Computers Math

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

Engineers at the University of Waterloo have discovered a new way to program robots to help people with dementia locate medicine, glasses, phones and other objects they need but have lost.
And while the initial focus is on assisting a specific group of people, the technology could someday be used by anyone who has searched high and low for something they’ve misplaced.
“The long-term impact of this is really exciting,” said Dr. Ali Ayub, a post-doctoral fellow in electrical and computer engineering. “A user can be involved not just with a companion robot but a personalized companion robot that can give them more independence.”
Ayub and three colleagues were struck by the rapidly rising number of people coping with dementia, a condition that restricts brain function, causing confusion, memory loss and disability. Many of these individuals repeatedly forget the location of everyday objects, which diminishes their quality of life and places additional burdens on caregivers.
Engineers believed a companion robot with an episodic memory of its own could be a game-changer in such situations. And they succeeded in using artificial intelligence to create a new kind of artificial memory.
The research team began with a Fetch mobile manipulator robot, which has a camera for perceiving the world around it.
Next, using an object-detection algorithm, they programmed the robot to detect, track and keep a memory log of specific objects in its camera view through stored video. With the robot capable of distinguishing one object from another, it can record the time and date objects enter or leave its view.
Researchers then developed a graphical interface to enable users to choose objects they want to be tracked and, after typing the objects’ names, search for them on a smartphone app or computer. Once that happens, the robot can indicate when and where it last observed the specific object.
Tests have shown the system is highly accurate. And while some individuals with dementia might find the technology daunting, Ayub said caregivers could readily use it.
Moving forward, researchers will conduct user studies with people without disabilities, then people with dementia.
A paper on the project, Where is my phone? Towards developing an episodic memory model for companion robots to track users’ salient objects, was presented at the recent 2023 ACM/IEEE International Conference on Human-Robot Interaction. More

A Cornell University-led experiment in which two people play a modified version of Tetris revealed that players who get fewer turns perceived the other player as less likable, regardless of whether a person or an algorithm allocated the turns.
Most studies on algorithmic fairness focus on the algorithm or the decision itself, but researchers sought to explore the relationships among the people affected by the decisions.
“We are starting to see a lot of situations in which AI makes decisions on how resources should be distributed among people,” said Malte Jung, associate professor of information science, whose group conducted the study. “We want to understand how that influences the way people perceive one another and behave towards each other. We see more and more evidence that machines mess with the way we interact with each other.”
In an earlier study, a robot chose which person to give a block to and studied the reactions of each individual to the machine’s allocation decisions.
“We noticed that every time the robot seemed to prefer one person, the other one got upset,” said Jung. “We wanted to study this further, because we thought that, as machines making decisions becomes more a part of the world — whether it be a robot or an algorithm — how does that make a person feel?”
Using open-source software, Houston Claure — the study’s first author and postdoctoral researcher at Yale University — developed a two-player version of Tetris, in which players manipulate falling geometric blocks in order to stack them without leaving gaps before the blocks pile to the top of the screen. Claure’s version, Co-Tetris, allows two people (one at a time) to work together to complete each round.
An “allocator” — either human or AI, which was conveyed to the players — determines which player takes each turn. Jung and Claure devised their experiment so that players would have either 90% of the turns (the “more” condition), 10% (“less”) or 50% (“equal”).
The researchers found, predictably, that those who received fewer turns were acutely aware that their partner got significantly more. But they were surprised to find that feelings about it were largely the same regardless of whether a human or an AI was doing the allocating.
The effect of these decisions is what the researchers have termed “machine allocation behavior” — similar to the established phenomenon of “resource allocation behavior,” the observable behavior people exhibit based on allocation decisions. Jung said machine allocation behavior is “the concept that there is this unique behavior that results from a machine making a decision about how something gets allocated.”
The researchers also found that fairness didn’t automatically lead to better game play and performance. In fact, equal allocation of turns led, on average, to a worse score than unequal allocation.
“If a strong player receives most of the blocks,” Claure said, “the team is going to do better. And if one person gets 90%, eventually they’ll get better at it than if two average players split the blocks.” More

Students in Sweden are positive towards AI tools such as ChatGPT in education, but 62 percent believe that using chatbots during exams is cheating. However, where the boundary for cheating lies is highly unclear. This is shown in a survey from Chalmers University of Technology, which is the first large-scale study in Europe to investigate students’ attitudes towards artificial intelligence in higher education.
“I am afraid of AI and what it could mean for the future.”
“Don’t worry so much! Keep up with the development and adapt your teaching for the future.”
“ChatGPT and similar tools will revolutionise how we learn, and we will be able to come up with amazing things.”
These are three out of nearly two thousand optional comments from the survey which almost 6,000 students in Sweden recently participated in.
“The students express strong, diverse, and in many cases emotionally charged opinions,” says Hans Malmström, Professor at the Department of Communication and Learning in Science at Chalmers University of technology. He, together with his colleagues Christian Stöhr and Amy Wanyu Ou, conducted the study.

More than a third use ChatGPT regularly
A majority of the respondents believe that chatbots and AI language tools make them more efficient as students and argue that such tools improve their academic writing and overall language skills. Virtually all the responding students are familiar with ChatGPT, the majority use the tool, and 35 percent use the chatbot regularly.
Lack guidance — opposed a ban
Despite their positive attitude towards AI, many students feel anxious and lack clear guidance on how to use AI in the learning environments they are in. It is simply difficult to know where the boundary for cheating lies.
“Most students have no idea whether their educational institution has any rules or guidelines for using AI responsibly, and that is of course worrying. At the same time, an overwhelming majority is against a ban on AI in educational contexts,” says Hans Malmström.

No replacement for critical thinking
Many students perceive chatbots as a mentor or teacher that they can ask questions or get help from, for example, with explanations of concepts and summaries of ideas. The dominant attitude is that chatbots should be used as an aid, not replace students’ own critical thinking. Or as one student put it: “You should be able to do the same things as the AI, but it should help you do it. You should not use a calculator if you don’t know what the plus sign on it does.”
Aid in case of disabilities
Another important aspect that emerged in the survey was that AI serves as an effective aid for people with various disabilities. A student with ADD and dyslexia described how they had spent 20 minutes writing down their answer in the survey and then improved it by inputting the text into ChatGPT: “It’s like being color blind and suddenly being able to see all the beautiful colors.”
Giving students a voice
The researchers have now gathered a wealth of important information and compiled the results in an overview report.
“We hope and believe that the answers from this survey will give students a voice and the results will thus be an important contribution to our collective understanding of AI and learning,” says Christian Stöhr, Associate Professor at the Department of Communication and Learning in Science at Chalmers.
More about the study
“Chatbots and other AI for learning: A survey on use and views among university students in Sweden” was conducted in the following way: The researchers at Chalmers conducted the survey between 5 April and 5 May, 2023. Students at all universities in Sweden could participate. The survey was distributed through social media and targeted efforts from multiple universities and student organisations. In total, the survey was answered by 5,894 students.
Summary of results: 95 percent of students are familiar with ChatGPT, while awareness of other chatbots is very low. 56 percent are positive about using chatbots in their studies; 35 percent use ChatGTP regularly. 60 percent are opposed to a ban on chatbots, and 77 percent are against a ban on other AI tools (such as Grammarly) in education. More than half of the students do not know if their institution has guidelines for how AI can be used in education; one in four explicitly says that their institution lack such regulations. 62 percent believe that using chatbots during examinations is cheating. Students express some concern about AI development, and there is particular concern over the impact of chatbots on future education. More

In the future, artificial intelligence will play an important role in medicine. In diagnostics, successful tests have already been performed: for example, the computer can learn to categorise images with great accuracy according to whether they show pathological changes or not. However, it is more difficult to train an artificial intelligence to examine the time-varying conditions of patients and to calculate treatment suggestions — this is precisely what has now been achieved at TU Wien in cooperation with the Medical University of Vienna.
With the help of extensive data from intensive care units of various hospitals, an artificial intelligence was developed that provides suggestions for the treatment of people who require intensive care due to sepsis. Analyses show that artificial intelligence already surpasses the quality of human decisions. However, it is now important to also discuss the legal aspects of such methods.
Making optimal use of existing data
“In an intensive care unit, a lot of different data is collected around the clock. The patients are constantly monitored medically. We wanted to investigate whether these data could be used even better than before,” says Prof. Clemens Heitzinger from the Institute for Analysis and Scientific Computing at TU Wien (Vienna). He is also Co-Director of the cross-faculty “Center for Artificial Intelligence and Machine Learning” (CAIML) at TU Wien.
Medical staff make their decisions on the basis of well-founded rules. Most of the time, they know very well which parameters they have to take into account in order to provide the best care. However, the computer can easily take many more parameters than a human into account — and in some cases this can lead to even better decisions.
The computer as planning agent
“In our project, we used a form of machine learning called reinforcement learning,” says Clemens Heitzinger. “This is not just about simple categorisation — for example, separating a large number of images into those that show a tumour and those that do not — but about a temporally changing progression, about the development that a certain patient is likely to go through. Mathematically, this is something quite different. There has been little research in this regard in the medical field.”

The computer becomes an agent that makes its own decisions: if the patient is well, the computer is “rewarded.” If the condition deteriorates or death occurs, the computer is “punished.” The computer programme has the task of maximising its virtual “reward” by taking actions. In this way, extensive medical data can be used to automatically determine a strategy which achieves a particularly high probability of success.
Already better than a human
“Sepsis is one of the most common causes of death in intensive care medicine and poses an enormous challenge for doctors and hospitals, as early detection and treatment is crucial for patient survival,” says Prof. Oliver Kimberger from the Medical University of Vienna. “So far, there have been few medical breakthroughs in this field, which makes the search for new treatments and approaches all the more urgent. For this reason, it is particularly interesting to investigate the extent to which artificial intelligence can contribute to improve medical care here. Using machine learning models and other AI technologies are an opportunity to improve the diagnosis and treatment of sepsis, ultimately increasing the chances of patient survival.”
Analysis shows that AI capabilities are already outperforming humans: “Cure rates are now higher with an AI strategy than with purely human decisions. In one of our studies, the cure rate in terms of 90-day mortality was increased by about 3% to about 88%,” says Clemens Heitzinger.
Of course, this does not mean that one should leave medical decisions in an intensive care unit to the computer alone. But the artificial intelligence may run along as an additional device at the bedside — and the medical staff can consult it and compare their own assessment with the artificial intelligence’s suggestions. Such artificial intelligences can also be highly useful in education.
Discussion about legal issues is necessary
“However, this raises important questions, especially legal ones,” says Clemens Heitzinger. “One probably thinks of the question who will be held liable for any mistakes made by the artificial intelligence first. But there is also the converse problem: what if the artificial intelligence had made the right decision, but the human chose a different treatment option and the patient suffered harm as a result?” Does the doctor then face the accusation that it would have been better to trust the artificial intelligence because it comes with a huge wealth of experience? Or should it be the human’s right to ignore the computer’s advice at all times?
“The research project shows: artificial intelligence can already be used successfully in clinical practice with today’s technology — but a discussion about the social framework and clear legal rules are still urgently needed,” Clemens Heitzinger is convinced. More

Cornell University researchers have developed a robot, called ReMotion, that occupies physical space on a remote user’s behalf, automatically mirroring the user’s movements in real time and conveying key body language that is lost in standard virtual environments.
“Pointing gestures, the perception of another’s gaze, intuitively knowing where someone’s attention is — in remote settings, we lose these nonverbal, implicit cues that are very important for carrying out design activities,” said Mose Sakashita, a doctoral student of information science.
Sakashita is the lead author of “ReMotion: Supporting Remote Collaboration in Open Space with Automatic Robotic Embodiment,” which he presented at the Association for Computing Machinery CHI Conference on Human Factors in Computing Systems in Hamburg, Germany. “With ReMotion, we show that we can enable rapid, dynamic interactions through the help of a mobile, automated robot.”
The lean, nearly six-foot-tall device is outfitted with a monitor for a head, omnidirectional wheels for feet and game-engine software for brains. It automatically mirrors the remote user’s movements — thanks to another Cornell-made device, NeckFace, which the remote user wears to track head and body movements. The motion data is then sent remotely to the ReMotion robot in real-time.
Telepresence robots are not new, but remote users generally need to steer them manually, distracting from the task at hand, researchers said. Other options such as virtual reality and mixed reality collaboration can also require an active role from the user and headsets may limit peripheral awareness, researchers added.
In a small study, nearly all participants reported having a better connection with their remote teammates when using ReMotion compared to an existing telerobotic system. Participants also reported significantly higher shared attention among remote collaborators.
In its current form, ReMotion only works with two users in a one-on-one remote environment, and each user must occupy physical spaces of identical size and layout. In future work, ReMotion developers intend to explore asymmetrical scenarios, like a single remote team member collaborating virtually via ReMotion with multiple teammates in a larger room.
With further development, Sakashita says ReMotion could be deployed in virtual collaborative environments as well as in classrooms and other educational settings.
This research was funded in part by the National Science Foundation and the Nakajima Foundation. More

A team of researchers led by Rensselaer Polytechnic Institute’s Trevor David Rhone, assistant professor in the Department of Physics, Applied Physics, and Astronomy, has identified novel van der Waals (vdW) magnets using cutting-edge tools in artificial intelligence (AI). In particular, the team identified transition metal halide vdW materials with large magnetic moments that are predicted to be chemically stable using semi-supervised learning. These two-dimensional (2D) vdW magnets have potential applications in data storage, spintronics, and even quantum computing.
Rhone specializes in harnessing materials informatics to discover new materials with unexpected properties that advance science and technology. Materials informatics is an emerging field of study at the intersection of AI and materials science. His team’s latest research was recently featured on the cover of Advanced Theory and Simulations.
2D materials, which can be as thin as a single atom, were only discovered in 2004 and have been the subject of great scientific curiosity because of their unexpected properties. 2D magnets are significant because their long-range magnetic ordering persists when they are thinned down to one or a few layers. This is due to magnetic anisotropy. The interplay with this magnetic anisotropy and low dimensionality could give rise to exotic spin degrees of freedom, such as spin textures that can be used in the development of quantum computing architectures. 2D magnets also span the full range of electronic properties and can be used in high-performance and energy-efficient devices.
Rhone and team combined high-throughput density functional theory (DFT) calculations, to determine the vdW materials’ properties, with AI to implement a form of machine learning called semi-supervised learning. Semi-supervised learning uses a combination of labeled and unlabeled data to identify patterns in data and make predictions. Semi-supervised learning mitigates a major challenge in machine learning — the scarcity of labeled data.
“Using AI saves time and money,” said Rhone. “The typical materials discovery process requires expensive simulations on a supercomputer that can take months. Lab experiments can take even longer and can be more expensive. An AI approach has the potential to speed up the materials discovery process.”
Using an initial subset of 700 DFT calculations on a supercomputer, an AI model was trained that could predict the properties of many thousands of materials candidates in milliseconds on a laptop. The team then identified promising candidate vdW materials with large magnetic moments and low formation energy. Low formation energy is an indicator of chemical stability, which is an important requirement for synthesizing the material in a laboratory and subsequent industrial applications.
“Our framework can easily be applied to explore materials with different crystal structures, as well,” said Rhone. “Mixed crystal structure prototypes, such as a data set of both transition metal halides and transition metal trichalcogenides, can also be explored with this framework.”
“Dr. Rhone’s application of AI to the field of materials science continues to produce exciting results,” said Curt Breneman, dean of Rensselaer’s School of Science. “He has not only accelerated our understanding of 2D materials that have novel properties, but his findings and methods are likely to contribute to new quantum computing technologies.”
Rhone was joined in research by Romakanta Bhattarai and Haralambos Gavras of Renselaer; Bethany Lusch and Misha Salim of Argonne National Laboratory; Marios Mattheakis, Daniel T. Larson, and Efthimios Kaxiras of Harvard University; and Yoshiharu Krockenberger of NTT Basic Research Laboratories. More