College of Engineering and Computing

C-TASC

AI’s blind spots, one PhD student’s clear vision

Ryley McGinnis — Mon, 11 Aug 2025 18:36:42 +0000

AI’s blind spots, one PhD student’s clear vision Ryley McGinnis Mon, 08/11/2025 - 14:36

When artificial intelligence makes a mistake, the consequences can be annoying—or they can be life-altering. For one 911�� computer science PhD student, the stakes are clear: biased algorithms in health care aren’t just a technical flaw, they’re a human risk. And with internship experience teaching him crucial skills beyond the classroom, Fardin Ahsan Sakib is ready to make a difference in health care.

Sakib’s research in natural language processing (NLP) addresses concerns that many may have when working with emerging large language models (LLMs), but when it is applied to health decisions, there could be deeper negative risks. Tools like ChatGPT pull from vast amounts of information, but they have sometimes been known to “hallucinate,” or make up information.

“Sometimes these systems are taking a shortcut to get you an answer,” said Sakib. “But regardless of where it comes from, it can provide you incorrect information.” And in the health care space, these LLMs could aid doctors, but not when they could be driving improper care.

In electronic health records (EHRs), patients’ diagnoses, medical history, and demographic data are stored and organized. This information includes social determinants of health, like employment status, family situations, or housing conditions, hidden in records.

Fardin Ahsan Sakib is seeking to bring accurate and reliable large language models' capabilities to healthcare.

“These factors can affect up to 80% of health outcomes,” Sakib said. “So it's very important that we can extract them correctly from clinical notes and that the model is not introducing any bias.”

Sakib poses a scenario: A doctor is using an NLP tool to quickly assess a new patient’s health. The LLM is pulling from the patient’s records, and the physician asks the system, “Is this patient a smoker?” The system quickly sifts through the data, and says that yes, this patient is a smoker. The physician then proceeds to make care recommendations based on this information. Maybe they order a lung cancer screening, or they speak with the patient about smoking cessation resources.

But what if that answer was incorrect? The LLM decided to take a shortcut, because, based off the information it has, it discovered that most people with similar demographic data to the patient are indeed smokers, but this one isn’t.

“There are two things at play: bias and hallucination,” said Sakib. “First, algorithms can only use the information they have, and if that information is missing an entire racial or ethnic group, it can be biased. Second, these models can use this biased information to take shortcuts, leading to them delivering false information.”

Sakib isn't simply identifying these problems; he's developing solutions. His recent research on detecting and mitigating shortcut learning in health record processing was accepted for presentation at the 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025), one of the premier NLP conferences.

That desire to build reliable and trustworthy NLP tools has driven Sakib’s academic work and his professional pursuits.

At Brillient Corporation, where he interned last summer, Sakib worked on creating a retrieval augmented generation system that connected an LLM to the Food and Drug Administration’s (FDA) knowledge base to make accessing and retrieving information faster and grounded in facts. He and his colleagues submitted a patent for this effort.

This summer, he's deep into another high-impact internship at Amazon. “Amazon wants to automate support so that when you ask for help, a large language model can try to solve the problem before handing it off to a human,” he said.

Despite the different settings, he sees clear continuity between his internships and academic work. “The industry experience helps me in my research, and my research experience helps in industry. It goes both ways,” he said. “In academia, collaboration is usually focused within a lab or research group. In industry, I’ve worked with engineers, product managers, and domain experts all at once, spanning from health regulators to AWS cloud architects. That diversity of perspectives changes how you solve problems, and I’ve brought that mindset back into my research collaborations.”

It’s that dual perspective—academic precision paired with industry scale—that he plans to take with him into industry after graduation.

“911�� has prepared me for a lot. From day one, all of the professors have helped me grow as a researcher, a person, and as a team member,” said Sakib.

Topics