AI recommendations: This time it’s personal
Study shows adaptive decision support can fight over-reliance on AI
From doctors diagnosing symptoms to judges intervening in court cases, humans make complex decisions every day. Increasingly, artificial intelligence tools are being used to help in those decisions.
There’s an insidious downside to this type of “help.” Research shows that over-reliance on AI for decision-making can lead to worse or inaccurate choices, and moreover, loss of expertise: over time, a person learns less about the subject, creating a cycle of over-reliance and under-achievement.
Computer scientists at the Harvard John A. Paulson School of Engineering and Applied Sciences (SEAS) offer a potential solution to this cycle. In a recent paper, they argue that AI shouldn’t offer one-size-fits-all decision support, as is typical today, but rather should be adaptive, or able to adjust to the situation and to the uniqueness of each user.
Now, they've developed an AI recommendation model that incorporates reinforcement learning, a machine learning method in which an AI system learns which actions to take based on continuous feedback. This model doesn’t just spit out answers — it decides in the moment how and to what extent to help the human.
In online experiments with more than 1,000 participants, the researchers demonstrated that reinforcement learning improved human-AI performance more than any other type of AI support that’s used today.
The research, led by Zana Buçinca, a recent Harvard computer science Ph.D. graduate and current MIT faculty member, is published in ACM Transactions on Computer-Human Interaction and will be presented later this year at the ACM Symposium on User Interface Software and Technology (UIST).
“Given this worrisome trend of human over-reliance on AI, we wanted to instead design AI that accounts for and optimizes for how the human processes its advice,” said Buçinca, who co-authored the work with former advisor Krzysztof Gajos, the Yahn W. Bernier and N. Elizabeth McCaw Professor of Computer Science at Harvard, and Maja Malaya, a student at Technical University of Łódź in Poland.
How humans and AIs make decisions together
Earlier work by Buçinca and colleagues tested how humans and AIs work together to make decisions. They uncovered that humans tend to over-rely on AI recommendations by accepting incorrect suggestions, even when they could have made the right decision on their own.
Their previous work also uncovered individual differences in how people received information from AI. Some people naturally have more “need for cognition,” — that is, they enjoy and are motivated by analytical thinking – while others desire less to think deeply. This finding underscored the need for AI assistance to adjust to situational context, including the individual characteristics of the human decision-makers.
In the new system, the reinforcement learning agent observes the human-AI as a pair, accounting for the person’s dynamic assessment of the skill, their need for cognition, and how confident the AI model is. It chooses from several interaction strategies, such as showing a full recommendation; providing a partial explanation; or withholding an answer altogether. The system is trained to choose among these options to maximize a specified objective, whether that’s immediate accuracy to the task, or longer-term human learning.
To evaluate their approach, the team designed a decision task modeled on real health-care data. Human participants were shown vignettes of fictitious patients with different health needs and goals and were then asked to select the most appropriate prescription for exercise, like pilates, weight-lifting, etc.
Across two online experiments of 316 and 964 participants each, participants first completed a baseline assessment to measure their initial skill on the task, and they answered survey questions to measure need for cognition. They then made a series of decisions about different patients, but under different sets of conditions, such as reinforcement learning optimized for accuracy; reinforcement learning optimized to support longer-term learning; and baseline non-adaptive AI supports that always provided decision recommendations accompanied by explanations.
In both experiments, people interacting with reinforcement learning optimized for accuracy achieved significantly higher decision accuracy than those with non-adaptive support that gave them the answers. In many cases, the learned policies enabled human-AI complementarity, where the human-AI team outperformed both humans alone and the AI system alone.
The findings have implications for emerging AI regulation, Buçinca added. Many policy frameworks, including the European Union AI Act, call for human oversight of high‑stakes AI systems with the implicit assumption that adding a human decision-maker on top of an algorithm will mitigate errors.
Buçinca and Gajos’s research shows that, without intentional design of how the human and AI interact, human oversight of AI systems could be undermined. Reinforcement learning could be a practical tool to discover assistance strategies that both improve accuracy and support human skills, they contend.
“Our paper shows that psychological needs are not just intangible things beyond the direct grasp of computer scientists, but rather, they are things we can model and incorporate as objectives for our optimization algorithms,” Gajos said. “In other words, as respectable engineers, we can optimize for human happiness.”
Center for Human-driven AI Research and Methods at Harvard
The study is part of the research agenda of a new center at Harvard, the Center for Human-driven AI Research and Methods, or CHARM. The initiative brings together different areas of science to develop AI systems that advance human values, rather than simply automating tasks. Within CHARM, researchers like Gajos and Buçinca are focused on “worker-centric AI”: systems that not only help people do their jobs better today, but also support their long-term competence, autonomy, and sense of meaning at work.
The research received federal support from the National Science Foundation under grant No. IIS-2107391 and by the Office of Naval Research under agreement No. N00014-24-1-2726. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation or the Office of Naval Research.
Journal
ACM Transactions on Computer-Human Interaction
Method of Research
Observational study
Subject of Research
People
Article Title
Offline Reinforcement Learning for Adaptive Support in AI-Assisted Decision-Making
Digital Science launches Papers AI: an AI-native researcher workspace for writing, data and code
Papers AI gives researchers and students an AI-native, full-context home for their projects
image:
The new Papers AI from Digital Science is an AI-native workspace that understands the full context of your work, so you're not starting fresh each time.
view moreCredit: Digital Science.
Researchers and students now have a new AI-native workspace, with the launch of Digital Science's Papers AI – where they'll never have to re-explain their work every time they turn to AI for assistance.
Built by the team behind Overleaf, Papers AI is designed specifically for research workflows. Researchers can write, organize their materials and analyze data, with context-aware AI assisting across a whole project, rather than working in one document at a time. From making sense of a paper someone else has written to producing a document of their own, the assistant stays in context throughout the process.
One workspace for the whole project
Papers AI brings research writing, analysis, management and collaborators together in one project, instead of being scattered across separate tools – supporting Word, Markdown, Typst and LaTeX documents alongside Jupyter notebooks, CSV, and Kanban boards.
Papers AI’s fully integrated AI assistant is designed to read a project's drafts, datasets and references together, reducing the need to re-upload files or re-explain context each time a researcher returns to their work.
Data is never used to train models, and users have the option to use Papers AI completely locally, ensuring no data ever leaves their hardware.
Every AI-suggested change appears as a reviewable edit, which a researcher can accept, reject, or adjust before it becomes part of the document – nothing enters a project unreviewed. When changes are made in one file, other files are updated to reflect those changes, so the whole project stays in sync, with a clear record of what changed and why.
Key features include:
- Multi-format writing – draft and organize in Word, Markdown, Typst, or LaTeX within one project
- AI with full project context – the assistant reads drafts, datasets and references together, with no re-uploading or re-explaining required
- Reviewable AI edits – AI-suggested changes are shown alongside the original, for the researcher to accept, reject, or edit
- Code alongside writing – run Jupyter notebook cells and work with CSV data files in place and reuse outputs directly as figures and tables
- Offline-first and private by design – projects live on-device unless a researcher chooses to share or sync them
- Model flexibility – an external or local AI model can be connected instead of, or alongside, the built-in assistant
Designed with researchers, not just for them
Feedback from the research community has been critical to the development of Papers AI.
Amye Kenall, Chief Product Officer, Digital Science, said: "Researchers lose enormous amounts of time in the gaps between tools – re-explaining a project to an AI assistant that's already forgotten it, or ferrying results from one system to another. Papers AI supports the whole research workflow in one place, and it never changes anything quietly: every edit is tracked and every file stays in sync, so researchers stay in control of their own work."
Juan Castro, Principal AI Scientist, Digital Science, said: "Most AI tools only see the page in front of you. Papers AI's assistant sees the whole project – the draft, the dataset, the notebook, the references – so it can actually help with the parts of research that take the most time, not just the writing. Trust matters as much as capability here – researchers need to know they're still the ones making the final call on their own work, and that's something we've built into Papers AI from the start."
Papers AI is available now at papers.ai, with free and paid options, and sign-up handled directly in the app.
Note to editors: The Papers name was previously used for part of ReadCube’s reference management solutions, which is now consolidated under the ReadCube brand. That product is unrelated to Papers AI.
About Digital Science
Digital Science is an AI-focused technology company providing innovative solutions to complex challenges faced by researchers, universities, governments, funders, industry, and publishers. We work in partnership to advance global research for the benefit of society. Through our brands – Altmetric, Dimensions, Figshare, IFI CLAIMS Patent Services, metaphacts, Overleaf, ReadCube, Symplectic, and Writefull – we believe when we solve problems together, we drive progress for all. Visit digital-science.com and follow Digital Science on Bluesky, on X or on LinkedIn.
Seoul National University of Science and Technology develops an AI framework for long-term bridge damage monitoring
Researchers create an automated computer vision system that tracks and measures bridge damage over time using routine drone inspections
Seoul National University of Science & Technology
image:
Researchers develop new AI based framework that allows long-term monitoring of bridge damage with less than 5% measurement error
view moreCredit: Assistant Professor Hyunjun Kim from Seoul National University of Science and Technology
Bridges form an important part of the road systems. Monitoring them is essential for ensuring structural safety as overtime these bridges develop cracks, concrete spalling, and water leakage due to traffic loads, weather, and environmental exposure. Traditional visual inspections are often labor-intensive, costly and sometimes hazardous. Computer vision has made automated bridge inspections more practical. However, comparing damage captured months apart remains difficult because images are typically taken from different positions and viewing angles.
Now, researchers from Seoul National University of Science and Technology, South Korea, led by Assistant Professor Hyunjun Kim, developed an automated computer vision framework that tracks structural damage over time using drone images collected during routine bridge inspections. Their findings were made available online in Structural Health Monitoring on April 27, 2026.
This study enables long-term monitoring by combining artificial intelligence with three-dimensional (3D) bridge reconstruction. The 3D model of the bridge is built using images captured during the initial inspection. Images from subsequent inspections are then automatically aligned with this model using hierarchical localization and image clustering, allowing the same damage to be identified and compared over time even when photographs are taken from different camera angles or distances. Additionally, the framework uses Global Navigation Satellite System data to convert image measurements into real-world dimensions.
"Long-term structural monitoring requires more than simply detecting damage, it requires understanding how that damage evolves," says Dr. Kim. "Our framework allows engineers to visualize damage progression and measure its severity using images collected during routine inspections."
The system was validated over 120 days by monitoring an in-service prestressed concrete bridge using drone imagery. The framework succeeded in tracking the progression of cracks, spalling, and water leakage throughout the study period despite changes in camera viewpoint. It was able to measure the damaged areas with a maximum error of only 4.61% as compared to conventional manual measurements.
Unlike traditional inspection methods which could analyze damage only at a single point in time or which requires rebuilding 3D models for every inspection, this study suggests continuous monitoring of damage using a single reference model. This approach improves consistency while reducing the computational effort needed for long-term assessments. Although the current method is best suited for relatively flat structural components and its accuracy might differ on highly curved surfaces, it can report most bridge elements commonly encountered during routine inspections.
"We believe this framework can help engineers make more informed maintenance decisions and contribute to extending the service life of critical infrastructure." says Dr. Kim
As transportation agencies face aging infrastructure and increasing maintenance demands, researchers believe their study could support predictive maintenance strategies, improving public safety while reducing long-term inspection and repair costs. In future this technique could also be adapted to monitor other infrastructures such as tunnels, dams, and elevated rail systems.
***
Reference
DOI: 10.1177/14759217261443618
About the institute Seoul National University of Science and Technology (SEOULTECH)
Seoul National University of Science and Technology, commonly known as 'SEOULTECH,' is a national university located in Nowon-gu, Seoul, South Korea. Founded in April 1910, around the time of the establishment of the Republic of Korea, SEOULTECH has grown into a large and comprehensive university with a campus size of 504,922 m2.
It comprises 10 undergraduate schools, 35 departments, 6 graduate schools, and has an enrollment of approximately 14,595 students.
Website: https://en.seoultech.ac.kr/
About Dr. Hyunjun Kim
Dr. Hyunjun Kim is an Assistant Professor in the Department of Civil Engineering at Seoul National University of Science and Technology, South Korea. His research focuses on developing smart technologies for the maintenance and management of civil infrastructure, with particular emphasis on computer vision-based visual inspection and system identification to improve the structural integrity, durability, and resilience of infrastructure systems. He earned his Ph.D. in Civil Engineering from Ulsan National Institute of Science and Technology (UNIST) in 2020 and completed postdoctoral research in the laboratory of Prof. Billie F. Spencer Jr. at the University of Illinois Urbana-Champaign.
Journal
Structural Health Monitoring
Method of Research
Experimental study
Subject of Research
Not applicable
Article Title
Long-term monitoring of damage progression using multi-view images from routine inspections of bridges
Solving the increasing complexity of AI problem-solving one bite at a time
image:
A proverb claims that the best way to eat an elephant is one bite at a time — meaning that a large, overwhelming problem is best approached by breaking it into smaller pieces. Now, a research team based on Tsinghua University have applied the same principle to increasingly complex AI with a framework dubbed the calculus of intelligence.
view moreCredit: iFuture, Tsinghua University Press
“How do you eat an elephant? One bite at a time,” goes the proverb that dates back to the early 1920s. Turns out the same idea may apply to advancing artificial intelligence (AI), according to a team from the Institute for Interdisciplinary Information Sciences (IIIS) at Tsinghua University.
The researchers developed a framework they call the “calculus of intelligence,” or COIN, capable of breaking down the increasingly complex problems agentic AI — the systems expected to operate with minimal human oversight to do things like monitor for cybersecurity breaches, write code and much more — is tasked with solving. They published their approach on July 17 in iFuture.
“The calculus of intelligence is a mathematical framework for decomposing a complex task into smaller, well-defined subtasks that are simple enough to solve, and then combining their solutions into a coherent whole,” said Yang Yuan, associate professor and corresponding author on the paper. “Just as classical calculus can calculate the area under a complicated curve by dividing it into tiny pieces and adding them up, this framework provides a way to build and understand complex systems through the step-by-step composition of numerous simple components.”
The framework translates splices up the overarching AI goal into smaller tasks within a logic system called a Grothendieck topos, which Yuan described as a mathematical rulebook for facilitating the coordination of building something as complex as a large airplane.
“It must be divided into many local components — such as the wings, engines and control systems — which are designed by different teams,” Yuan said, explaining that these limited perspectives are the local views, with each team only seeing the information relevant to its own task rather than every detail of the entire airplane, while still retaining the shared information needed to connect its component to the rest of the system. “This allows different teams to work independently while ensuring that the components they produce match at their shared boundaries.”
The Grothendieck topos expresses the structure of local design, shared interfaces and global coordination in mathematical language. It specifies what each local component can see, what requirements it must meet and how different components must agree where they overlap. According to Yuan, COIN takes the next step by providing the rules for decomposing tasks and reliably recomposing local solutions that comply with the larger system’s rulebook.
“Intelligence resides in structure, and structure can be decomposed, learned and recompose,” Yuan said, noting that complexity does not mean incomprehensibility. “Many systems that appear overwhelmingly complex may simply be waiting for the right decomposition. A good structure can turn a difficult global problem into a collection of clearly bounded local problems that can be analyzed step by step. Through appropriate structural decomposition, many complex tasks can become easier to understand, execute and verify.”
This structure can also lend itself to making extremely large systems with limited intelligence, Yuan pointed out.
“The truly transformative future may not be an infinitely powerful individual intelligence — how some may imagine AI — but rather humans using many limited intelligences to construct systems far beyond the scale that any single person or model could independently understand or complete,” Yuan said.
According to Yuan, COIN is a step in that direction.
“The broader goal is to develop a common mathematical language for intelligence: One that can describe what models learn, how complex tasks are decomposed and how many limited intelligences can be organized into larger systems,” Yuan said.
Andrew Chi-Chih Yao, professor and dean of IIIS, co-authored this paper. Yuan and Yao are also affiliated with the Shanghai QiZhi Institute.
About iFuture
iFuture is a premier open-access journal published by Tsinghua University Press on the SciOpen platform, with academic support from the Institute for Interdisciplinary Information Sciences at Tsinghua University. Led by Turing Award Laureate Prof. Andrew Chi-Chih Yao as Editor-in-Chief, the journal is the core component of the AI Open Alliance. Its core mission is to break through AI’s theoretical bottlenecks and foundational infrastructure.
Journal
iFuture
Article Title
Calculus of intelligence: A topos-monadic framework for agentic workflows
WVU study explores AI’s role in training tomorrow’s psychiatrists
image:
As the use of artificial intelligence for mental health conditions grows for both patients seeking support and psychiatrists treating them, WVU researchers put ChatGPT-5 Pro to the test to determine its abilities on the psychiatry educational front.
view moreCredit: WVU Photo/Davidson Chan
As the use of artificial intelligence for mental health conditions grows for both patients seeking support and psychiatrists treating them, West Virginia University researchers put ChatGPT-5 Pro to the test to determine its abilities on the psychiatry educational front.
Their research showed that human supervision is still a top priority when it comes to ensuring patient safety and ethical use of the tool for medical training.
The WVU School of Medicine team of scientists and physicians collaborated on the study to develop clinical education materials focused on patient chatbot use for students and residents preparing for careers in psychiatry. While psychiatry training has traditionally relied on textbooks and patient interactions, the emergence of people using AI for their mental health concerns has left a knowledge gap.
“With AI, we are connected no matter what, that’s the reality,” said Dr. Wanhong Zheng, professor in the WVU School of Medicine Department of Behavioral Medicine and Psychiatry at the WVU Rockefeller Neuroscience Institute, who co-led the study.
“That brings us to the question of how we can incorporate that into our real-life medical training. Students and residents can read about these new case reports and mental health concerns of people using AI in medical journals, the media, and social media, but they don’t typically see them often enough in real-life clinical settings. We designed this study to cover major psychiatric problems that are known to be related to AI chatbot use.”
The conditions included schizophrenia spectrum disorder, anxiety spectrum disorders, mood spectrum disorders, and major depression and psychosis.
Their findings showed ChatGPT-5 Pro was able to generate realistic and useful psychiatry vignettes with strong diagnostic details and explanations. However, safety evaluations underscored the need for a human-centered approach when using AI to create them.
Their work was recently published in npj Digital Medicine, a Nature Portfolio journal.
To provide simulated training cases, Gangqing “Michael” Hu, who co-led the study and is an associate professor in the WVU School of Medicine Department of Microbiology, Immunology, and Cell Biology, asked ChatGPT-5 Pro to generate clinical vignettes — brief scenarios that include symptoms, history, and behavior — of patients using chatbots for mental health support. The prompts asked the model to include realistic chatbot interactions, diagnostic details, and a multiple-choice question with explanatory answers.
Mohammad Iqbal Nouyed, a postdoctoral fellow in the School of Medicine Department of Microbiology, Immunology, and Cell Biology, assisted Hu with data analysis and statistics for the study.
“It’s important for the physicians to understand the patients’ use of AI tools as well,” Hu said. “You have to consider if the chatbot is contributing to the disease or if it is amplifying the signs and symptoms. For example, if a chatbot validates a patient’s unusual belief instead of challenging it, it could reinforce the symptoms and make them harder to interrupt. We’ve also seen case reports of people developing delusional beliefs after long chatbot interactions, but the causal role of the chatbot is still unclear.”
Hu said the chatbot, by itself, has a strong knowledge base, but the tricky part is how to interact with it.
“A chatbot can be warm and agreeable, but that can be risky if it keeps validating a patient’s unusual beliefs,” Hu explained. “A lot of research should be done on that path, not only about the chatbot competency, but also for safe communication between the chatbot and the person using it.”
Zheng and board-certified psychiatrists Dr. Dilip Chandran, associate professor, and Dr. Daniel Elswick, professor, associate residency director, and vice chair for education, in the WVU School of Medicine Department of Behavioral Medicine and Psychiatry, evaluated the vignettes across four domains: language, accuracy of diagnosis, safety and ethics, and whether the model included enough information to benefit students’ learning.
Scores for language, diagnostic accuracy, and educational value were high. However, the study recommends that if the vignettes are incorporated into digital psychiatry curricula, faculty should moderate discussions and include safeguards such as structured debriefing that reviews diagnostic formulation, patient risk assessment and management, and advice on patient chatbot use.
“We still have concerns about safety,” Zheng said. “For example, if the patient has been using a chatbot and is having some kind of severe depressive symptoms, we want to see whether there are any safety concerns such as suicidal or homicidal evaluations involved. Those are key things we teach our trainees and we want to make sure they assess patient safety first. We want the cases to be relevant and match teaching goals and objectives.”
Zheng said he and the team of evaluators in the Department of Behavioral Medicine and Psychiatry think the AI tool will help medical residents learn more about the new role of chatbot in terms of a psychiatric conditions or clinical course trajectory.
“We feel like this can be incorporated in our didactics which is something to consider for the future because we want to make sure our physicians are AI competent,” Zheng said. “This aligns very well with our goal to not only helping us to learn more about this emerging problem, but it also fits with the Accreditation Council for Graduate Medical Education goal with AI competency.”
As more patients and psychiatrists turn to AI in mental healthcare, Hu and Zheng agree that more research is needed before it’s standardized for mental health purposes.
“One of the next steps could be to collect and analyze patterns from cases reported in medical journals and news media,” Hu said. “That information can help us guide AI to create better training cases. We also need to think about clinical variations, such as patient age, medical history, or existing medical conditions, because those factors can affect how we interpret the chatbot’s role and develop coping strategies.”
Zheng said he sees AI as something connected to our daily lives that requires great care.
“We are not asking people not to use AI, but recognizing the pathological use or potential harm is very important. For people with mental problems, I recommend that AI cannot replace professionals.”
Journal
npj Digital Medicine
Article Title
Evaluation of artificial intelligence-generated vignettes depicting patient chatbot use in psychiatric contexts
Albanese & Chen receive funding for conference aimed at creating growing secure open-source ecosystems in era of AI
George Mason University
Massimiliano Albanese, Professor and Associate Chair for Research, Information Sciences and Technology, Executive Director, Institute for Digital Innovation, College of Engineering and Computing (CEC), and Songqing Chen, Professor, Computer Science, CEC, received funding for: “POSE: Conference: A Community-Wide Convening to Grow Secure Open-Source Ecosystems in the Era of AI.”
The event will be a two-day, community-wide convening aimed at strengthening secure and viable open-source ecosystems in the era of artificial intelligence (AI).
It will bring together participants from academia, industry, government, nonprofit organizations, and open-source communities.
One of the goals will be to identify shared priorities and practical opportunities for improving the long-term stewardship of open-source ecosystems as AI-assisted development, AI agents, and other emerging technologies reshape how software is created, maintained, and used.
The workshop will use facilitated plenary discussions, structured breakout sessions, and hands-on working activities to examine technical, organizational, legal, and other dimensions of secure and viable open-source ecosystems.
Expected outcomes include a synthesis of key themes, identified gaps and opportunities, recommended next steps, and broadly disseminated materials that support future activities and contribute to the advancement of secure and resilient open-source ecosystems.
Albanese and Chen received $438,568 from the National Science Foundation for this project. Funding began in July 2026 and will end in late June 2027.
###
George Mason University is Virginia’s largest public research university. Located near Washington, D.C., George Mason enrolls more than 40,000 students from 130 countries and all 50 states. George Mason has grown rapidly over the past half century and is recognized for its innovation and entrepreneurship, remarkable diversity, and commitment to accessibility. In 2023, the university launched Mason Now: Power the Possible, a $1 billion comprehensive campaign to support student success, research, innovation, community, and stewardship. Learn more at GMU.EDU.
Consumer perspectives on trust in and benefits of artificial intelligence in health care
JAMA Network Open
About the Study: This qualitative study of consumer perspectives on AI in health care found that social license for AI is a conditional and dynamic construct not a fixed state; structural, performance, and relational factors intersected to shape social license. The findings provide evidence-based recommendations for stakeholders designing and implementing AI in clinical settings, highlighting the need for AI tools designed to support both consumers and clinicians in delivering care that is personalized, empathetic, and responsive to patient needs.
Corresponding Author: Tuan Duong, MSc, MD, Faculty of Health, Medicine, and Behavioural Science, Queensland Digital Health Centre, The University of Queensland, Brisbane, Queensland 4006, Australia (tuan.duong@uq.edu.au).
10.1001/jamanetworkopen.2026.26916
To access the embargoed study: Visit our JAMA Network Media Center at this link https://media.jamanetwork.com/
Link to the article in your story
We encourage you to link out to this article in your story using the link below. It includes an access token that will give free access to the article for your readers up to one year after publication. (The link will be live after the article publishes and embargo is lifted.)
Editor's Note: Please see the article for additional information, including full author list, author contributions and affiliations, conflict of interest and financial disclosures, and funding and support.
Journal
JAMA Network Open
AI identifies previously unrecognized health insights in routine sleep studies
Interdisciplinary research team develops foundation model that identifies patient groups with markedly different long-term health risks
Cleveland Clinic
August 3, 2026, 5:00 AM EDT: A novel AI model can use information collected during routine sleep studies to identify patients’ long-term health risks, according to a new study published in Nature Communications. Developed by a multidisciplinary research team, the model uncovered hidden sleep patterns linked to risks including heart disease, cognitive decline and death.
The findings also suggest that routine medical tests may contain substantially more physiologic information than current clinical practice extracts from them. In this case, AI identified meaningful signals in standard overnight sleep study data that are not captured by conventional summary measures alone.
The research revealed clinically meaningful patient subtypes with sharply different long-term health risks. Patients in the highest-risk group had twice the mortality risk over the next five years compared to those in the lowest-risk group, a distinction that was not captured by the standard clinical measure used to assess sleep apnea severity, the apnea-hypopnea index.
Each year, an estimated 1 to 4 million polysomnograms, or in-lab sleep studies, are performed in the United States, typically to evaluate sleep apnea. While these studies collect rich data on each patient’s brains, lungs, muscles and heart, clinicians historically have focused on a small subset of that information to grade sleep apnea severity.
“For decades we have distilled an overnight sleep study into a handful of summary measures,” said Reena Mehra, M.D., professor of medicine at the University of Washington and the study’s senior clinical author. “AI gives us the opportunity to move beyond those summaries and learn from the full richness of sleep physiology.”
The model was developed by a collaborative team of sleep physicians, AI researchers, data scientists and neuroscientists brought together through the Discovery Accelerator, a 10-year joint research partnership between Cleveland Clinic and IBM aimed at advancing the pace of discovery in life sciences through AI and quantum computing.
Using data from the Cleveland Clinic Sleep Signals, Testing, and Reports Linked to Patient Traits (STARLIT) registry, the researchers grouped patients into five risk categories. The model also predicted outcomes well for men and women, while the apnea hypopnea index has historically performed better in men. The findings were independently confirmed in a nationwide patient cohort.
“Modern AI lets us recover much more of the information contained in a night’s worth of sleep physiology, revealing clinically meaningful patient groups with very different long-term health risks,” said Jeffrey Rogers, Ph.D., the corresponding author and professor adjunct, neurosurgery, Yale School of Medicine. “These findings demonstrate that routine medical tests can contain substantially more physiologic information than current clinical practice extracts from them.”
The model could also help researchers better understand how sleep impacts health outcomes. By looking beyond traditional measures, the approach uses AI to detect latent physiologic features invisible to the human eye and extract prognostic biomarkers that help stratify risk for cardiovascular and neurologic disease, and survival, opening the door to earlier and more personalized care.
“Sleep is foundational to health and wellness,” said Matheus Lima Diniz Araujo, Ph.D., a sleep researcher at Cleveland Clinic. “Nearly 70 million Americans live with chronic disorders of sleep and wakefulness, affecting daily functioning and overall health. This discovery offers a more personalized approach to sleep medicine, by potentially expanding the value of routine sleep testing and reinforcing the key role sleep plays in chronic disease.”
Carl Saab, Ph.D., a professor of biomedical engineering and Chief Scientist of Cleveland Clinic’s Discovery Accelerator, said, “The next step is to validate these findings in diverse populations and expand collaborations among medical and technical experts, industry partners and professional society stakeholders.”
“Sleep is increasingly recognized as a critical component of health, yet the physiological information captured during sleep remains largely underused,” said Erhan Bilal, Ph.D., lead author of the study. “Because everyone sleeps, sleep studies offer a remarkable window into human health that extends far beyond the diagnosis of sleep disorders. Our work shows how foundation models can begin to unlock the richness of these complex signals. And this is only the beginning.”
The research team included Erhan Bilal, Ph.D.; Matheus Lima Diniz Araujo, Ph.D.; Kristen Beck, Ph.D.; Catherine Heinzinger, D.O.; Samer Ghosn, B.S.; Nancy Foldvary-Schaefer, D.O.; Carl Saab, Ph.D.; Jeffrey Rogers, Ph.D.; and Reena Mehra, M.D.
PaveX receives NSF funding to help
agencies plan road repairs sooner, more
efficiently
$305,000 SBIR grant will develop AI-sensor-based system to collect street-level data and produce road condition ratings
image:
High-tech startup PaveX has received a $305,000 Phase I Small Business Innovation Research grant from the National Science Foundation. It will fund an 18-month project to develop the company’s patent-pending, vehicle-mounted system to assess road infrastructure.
view moreCredit: Purdue University photo/Kevin Crisp
WEST LAFAYETTE, Ind. — PaveX, a Purdue University-related startup that uses advanced AI to automate road condition assessments and help public works departments maintain infrastructure, has received a $305,000 Phase I Small Business Innovation Research grant from the National Science Foundation (NSF) to develop its technology.
Mohammad Jahanshahi, CEO and founder, said the broad impact of the 18-month project is to develop the company’s patent-pending, low-cost, vehicle-mounted sensing approach that can be deployed widely to state, national and international infrastructure agencies.
“The system will collect street-level data and produce practical road condition ratings that help agencies plan repairs sooner and more efficiently,” he said. “By enabling more timely maintenance decisions, the project can reduce vehicle damage costs while improving safety.”
Jahanshahi is an associate professor in Purdue’s Lyles School of Civil and Construction Engineering with a courtesy appointment in the Elmore Family School of Electrical and Computer Engineering. He developed the AI platform and disclosed it to the Purdue Innovates Office of Technology Commercialization, which applied for a patent to protect the intellectual property and granted PaveX the exclusive license to commercialize the technology.
Project details and success targets
Jahanshahi said the project will investigate PaveX’s automation framework.
“The framework enables comprehensive, lane-level road coverage using low-cost mobile sensors while reducing two key bottlenecks: manual route planning and the manual validation of uncertain pavement-distress detections,” he said.
Jahanshahi said the project also will improve distress detection reliability by developing a Bayesian multiframe data fusion framework that integrates redundant observations across consecutive frames.
“Rather than treat each image independently, the fusion approach will combine evidence over time to suppress false positives and strengthen consistent detections, improving accuracy and reducing reliance on human review,” he said.
Phase I success targets include measurable reductions in route planning overhead and improved detection performance under real-world variability.
PaveX national milestones
Since January 2025, PaveX has surveyed and assessed more than 6,200 miles of roads in Indiana, Michigan, North Carolina, Utah and California.
Jahanshahi said the NSF-funded project is an important step toward scaling PaveX’s technology for widespread development across transportation networks.
“In the future, PaveX could be integrated into autonomous vehicles, enabling them to continuously collect roadway-condition data as they travel and provide agencies with more frequent, comprehensive infrastructure assessments,” he said.
About PaveX
PaveX merges cutting-edge AI technology with advanced sensors to provide a seamless, cost-effective approach to road condition monitoring. Our mission is to empower communities with smarter, safer infrastructure.
About Purdue Innovates Office of Technology Commercialization
The Purdue Innovates Office of Technology Commercialization operates one of the most comprehensive technology transfer programs among leading research universities in the U.S. Services provided by this office support the economic development initiatives of Purdue University and benefit the university’s academic activities through commercializing, licensing and protecting Purdue intellectual property. In fiscal year 2025, the office reported 161 deals executed with 269 technologies licensed, 479 invention disclosures received, and 267 U.S. and international patents received. The office is managed by the Purdue Research Foundation, a private, nonprofit foundation created to advance the mission of Purdue University. Contact otcip@prf.org for more information.
About Purdue University
Purdue University is a research institution ranked among the top 10 public universities in the United States. More than 106,000 students study at Purdue across multiple campuses, including more than 57,000 at our main campus locations in West Lafayette and Indianapolis. As a land-grant university committed to affordability and accessibility, Purdue’s main campus has frozen tuition 14 years in a row, enabling more students than ever to graduate debt-free.
Study shows AI could support low-cost foot health technology
Researchers have produced an AI model capable of reconstructing detailed foot pressure maps using information about foot shape and a small number of anatomical pressure points
University of Queensland
video:
Plantar pressure analysis is widely used to assess a person's foot function and balance, gait mechanics and foot health, and to inform the design of orthotics that aim to minimise the risk and progression of foot-related pathologies.
view moreCredit: Healthia Ltd
Researchers have demonstrated AI tools could play an important part in the development of simpler, lower-cost technologies for monitoring foot health, particularly in settings where access to specialised equipment may be limited.
A collaboration between The University of Queensland, iOrthotics and Healthia Limited has produced an AI model capable of reconstructing detailed foot pressure maps using information about foot shape and a small number of anatomical pressure points.
Applied mechanics engineer UQ Emeritus Professor Martin Veidt said plantar pressure analysis was widely used to assess a person's foot function and balance, gait mechanics and foot health, and to inform the design of orthotics that aim to minimise the risk and progression of foot-related pathologies.
“But existing measurement methods have limitations and are often costly and inaccessible for people living in rural and remote regions," Professor Veidt said.
UQ materials engineer Dr Stuart McDonald said in-shoe systems offer greater mobility and extended pressure monitoring but typically rely on a large number of sensors which can increase cost, complexity and power requirements.
“This study looked at the potential for AI to overcome some of the challenges associated with traditional plantar pressure monitoring systems,” Dr McDonald said.
The researchers from UQ's School of Mechanical and Mining Engineering collaborated with iOrthotics and Healthia Limited to explore how a multimodal deep learning system could help unlock practical and more accessible methods to reconstruct dense plantar pressure information from sparse sensing.
Using anatomical foot information and plantar pressure measurements from 35 study participants, UQ PhD student Chongguang Wang was able to build an artificial neural network framework capable of generating accurate, high-resolution pressure maps using significantly fewer physical sensors.
The proposed deep learning model achieved its best performance using just 16 anatomical ‘landmarks’ from the bottom of the foot and achieved a comparable result using only 2 landmarks, indicating promising reconstruction performance even under very limited sensing conditions.
"This research demonstrates that by combining information about foot shape with only a small number of anatomical inputs, it is possible to reconstruct detailed plantar pressure distributions with a high degree of accuracy,” Professor Veidt said.
"The beauty is that data collection could feasibly take place anywhere, including in isolated communities where health outcomes are poor and services may be limited."
The UQ-led study was part of a suite of research initiated by iOrthotics and parent company Healthia, together with researchers from QUT, through a $2.2 million Federal Government Cooperative Research Centres Projects (CRC-P) grant to develop smarter orthotic technology for people living in rural and remote regions.
Healthia’s group chief education and research officer Kerrie Evans said the research formed part of a broader program at Healthia and iOrthotics exploring how emerging technologies could improve access to foot-health assessment and monitoring.
"Foot complications, including diabetic foot ulcers and amputations, continue to have a significant impact on individuals and health systems," Associate Professor Evans said.
"Our goal is to support the development of practical, affordable technologies that can help clinicians better understand foot function and identify potential problems earlier.”
"While more research is needed, particularly in clinical populations and real-world settings, these findings demonstrate the potential for AI to play an important role in future foot-health monitoring technologies”.
The research is published in Sensors.
Journal
Sensors
Method of Research
Experimental study
Subject of Research
People
Article Title
Multimodal Feature-Level Fusion CBAM U-Net for Static Plantar Pressure Prediction Using Plantar Geometry and Sparse Anatomical Landmarks
COI Statement
Authors Kerrie Evans, Dean Hartley were employed by the company Healthia Limited. Author Scott Morrison was employed by the company iOrthotics Pty Ltd. The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
No comments:
Post a Comment