Aurelia Butler

A University of Minnesota Twin Cities-led team has developed a first-of-its-kind, breakthrough method that makes it easier to create high-quality metal oxide thin films out of “stubborn” metals that have historically been difficult to synthesize in an atomically precise manner. This research paves the way for scientists to develop better materials for various next-generation applications including quantum computing, microelectronics, sensors, and energy catalysis.
The researchers’ paper is published in Nature Nanotechnology, a peer-reviewed, scientific journal run by Nature Publishing Group.
“This is truly remarkable discovery, as it unveils an unparalleled and simple way for navigating material synthesis at the atomic scale by harnessing the power of epitaxial strain,” said Bharat Jalan, senior author on the paper and a professor and Shell Chair in the University of Minnesota Department of Chemical Engineering and Materials Science. “This breakthrough represents a significant advancement with far-reaching implications in a broad range of fields. Not only does it provide a means to achieve atomically-precise synthesis of quantum materials, but it also holds immense potential for controlling oxidation-reduction pathways in various applications, including catalysis and chemical reactions occurring in batteries or fuel cells.”
“Stubborn” metals oxides, such as those based on ruthenium or iridium, play a crucial role in numerous applications in quantum information sciences and electronics. However, converting them into thin films has been a challenge for researchers due to the inherent difficulties in oxidizing metals using high-vacuum processes.
The fabrication of these materials has perplexed materials scientists for decades. While some researchers have successfully achieved oxidation, the methods used thus far have been costly, unsafe, or have resulted in poor material quality.
The University of Minnesota researchers’ solution? Give it a stretch.

While attempting to synthesize metal oxides using conventional molecular beam epitaxy, a low-energy technique that generates single layers of material in an ultra-high vacuum chamber, the researchers stumbled upon a groundbreaking revelation. They found that incorporating a concept called “epitaxial strain” — effectively stretching the metals at the atomic level — significantly simplifies the oxidation process of these stubborn metals.
“This enables the creation of technologically important metal oxides out of stubborn metals in ultra-high vacuum atmospheres, which has been a longstanding problem,” said Sreejith Nair, first author of the paper and a University of Minnesota chemical engineering Ph.D. student. “The current synthesis approaches have limits, and we need to find new ways to push those limits further so that we can make better quality materials. Our new method of stretching the material at the atomic scale is one way to improve the performance of the current technology.”
Although the University of Minnesota team used iridium and ruthenium as examples in this paper, their method has the potential to generate atomically-precise oxides of any hard-to-oxidize metal. With this groundbreaking discovery, the researchers aim to empower scientists worldwide to synthesize these novel materials.
The researchers worked closely with collaborators at Auburn University, the University of Delaware, Brookhaven National Laboratory, Argonne National Laboratory, and fellow University of Minnesota Department of Chemical Engineering and Materials Science Professor Andre Mkhoyan’s lab to verify their method.
“When we looked at these metal oxide films very closely using very powerful electron microscopes, we captured the arrangements of the atoms and determined their types,” Mkhoyan explained. “Sure enough, they were nicely and periodically arranged as they should be in these crystalline films.”
This research was funded primarily by the United States Department of Energy (DOE), the Air Force Office of Scientific Research (AFOSR), and the University of Minnesota’s Materials Research Science and Engineering Center (MRSEC).
In addition to Jalan, Nair, and Mkhoyan, the research team included University of Minnesota Twin Cities researchers Zhifei Yang, Dooyong Lee, and Silu Guo; Brookhaven National Laboratory researcher Jerzy Sadowski; Auburn University researchers Spencer Johnson, Ryan Comes, and Wencan Jin; University of Delaware researchers Abdul Saboor and Anderson Janotti; and Argonne National Laboratory researchers Yan Li and Hua Zhou. Parts of the work were carried out at the University of Minnesota’s Characterization Facility. More

Artificial intelligence has exploded across our news feeds, with ChatGPT and related AI technologies becoming the focus of broad public scrutiny. Beyond popular chatbots, biologists are finding ways to leverage AI to probe the core functions of our genes.
Previously, University of California San Diego researchers who investigate DNA sequences that switch genes on used artificial intelligence to identify an enigmatic puzzle piece tied to gene activation, a fundamental process involved in growth, development and disease. Using machine learning, a type of artificial intelligence, School of Biological Sciences Professor James T. Kadonaga and his colleagues discovered the downstream core promoter region (DPR), a “gateway” DNA activation code that’s involved in the operation of up to a third of our genes.
Building from this discovery, Kadonaga and researchers Long Vo ngoc and Torrey E. Rhyne have now used machine learning to identify “synthetic extreme” DNA sequences with specifically designed functions in gene activation. Publishing in the journal Genes & Development, the researchers tested millions of different DNA sequences through machine learning (AI) by comparing the DPR gene activation element in humans versus fruit flies (Drosophila). By using AI, they were able to find rare, custom-tailored DPR sequences that are active in humans but not fruit flies and vice versa. More generally, this approach could now be used to identify synthetic DNA sequences with activities that could be useful in biotechnology and medicine.
“In the future, this strategy could be used to identify synthetic extreme DNA sequences with practical and useful applications. Instead of comparing humans (condition X) versus fruit flies (condition Y) we could test the ability of drug A (condition X) but not drug B (condition Y) to activate a gene,” said Kadonaga, a distinguished professor in the Department of Molecular Biology. “This method could also be used to find custom-tailored DNA sequences that activate a gene in tissue 1 (condition X) but not in tissue 2 (condition Y). There are countless practical applications of this AI-based approach. The synthetic extreme DNA sequences might be very rare, perhaps one-in-a-million — if they exist they could be found by using AI.”
Machine learning is a branch of AI in which computer systems continually improve and learn based on data and experience. In the new research, Kadonaga, Vo ngoc (a former UC San Diego postdoctoral researcher now at Velia Therapeutics) and Rhyne (a staff research associate) used a method known as support vector regression to “train” machine learning models with 200,000 established DNA sequences based on data from real-world laboratory experiments. These were the targets presented as examples for the machine learning system. They then “fed” 50 million test DNA sequences into the machine learning systems for humans and fruit flies and asked them to compare the sequences and identify unique sequences within the two enormous data sets.
While the machine learning systems showed that human and fruit fly sequences largely overlapped, the researchers focused on the core question of whether the AI models could identify rare instances where gene activation is highly active in humans but not in fruit flies. The answer was a resounding “yes.” The machine learning models succeeded in identifying human-specific (and fruit fly-specific) DNA sequences. Importantly, the AI-predicted functions of the extreme sequences were verified in Kadonaga’s laboratory by using conventional (wet lab) testing methods.
“Before embarking on this work, we didn’t know if the AI models were ‘intelligent’ enough to predict the activities of 50 million sequences, particularly outlier ‘extreme’ sequences with unusual activities. So, it’s very impressive and quite remarkable that the AI models could predict the activities of the rare one-in-a-million extreme sequences,” said Kadonaga, who added that it would be essentially impossible to conduct the comparable 100 million wet lab experiments that the machine learning technology analyzed since each wet lab experiment would take nearly three weeks to complete.
The rare sequences identified by the machine learning system serve as a successful demonstration and set the stage for other uses of machine learning and other AI technologies in biology.
“In everyday life, people are finding new applications for AI tools such as ChatGPT. Here, we’ve demonstrated the use of AI for the design of customized DNA elements in gene activation. This method should have practical applications in biotechnology and biomedical research,” said Kadonaga. “More broadly, biologists are probably at the very beginning of tapping into the power of AI technology.” More

The number of qubits in superconducting quantum computers has risen rapidly during the last years, but further growth is limited by the need for ultra-cold operating temperatures. Connecting several smaller processors could create larger, more computationally powerful quantum computers, however doing so poses new challenges. A team of researchers led by Rishabh Sahu, Liu Qiu, and Johannes Fink from the Institute of Science and Technology Austria (ISTA) have now, for the first time, demonstrated quantum entanglement between optical and microwave photons that could lay the foundation for such a future quantum network.
Quantum computers promise to solve challenging tasks in material science and cryptography that will remain out of reach even for the most powerful conventional supercomputers in the future. Yet, this will likely require millions of high-quality qubits due to the required error correction.
Progress in superconducting processors advances quickly with a current qubit count in the few hundreds. The advantages of this technology are the fast computing speed and its compatibility with microchip fabrication, but the need for ultra-cold temperatures ultimately confines the processor in size and prevents any physical access once it is cooled down.
A modular quantum computer with multiple separately cooled processor nodes could solve this. However, single microwave photons — the particles of light that are the native information carriers between superconducting qubits within the processors — are not suitable to be sent through a room temperature environment between the processors. The world at room temperature is bustling with heat, which easily disturbs the microwave photons and their fragile quantum properties like entanglement.
Researchers from the Fink group at the Institute of Science and Technology Austria (ISTA), together with collaborators from TU Wien and the Technical University of Munich, demonstrated an important technological step to overcome these challenges. They entangled low-energy microwave with high-energy optical photons for the very first time. Such an entangled quantum state of two photons is the foundation to wire up superconducting quantum computers via room temperature links. This has implications not only for scaling up existing quantum hardware but it is also needed to realize interconnects to other quantum computing platforms as well as for novel quantum-enhanced remote sensing applications. Their results have been published in the journal Science.
Cooling Away the Noise
Rishabh Sahu, a postdoc in the Fink group and one of the first authors of the new study, explains, “One major problem for any qubit is noise. Noise can be thought of as any disturbance to the qubit. One major source of noise is the heat of the material the qubit is based on.”

Heat causes atoms in a material to jostle around rapidly. This is disruptive to quantum properties like entanglement, and as a result, it would make qubits unsuitable for computation. Therefore, to remain functional, a quantum computer must have its qubits isolated from the environment, cooled to extremely low temperatures, and kept within a vacuum to preserve their quantum properties.
For superconducting qubits, this happens in a special cylindrical device that hangs from the ceiling, called a “dilution refrigerator” in which the “quantum” part of the computation takes place. The qubits at its very bottom are cooled down to only a few thousandths of a degree above absolute zero temperature — at about -273 degrees Celsius. Sahu excitedly adds, “This makes these fridges in our labs the coldest locations in the whole universe, even colder than space itself.”
The refrigerator has to continuously cool the qubits but the more qubits and associated control wiring are added, the more heat is generated and the harder it is to keep the quantum computer cool. “The scientific community predicts that at around 1,000 superconducting qubits in a single quantum computer, we reach the limits of cooling,” Sahu cautions. “Just scaling up is not a sustainable solution to construct more powerful quantum computers.”
Fink adds, “Larger machines are in development but each assembly and cooldown then becomes comparable to a rocket launch, where you only find out about problems once the processor is cold and without the ability to intervene and correct such problems.”
Quantum Waves
“If a dilution fridge cannot sufficiently cool more than a thousand superconducting qubits at once, we need to link several smaller quantum computers to work together,” Liu Qiu, postdoc in the Fink group and another first author of the new study, explains. “We would need a quantum network.”

Linking together two superconducting quantum computers, each with its own dilution refrigerator, is not as straightforward as connecting them with an electrical cable. The connection needs special consideration to preserve the quantum nature of the qubits.
Superconducting qubits work with tiny electrical currents that move back and forth in a circuit at frequencies about ten billion times per second. They interact using microwave photons — particles of light. Their frequencies are similar to the ones used by cellphones.
The problem is that even a small amount of heat would easily disturb single microwave photons and their quantum properties needed to connect the qubits in two separate quantum computers. When passing through a cable outside the refrigerator, the heat of the environment would render them useless.
“Instead of the noise-prone microwave photons that we need to do the computations within the quantum computer, we want to use optical photons with much higher frequencies similar to visible light to network quantum computers together,” Qiu explains. These optical photons are the same kind sent through optical fibers that deliver high-speed internet to our homes. This technology is well understood and much less susceptible to noise from heat. Qiu adds, “The challenge was how to have the microwave photons interact with the optical photons and how to entangle them.”
Splitting Light
In their new study, the researchers used a special electro-optic device: an optical resonator made from a nonlinear crystal, which changes its optical properties in the presence of an electric field. A superconducting cavity houses this crystal and enhances this interaction.
Sahu and Qiu used a laser to send billions of optical photons into the electro-optic crystal for a fraction of a microsecond. In this way, one optical photon splits into a pair of new entangled photons: an optical one with only slightly less energy than the original one and a microwave photon with much lower energy.
“The challenge of this experiment was that the optical photons have about 20,000 times more energy than the microwave photons,” Sahu explains, “and they bring a lot of energy and therefore heat into the device that can then destroy the quantum properties of the microwave photons. We have worked for months tweaking the experiment and getting the right measurements.” To solve this problem, the researchers built a bulkier superconducting device compared to previous attempts. This not only avoids a breakdown of superconductivity, but it also helps to cool the device more effectively and to keep it cold during the short timescales of the optical laser pulses.
“The breakthrough is that the two photons leaving the device — the optical and the microwave photon — are entangled,” Qiu explains. “This has been verified by measuring correlations between the quantum fluctuations of the electromagnetic fields of the two photons that are stronger than can be explained by classical physics.”
“We are now the first to entangle photons of such vastly different energy scales.” Fink says, “This is a key step to creating a quantum network and also useful for other quantum technologies, such as quantum-enhanced sensing.” More

A new paper on wireless connectivity from the lab of Dinesh Bharadia, an affiliate of the UC San Diego Qualcomm Institute (QI) and faculty member with the Jacobs School of Engineering’s Department of Electrical and Computer Engineering, introduces a new technique for increasing access to the 5G-and-beyond millimeter wave (mmWave) network.
“Energy grids and mmWave/sub-THz networks share a remarkable similarity; both face fundamental challenges in efficient distribution,” said Bharadia. “Just as energy grids generate substantial amounts of energy but encounter significant hurdles in efficiently delivering it to homes, the utilization of mmWave/sub-THz networks for seamless data connectivity presents a similar predicament. Despite abundantly available bandwidth in these spectra, the efficient distribution of data with these spectra to user devices remains a formidable challenge.”
The paper, “mmFlexible: Flexible Directional Frequency Multiplexing for Multi-user mmWave Networks,” was presented by Ph.D. student and lead author Ish Kumar Jain at the IEEE International Conference on Computer Communications in New York on Wednesday, May 17.
Sharing Access
With the introduction of more automation and greater speeds and processing power behind wireless networks, the infrastructure that connects people to these resources has fallen behind.
Jain was drawn to the challenge of creating a device that could bridge this gap and give people greater access to the 5G mmWave network.

5G mmWave systems use radio frequencies to connect everything from “smart” cars to handheld devices and virtual reality sets to wireless networks. The advancement from 4G to 5G allows for higher speeds and bandwidth overall.
Part of the problem, Jain says, is that the jump from 4G to 5G opened up far more resources and processing power than existing infrastructure could handle. mmWave systems depend on a “pencil beam” distribution model in which a base station sends out a single beam of coverage, like shining a light in the dark. Everyone within that beam has access to all resources that the 5G mmWave network has to offer, regardless of whether their devices can process them.
This can lead to a waste of bandwidth that might otherwise have been leveraged by users in other regions. Even shifting this beam, like a lighthouse rotating slowly at timed intervals, creates lag for those who fall beyond its range.
To address the combined issues of wasted bandwidth and lag, Jain, Rohith Reddy Vennam and Raghav Subbaraman, also Ph.D. students in Bharadia’s Wireless Communication, Sensing and Networking Group (WCSNG), set out to determine whether they could create an antenna array that served users in multiple directions without sacrificing distance and power.
The team designed a prototype device that works in concert with a novel array of antennas to divide a single frequency band into multiple usable beams. Called a delay phased array, this antenna arrangement leverages 5G mmWave’s sheer amount of bandwidth to connect multiple regions to the network and can be tailored to deliver greater connection to those who need it.
This new, programmable array can also be built using existing technologies and scaled up with a very high number of antennas to support all future devices.
Through experiments run in QI’s Atkinson Hall on the UC San Diego campus, the team found mmFlexible decreased lag by 60-150%.
“It’s very exciting to see new generations of applications coming up,” said Jain. “But I feel, in the future, the number of [wireless] devices will grow and so will their demand for wireless spectrum. These are the key things that motivate me to further explore these innovative techniques.” More

To study ocean currents, scientists release GPS-tagged buoys in the ocean and record their velocities to reconstruct the currents that transport them. These buoy data are also used to identify “divergences,” which are areas where water rises up from below the surface or sinks beneath it.
By accurately predicting currents and pinpointing divergences, scientists can more precisely forecast the weather, approximate how oil will spread after a spill, or measure energy transfer in the ocean. A new model that incorporates machine learning makes more accurate predictions than conventional models do, a new study reports.
A multidisciplinary research team including computer scientists at MIT and oceanographers has found that a standard statistical model typically used on buoy data can struggle to accurately reconstruct currents or identify divergences because it makes unrealistic assumptions about the behavior of water.
The researchers developed a new model that incorporates knowledge from fluid dynamics to better reflect the physics at work in ocean currents. They show that their method, which only requires a small amount of additional computational expense, is more accurate at predicting currents and identifying divergences than the traditional model.
This new model could help oceanographers make more accurate estimates from buoy data, which would enable them to more effectively monitor the transportation of biomass (such as Sargassum seaweed), carbon, plastics, oil, and nutrients in the ocean. This information is also important for understanding and tracking climate change.
“Our method captures the physical assumptions more appropriately and more accurately. In this case, we know a lot of the physics already. We are giving the model a little bit of that information so it can focus on learning the things that are important to us, like what are the currents away from the buoys, or what is this divergence and where is it happening?” says senior author Tamara Broderick, an associate professor in MIT’s Department of Electrical Engineering and Computer Science (EECS) and a member of the Laboratory for Information and Decision Systems and the Institute for Data, Systems, and Society.

Broderick’s co-authors include lead author Renato Berlinghieri, an electrical engineering and computer science graduate student; Brian L. Trippe, a postdoc at Columbia University; David R. Burt and Ryan Giordano, MIT postdocs; Kaushik Srinivasan, an assistant researcher in atmospheric and ocean sciences at the University of California at Los Angeles; Tamay Özgökmen, professor in the Department of Ocean Sciences at the University of Miami; and Junfei Xia, a graduate student at the University of Miami. The research will be presented at the International Conference on Machine Learning.
Diving into the data
Oceanographers use data on buoy velocity to predict ocean currents and identify “divergences” where water rises to the surface or sinks deeper.
To estimate currents and find divergences, oceanographers have used a machine-learning technique known as a Gaussian process, which can make predictions even when data are sparse. To work well in this case, the Gaussian process must make assumptions about the data to generate a prediction.
A standard way of applying a Gaussian process to oceans data assumes the latitude and longitude components of the current are unrelated. But this assumption isn’t physically accurate. For instance, this existing model implies that a current’s divergence and its vorticity (a whirling motion of fluid) operate on the same magnitude and length scales. Ocean scientists know this is not true, Broderick says. The previous model also assumes the frame of reference matters, which means fluid would behave differently in the latitude versus the longitude direction.

“We were thinking we could address these problems with a model that incorporates the physics,” she says.
They built a new model that uses what is known as a Helmholtz decomposition to accurately represent the principles of fluid dynamics. This method models an ocean current by breaking it down into a vorticity component (which captures the whirling motion) and a divergence component (which captures water rising or sinking).
In this way, they give the model some basic physics knowledge that it uses to make more accurate predictions.
This new model utilizes the same data as the old model. And while their method can be more computationally intensive, the researchers show that the additional cost is relatively small.
Buoyant performance
They evaluated the new model using synthetic and real ocean buoy data. Because the synthetic data were fabricated by the researchers, they could compare the model’s predictions to ground-truth currents and divergences. But simulation involves assumptions that may not reflect real life, so the researchers also tested their model using data captured by real buoys released in the Gulf of Mexico.
In each case, their method demonstrated superior performance for both tasks, predicting currents and identifying divergences, when compared to the standard Gaussian process and another machine-learning approach that used a neural network. For example, in one simulation that included a vortex adjacent to an ocean current, the new method correctly predicted no divergence while the previous Gaussian process method and the neural network method both predicted a divergence with very high confidence.
The technique is also good at identifying vortices from a small set of buoys, Broderick adds.
Now that they have demonstrated the effectiveness of using a Helmholtz decomposition, the researchers want to incorporate a time element into their model, since currents can vary over time as well as space. In addition, they want to better capture how noise impacts the data, such as winds that sometimes affect buoy velocity. Separating that noise from the data could make their approach more accurate.
“Our hope is to take this noisily observed field of velocities from the buoys, and then say what is the actual divergence and actual vorticity, and predict away from those buoys, and we think that our new technique will be helpful for this,” she says.
This research is supported, in part, by the Office of Naval Research, a National Science Foundation (NSF) CAREER Award, and the Rosenstiel School of Marine, Atmospheric, and Earth Science at the University of Miami. More

The theory of relativity works well when you want to explain cosmic-scale phenomena — such as the gravitational waves created when black holes collide. Quantum theory works well when describing particle-scale phenomena — such as the behavior of individual electrons in an atom. But combining the two in a completely satisfactory way has yet to be achieved. The search for a “quantum theory of gravity” is considered one of the significant unsolved tasks of science.
This is partly because the mathematics in this field is highly complicated. At the same time, it is tough to perform suitable experiments: One would have to create situations in which phenomena of both the relativity theory play an important role, for example, a spacetime curved by heavy masses, and at the same time, quantum effects become visible, for example the dual particle and wave nature of light.
At the TU Wien in Vienna, Austria, a new approach has now been developed for this purpose: A so-called “quantum simulator” is used to get to the bottom of such questions: Instead of directly investigating the system of interest (namely quantum particles in curved spacetime), one creates a “model system” from which one can then learn something about the system of actual interest by analogy. The researchers have now shown that this quantum simulator works excellently. The findings of this international collaboration involving physicists from the University of Crete, Nanyang Technological University, and FU Berlin are now published in the scientific journal Proceedings of the National Academy of Sciences of the USA (PNAS).
Learning from one system about another
The basic idea behind the quantum simulator is simple: Many physical systems are similar. Even if they are entirely different kinds of particles or physical systems on different scales that, at first glance, have little to do with each other, these systems may obey the same laws and equations at a deeper level. This means one can learn something about a particular system by studying another.
“We take a quantum system that we know we can control and adjust very well in experiments,” says Prof. Jörg Schmiedmayer of the Atomic Institute at TU Wien. “In our case, these are ultracold atomic clouds held and manipulated by an atom chip with electromagnetic fields.” Suppose you properly adjust these atomic clouds so that their properties can be translated into another quantum system. In that case, you can learn something about the other system from the measurement of the atomic cloud model system — much like you can learn something about the oscillation of a pendulum from the oscillation of a mass attached to a metal spring: They are two different physical systems, but one can be translated into the other.

The gravitational lensing effect
“We have now been able to show that we can produce effects in this way that can be used to resemble the curvature of spacetime,” says Mohammadamin Tajik of the Vienna Center for Quantum Science and Technology (VCQ) — TU Wien, first author of the current paper. In the vacuum, light propagates along a so-called “light cone.” The speed of light is constant; at equal times, the light travels the same distance in each direction. However, if the light is influenced by heavy masses, such as the sun’s gravitation, these light cones are bent. The light’s paths are no longer perfectly straight in curved spacetimes. This is called “gravitational lens effect.”
The same can now be shown in atomic clouds. Instead of the speed of light, one examines the speed of sound. “Now we have a system in which there is an effect that corresponds to spacetime curvature or gravitational lensing, but at the same time, it is a quantum system that you can describe with quantum field theories,” says Mohammadamin Tajik. “With this, we have a completely new tool to study the connection between relativity and quantum theory.”
A model system for quantum gravity
The experiments show that the shape of light cones, lensing effects, reflections, and other phenomena can be demonstrated in these atomic clouds precisely as expected in relativistic cosmic systems. This is not only interesting for generating new data for basic theoretical research — solid-state physics and the search for new materials also encounter questions that have a similar structure and can therefore be answered by such experiments.
“We now want to control these atomic clouds better to determine even more far-reaching data. For example, interactions between the particles can still be changed in a very targeted way,” explains Jörg Schmiedmayer. In this way, the quantum simulator can recreate physical situations that are so complicated that they cannot be calculated even with supercomputers.
The quantum simulator thus becomes a new, additional source of information for quantum research — in addition to theoretical calculations, computer simulations, and direct experiments. When studying the atomic clouds, the research team hopes to come across new phenomena that may have been entirely unknown up to now, which also take place on a cosmic, relativistic scale — but without a look at tiny particles, they might never have been discovered. More

Artificial intelligence could be a useful tool in mental health treatment, according to the results of a new pilot study led by University of Illinois Chicago researchers.
The study, which was the first to test an AI voice-based virtual coach for behavioral therapy, found changes in patients’ brain activity along with improved depression and anxiety symptoms after using Lumen, an AI voice assistant that delivered a form of psychotherapy.
The UIC team says the results, which are published in the journal Translational Psychiatry, offer encouraging evidence that virtual therapy can play a role in filling the gaps in mental health care, where waitlists and disparities in access are often hurdles that patients, particularly from vulnerable communities, must overcome to receive treatment.
“We’ve had an incredible explosion of need, especially in the wake of COVID, with soaring rates of anxiety and depression and not enough practitioners,” said Dr. Olusola A. Ajilore, UIC professor of psychiatry and co-first author of the paper. “This kind of technology may serve as a bridge. It’s not meant to be a replacement for traditional therapy, but it may be an important stop-gap before somebody can seek treatment.”
Lumen, which operates as a skill in the Amazon Alexa application, was developed by Ajilore and study senior author Dr.Jun Ma, the Beth and George Vitoux Professor of Medicine at UIC, along with collaborators at Washington University in St. Louis and Pennsylvania State University, with the support of a $2 million grant from the National Institute of Mental Health.
The UIC researchers recruited over 60 patients for the clinical study exploring the application’s effect on mild-to-moderate depression and anxiety symptoms, and activity in brain areas previously shown to be associated with the benefits of problem-solving therapy.

Two-thirds of the patients used Lumen on a study-provided iPad for eight problem-solving therapy sessions, with the rest serving as a “waitlist” control receiving no intervention.
After the intervention, study participants using the Lumen app showed decreased scores for depression, anxiety and psychological distress compared with the control group. The Lumen group also showed improvements in problem-solving skills that correlated with increased activity in the dorsolateral prefrontal cortex, a brain area associated with cognitive control. Promising results for women and underrepresented populations also were found.
“It’s about changing the way people think about problems and how to address them, and not being emotionally overwhelmed,” Ma said. “It’s a pragmatic and patient-driven behavior therapy that’s well established, which makes it a good fit for delivery using voice-based technology.”
A larger trial comparing the use of Lumen with both a control group on a waitlist, and patients receiving human-coached problem-solving therapy is currently being conducted by the researcher. They stress that the virtual coach doesn’t need to perform better than a human therapist to fill a desperate need in the mental health system.
“The way we should think about digital mental health service is not for these apps to replace humans, but rather to recognize what a gap we have between supply and demand, and then find novel, effective and safe ways to deliver treatments to individuals who otherwise do not have access, to fill that gap,” Ma said.
Co-first author of the study is Thomas Kannampallil at Washington University in St. Louis.
Other co-investigators include Aifeng Zhang, Nan Lv, Nancy E. Wittels, Corina R. Ronneberg, Vikas Kumar, Susanth Dosala, Amruta Barve, Kevin C. Tan, Kevin K. Cao, Charmi R. Patel and Emily A. Kringle, all of UIC; Joshua Smyth and Jillian A. Johnson at Pennsylvania State University; and Lan Xiao at Stanford University. More

The latest version of ChatGPT passed a radiology board-style exam, highlighting the potential of large language models but also revealing limitations that hinder reliability, according to two new research studies published in Radiology, a journal of the Radiological Society of North America (RSNA).
ChatGPT is an artificial intelligence (AI) chatbot that uses a deep learning model to recognize patterns and relationships between words in its vast training data to generate human-like responses based on a prompt. But since there is no source of truth in its training data, the tool can generate responses that are factually incorrect.
“The use of large language models like ChatGPT is exploding and only going to increase,” said lead author Rajesh Bhayana, M.D., FRCPC, an abdominal radiologist and technology lead at University Medical Imaging Toronto, Toronto General Hospital in Toronto, Canada. “Our research provides insight into ChatGPT’s performance in a radiology context, highlighting the incredible potential of large language models, along with the current limitations that make it unreliable.”
ChatGPT was recently named the fastest growing consumer application in history, and similar chatbots are being incorporated into popular search engines like Google and Bing that physicians and patients use to search for medical information, Dr. Bhayana noted.
To assess its performance on radiology board exam questions and explore strengths and limitations, Dr. Bhayana and colleagues first tested ChatGPT based on GPT-3.5, currently the most commonly used version. The researchers used 150 multiple-choice questions designed to match the style, content and difficulty of the Canadian Royal College and American Board of Radiology exams.
The questions did not include images and were grouped by question type to gain insight into performance: lower-order (knowledge recall, basic understanding) and higher-order (apply, analyze, synthesize) thinking. The higher-order thinking questions were further subclassified by type (description of imaging findings, clinical management, calculation and classification, disease associations).

The performance of ChatGPT was evaluated overall and by question type and topic. Confidence of language in responses was also assessed.
The researchers found that ChatGPT based on GPT-3.5 answered 69% of questions correctly (104 of 150), near the passing grade of 70% used by the Royal College in Canada. The model performed relatively well on questions requiring lower-order thinking (84%, 51 of 61), but struggled with questions involving higher-order thinking (60%, 53 of 89). More specifically, it struggled with higher-order questions involving description of imaging findings (61%, 28 of 46), calculation and classification (25%, 2 of 8), and application of concepts (30%, 3 of 10). Its poor performance on higher-order thinking questions was not surprising given its lack of radiology-specific pretraining.
GPT-4 was released in March 2023 in limited form to paid users, specifically claiming to have improved advanced reasoning capabilities over GPT-3.5.
In a follow-up study, GPT-4 answered 81% (121 of 150) of the same questions correctly, outperforming GPT-3.5 and exceeding the passing threshold of 70%. GPT-4 performed much better than GPT-3.5 on higher-order thinking questions (81%), more specifically those involving description of imaging findings (85%) and application of concepts (90%).
The findings suggest that GPT-4’s claimed improved advanced reasoning capabilities translate to enhanced performance in a radiology context. They also suggest improved contextual understanding of radiology-specific terminology, including imaging descriptions, which is critical to enable future downstream applications.

“Our study demonstrates an impressive improvement in performance of ChatGPT in radiology over a short time period, highlighting the growing potential of large language models in this context,” Dr. Bhayana said.
GPT-4 showed no improvement on lower-order thinking questions (80% vs 84%) and answered 12 questions incorrectly that GPT-3.5 answered correctly, raising questions related to its reliability for information gathering.
“We were initially surprised by ChatGPT’s accurate and confident answers to some challenging radiology questions, but then equally surprised by some very illogical and inaccurate assertions,” Dr. Bhayana said. “Of course, given how these models work, the inaccurate responses should not be particularly surprising.”
ChatGPT’s dangerous tendency to produce inaccurate responses, termed hallucinations, is less frequent in GPT-4 but still limits usability in medical education and practice at present.
Both studies showed that ChatGPT used confident language consistently, even when incorrect. This is particularly dangerous if solely relied on for information, Dr. Bhayana notes, especially for novices who may not recognize confident incorrect responses as inaccurate.
“To me, this is its biggest limitation. At present, ChatGPT is best used to spark ideas, help start the medical writing process and in data summarization. If used for quick information recall, it always needs to be fact-checked,” Dr. Bhayana said. More