bioinnovationbeacon.com

Category: AI in Biotechnology

  • Artificial intelligence tools make education materials more patient friendly

    Artificial intelligence tools make education materials more patient friendly

    Artificial intelligence (AI) tools significantly improve the readability of online patient education materials (PEMs), making them more accessible, a new study shows.

    Led by researchers at NYU Langone Health, the study focused on the readability of PEMs available on the websites of the American Heart Association (AHA), American Cancer Society (ACS), and American Stroke Association (ASA). According to the researchers, these materials help patients make decisions about their health care but often exceed the recommended reading level of grade 6, making them difficult for many patients to understand.

    For the study, researchers evaluated the capabilities of three large language models (LLMs) — ChatGPT, Gemini, and Claude — to optimize the readability of PEMs without compromising accuracy. These generative AI tools are designed to simplify complex texts by predicting the next word in a sentence based on extensive Internet data. This next-word prediction gives such models the ability to rewrite any article in simpler language as directed.

    Published online April 10 in the Journal of Medical Internet Research, the study involved 60 randomly selected PEMs from the AHA, ACS, and ASA websites. Researchers prompted the LLMs to simplify the reading level of the materials. Results showed that the original readability scores were significantly above the recommended level of grade 6, with mean grade-level scores of 10.7, 10, and 9.6, respectively.

    After optimization by the LLMs, readability scores improved significantly across all three websites. ChatGPT improved readability to a mean grade level of 7.6, Gemini to 6.6, and Claude to 5.6. Word counts were also significantly reduced, making the materials more concise.

    “Our study shows that widely used large language models have the potential to transform patient education materials into more readable content, which is essential for patient empowerment and better health outcomes,” said study senior author Jonah Feldman, MD, medical director of transformation and informatics at NYU Langone.

    “Our findings demonstrate that even expert-composed education materials, which are already patient-directed, can benefit from AI-driven improvements,” said Feldman, who also serves as an assistant professor at NYU Grossman Long Island School of Medicine.

    This study, the researchers say, provides an example of how healthcare organizations can apply AI to make clinical communication more patient friendly. Prior studies demonstrated the capabilities of AI models to create patient-focused explanations of heart test results, to draft responses to electronic advice queries, and to generate human-friendly summaries of complex medical reports.

    “The breadth of possible AI offerings shows how technology can be leveraged to transform the patient experience across health care systems, and not just in the United States,” said study co-author Paul Testa, MD, JD, MPH, chief health informatics officer at NYU Langone.

    “These studies are not just theoretical — after demonstrating their effectiveness, we are actively putting these AI tools into practice,” said Testa, who is also a clinical professor at NYU Grossman School of Medicine.

    According to Testa, the NYU Langone team is already using the same AI tools in a randomized controlled trial that incorporates AI-generated, patient-friendly summaries for hospital discharge instructions, with the goal to evaluate their effectiveness in improving patient comprehension and satisfaction. The researchers hope to show that providing clear and accessible discharge instructions will help ensure better postdischarge care and smoother transitions.

    “Generating real-world evidence through randomized trials is crucial for validating the effectiveness of AI tools in clinical settings,” said study co-author Jonah Zaretsky, MD, associate chief of medicine at NYU Langone Hospital — Brooklyn. “This approach ensures that the AI-generated documentation is not only accurate but also genuinely beneficial for patients and their families,” added Zaretsky, a clinical assistant professor at NYU Grossman School of Medicine.

    The study was self-funded by NYU Langone. Besides Feldman, Testa, and Zaretsky, NYU Langone researchers involved in the study were lead author John Will, and co-authors Mahin Gupta and Aliesha Dowlath.

  • AI tool grounded in evidence-based medicine outperformed other AI tools — and most doctors- on USMLE exams

    AI tool grounded in evidence-based medicine outperformed other AI tools — and most doctors- on USMLE exams

    A powerful clinical artificial intelligence tool developed by University at Buffalo biomedical informatics researchers has demonstrated remarkable accuracy on all three parts of the United States Medical Licensing Exam (Step exams), according to a paper published today (April 22) in JAMA Network Open.

    Achieving higher scores on the USMLE than most physicians and all other AI tools so far, Semantic Clinical Artificial Intelligence (SCAI, pronounced “Sky”) has the potential to become a critical partner for physicians, says lead author Peter L. Elkin, MD, chair of the Department of Biomedical Informatics in the Jacobs School of Medicine and Biomedical Sciences at UB and a physician with UBMD Internal Medicine.

    Elkin says SCAI is the most accurate clinical AI tool available to date, with the most advanced version scoring 95.2% on Step 3 of the USMLE, while a GPT4 Omni tool scored 90.5% on the same test.

    “As physicians, we are used to using computers as tools,” he explains, “but SCAI is different; it can add to your decision-making and thinking based on its own reasoning.”

    The tool can respond to medical questions posed by clinicians or the public at https://halsted.compbio.buffalo.edu/chat/.

    The researchers tested the model against the USMLE, required for licensing physicians nationwide, which assesses the physician’s ability to apply knowledge, concepts and principles, and to demonstrate fundamental patient-centered skills. Any questions with a visual component were eliminated.

    Elkin explains that most AI tools function by using statistics to find associations in online data that allow them to answer a question. “We call these tools generative artificial intelligence,” he says. “Some have postulated that they are just plagiarizing what’s on the internet because the answers they give you are what others have written.” However, these AI models are now becoming partners in care rather than simple tools for clinicians to utilize in their practice, he says.

    “But SCAI answers more complex questions and performs more complex semantic reasoning,” he says, “We have created knowledge sources that can reason more the way people learn to reason while doing their training in medical school.”

    The team started with a natural language processing software they had previously developed. They added vast amounts of authoritative clinical information gleaned from widely disparate sources ranging from recent medical literature and clinical guidelines to genomic data, drug information, discharge recommendations, patient safety data and more. Any data that might be biased, such as clinical notes, were not included.

    13 million medical facts

    SCAI contains 13 million medical facts, as well as all the possible interactions between those facts. The team used basic clinical facts known as semantic triples (subject-relation-object, such as “Penicillin treats pneumococcal pneumonia”) to create semantic networks. The tool can then represent these semantic networks so that it is possible to draw logical inferences from them.

    “We have taught large language models how to use semantic reasoning,” says Elkin.

    Other techniques that contributed to SCAI include knowledge graphs that are designed to find new links in medical data as well as previously “hidden” patterns, as well as retrieval-augmented generation, which allows the large language model to access and incorporate information from external knowledge databases before responding to a prompt. This reduces “confabulation,” the tendency for AI tools to always respond to a prompt even when it doesn’t have enough information to go on.

    Elkin adds that using formal semantics to inform the large language model provides important context necessary for SCAI to understand and respond more accurately to a particular question.

    ‘It can have a conversation with you’

    “SCAI is different from other large language models because it can have a conversation with you and as a human-computer partnership can add to your decision-making and thinking based on its own reasoning,” Elkin says.

    He concludes: “By adding semantics to large language models, we are providing them with the ability to reason similarly to the way we do when practicing evidence-based medicine.”

    Because it can access such vast amounts of data, SCAI also has the potential to improve patient safety, improve access to care and “democratize specialty care,” Elkin says, by making medical information on specialties and subspecialties accessible to primary care providers and even to patients.

    While the power of SCAI is impressive, Elkin stresses its role will be to augment, not replace, physicians.

    “Artificial intelligence isn’t going to replace doctors,” he says, “but a doctor who uses AI may replace a doctor who does not.”

    In addition to Elkin, UB co-authors from the Department of Biomedical Informatics are Guresh Mehta; Frank LeHouillier; Melissa Resnick, PhD; Crystal Tomlin, PhD; Skyler Resendez, PhD; and Jiaxing Liu.

    Sarah Mullin, PhD, of Roswell Park Comprehensive Cancer Center, and Jonathan R. Nebeker, MD, and Steven H. Brown, MD, both of the Department of Veterans Affairs, also are co-authors.

    The work was funded by grants from the National Institutes of Health and the Department of Veterans Affairs.

  • Our DNA is at risk of hacking, warn scientists

    Our DNA is at risk of hacking, warn scientists

    Next-generation DNA sequencing (NGS) — the same technology which is powering the development of tailor-made medicines, cancer diagnostics, infectious disease tracking, and gene research — could become a prime target for hackers.

    A study, published in IEEE Access, highlights growing concerns over how this powerful sequencing tool — if left unsecured — could be exploited for data breaches, privacy violations, and even future biothreats.

    Led by Dr Nasreen Anjum from the University of Portsmouth’s School of Computing, it is the first comprehensive research study of cyber-biosecurity threats across the entire NGS workflow.

    NGS is a cornerstone of modern biotechnology, enabling rapid and cost-effective DNA and RNA sequencing that supports important advances in cancer research, drug development, agricultural innovation and forensic science.

    The sequencing process within NGS workflow involves a series of complex, interdependent steps — from sample preparation and sequencing to data analysis and interpretation — each involving highly specialised instruments, technologies, software, and connected systems.

    While the steps are essential for generating accurate results, they also open up multiple points of vulnerability. As many DNA datasets are openly accessible online, the study warns it is possible for cybercriminals to misuse the information for surveillance, manipulation, or malicious experimentation.

    Dr Anjum said: “Our work is a wake-up call. Protecting genomic data isn’t just about encryption — it’s about anticipating attacks that don’t yet exist. We need a paradigm shift in how we secure the future of precision medicine.”

    The research was carried out in collaboration with colleagues from the Department of Computer Science at Anglia Ruskin University in Cambridge, the University of Gloucestershire’s Department of Cyber Security and Computing, Narjan University’s Department of Computer Science and Emerging Research Laboratory, and the Department of Microbiology at Shaheed Benazir Bhutto Women University.

    Dr Mahreen-Ul-Hassan, microbiologist and co-author from the Shaheed Benazir Bhutto Women University, said: “Genomic data is one of the most personal forms of data we have. If compromised, the consequences go far beyond a typical data breach.”

    The research team identified new and emerging methods that hackers and those with malicious intent could use to exploit or attack systems, such as synthetic DNA-encoded malware, AI-driven manipulation of genome data, and identity tracing through re-identification techniques. These threats go beyond typical data breaches, posing risks to individual privacy, scientific integrity, and national security.

    Dr Anjun added: “Despite its importance, cyber-biosecurity remains one of the most neglected and poorly understood research disciplines and is leaving a critical gap in global biosecurity. To make sure our DNA information stays safe and is used only for good, we’re urging more research and collaboration to find ways to keep this powerful technology secure.

    “Governments, regulatory bodies, funding agencies, and academic institutions must prioritise this field and invest in dedicated research, education, and policy development before it’s too late.

    “Without coordinated action, genomic data could be exploited for surveillance, discrimination, or even bioterrorism. Current protections are fragmented, and vital collaboration between disciplines is lacking. Key to successful prevention will be interdisciplinary cooperation between computer scientists, bioinformaticians, biotechnologists, and security professionals — groups that rarely work together but must align.

    “Our research lays the foundations for improving biosecurity by providing a single, clear list of all the possible threats in the entire next-generation sequencing process.”

    The paper also recommends practical solutions, including secure sequencing protocols, encrypted storage, and AI-powered anomaly detection, creating a foundation for much stronger cyber-biosecurity.

    The study was funded by the British Council’s UK-Saudi Challenge Fund and a Quality Related Research Grant from the University of Portsmouth.

  • Powerful new software platform could reshape biomedical research by making data analysis more accessible

    Powerful new software platform could reshape biomedical research by making data analysis more accessible

    A powerful new software platform called the Playbook Workflow Builder is set to transform biomedical research by allowing scientists to conduct complex and customized data analyses without advanced programming skills. An article that describes the new platform was published in the April 3online issue of the journal PLOS Computational Biology.

    Developed by a multi-institutional team that was led by Icahn School of Medicine at Mount Sinai investigators as part of the National Institutes of Health Common Fund Data Ecosystem (CFDE) program, researchers from across the United States developed the web-based platform that enables scientists to analyze and visualize their own data independently through an intuitive, interactive interface.

    Traditionally, experimental biologists rely heavily on bioinformaticians to process and analyze the large datasets they collect. The new software platform changes this paradigm by offering a modular, user-friendly system where scientists can design custom workflows using pre-built analytical components — akin to assembling “LEGO® pieces,” say the investigators.

    “This approach could fundamentally reinvent data analysis in biomedical research and beyond. By allowing researchers to conduct sophisticated analyses without requiring advanced programming skills, it removes a major barrier to data-driven discovery,” says first author Daniel J.B. Clarke, MS, a data scientist in the lab of senior corresponding author Avi Ma’ayan, PhD, Professor, Pharmacological Sciences, and Director of the Mount Sinai Center for Bioinformatics at the Icahn School of Medicine at Mount Sinai. “With greater accessibility and efficiency, scientists across disciplines can explore complex datasets more easily, uncover new insights faster, and ultimately accelerate the pace of scientific publication and discovery.”

    Users of the Playbook Workflow Builder engage with an intuitive user interface, clicking on cards to construct their data analysis workflows. Alternatively, users can interface with a large language model-powered chatbot to design and build their data analysis pipelines. As they upload their data and input their instructions, the system automatically generates detailed documentation, including interactive figures, clear figure legends, and step-by-step automatically generated method descriptions.

    This ensures that the entire workflow is well-organized and easy to follow. Additionally, researchers can export their workflows in multiple formats, making it easier to share their work with colleagues, which also enhances reproducibility across different studies.

    Currently, the team is actively working to expand the platform’s capabilities by adding more analytical components, refining the user experience, and fostering a larger community of users and contributors.

    “We see this as just the beginning. By reducing reliance on bioinformatics specialists, this platform has the potential to dramatically accelerate the pace of research by enabling experimental biologists to perform sophisticated analyses independently,” says Dr. Ma’ayan “As more scientists adopt this tool, we expect it to streamline workflows, speed up the sharing of findings, and inspire new AI-driven systems that further enhance scientific reasoning and data interpretation. In the years to come, this could become an essential resource for biomedical research, unlocking new insights and driving innovation across the field.”

  • New AI tool generates high-quality images faster than state-of-the-art approaches

    New AI tool generates high-quality images faster than state-of-the-art approaches

    The ability to generate high-quality images quickly is crucial for producing realistic simulated environments that can be used to train self-driving cars to avoid unpredictable hazards, making them safer on real streets.

    But the generative AI techniques increasingly being used to produce such images have drawbacks. One popular type of model, called a diffusion model, can create stunningly realistic images but is too slow and computationally intensive for many applications. On the other hand, the autoregressive models that power LLMs like ChatGPT are much faster, but they produce poorer-quality images that are often riddled with errors.

    Researchers from MIT and NVIDIA developed a new approach that brings together the best of both methods. Their hybrid image-generation tool uses an autoregressive model to quickly capture the big picture and then a small diffusion model to refine the details of the image.

    Their tool, known as HART (short for Hybrid Autoregressive Transformer) can generate images that match or exceed the quality of state-of-the-art diffusion models, but do so about nine times faster.

    The generation process consumes fewer computational resources than typical diffusion models, enabling HART to run locally on a commercial laptop or smartphone. A user only needs to enter one natural language prompt into the HART interface to generate an image.

    HART could have a wide range of applications, such as helping researchers train robots to complete complex real-world tasks and aiding designers in producing striking scenes for video games.

    “If you are painting a landscape, and you just paint the entire canvas once, it might not look very good. But if you paint the big picture and then refine the image with smaller brush strokes, your painting could look a lot better. That is the basic idea with HART,” says Haotian Tang PhD ’25, co-lead author of a new paper on HART.

    He is joined by co-lead author Yecheng Wu, an undergraduate student at Tsinghua University; senior author Song Han, an associate professor in the Department of Electrical Engineering and Computer Science (EECS), a member of the MIT-IBM Watson AI Lab, and a distinguished scientist of NVIDIA; as well as others at MIT, Tsinghua University, and NVIDIA. The research will be presented at the International Conference on Learning Representations.

    The best of both worlds

    Popular diffusion models, such as Stable Diffusion and DALL-E, are known to produce highly detailed images. These models generate images through an iterative process where they predict some amount of random noise on each pixel, subtract the noise, then repeat the process of predicting and “de-noising” multiple times until they generate a new image that is completely free of noise.

    Because the diffusion model de-noises all pixels in an image at each step, and there may be 30 or more steps, the process is slow and computationally expensive. But because the model has multiple chances to correct details it got wrong, the images are high-quality.

    Autoregressive models, commonly used for predicting text, can generate images by predicting patches of an image sequentially, a few pixels at a time. They can’t go back and correct their mistakes, but the sequential prediction process is much faster than diffusion.

    These models use representations known as tokens to make predictions. An autoregressive model utilizes an autoencoder to compress raw image pixels into discrete tokens as well as reconstruct the image from predicted tokens. While this boosts the model’s speed, the information loss that occurs during compression causes errors when the model generates a new image.

    With HART, the researchers developed a hybrid approach that uses an autoregressive model to predict compressed, discrete image tokens, then a small diffusion model to predict residual tokens. Residual tokens compensate for the model’s information loss by capturing details left out by discrete tokens.

    “We can achieve a huge boost in terms of reconstruction quality. Our residual tokens learn high-frequency details, like edges of an object, or a person’s hair, eyes, or mouth. These are places where discrete tokens can make mistakes,” says Tang.

    Because the diffusion model only predicts the remaining details after the autoregressive model has done its job, it can accomplish the task in eight steps, instead of the usual 30 or more a standard diffusion model requires to generate an entire image. This minimal overhead of the additional diffusion model allows HART to retain the speed advantage of the autoregressive model while significantly enhancing its ability to generate intricate image details.

    “The diffusion model has an easier job to do, which leads to more efficiency,” he adds.

    Outperforming larger models

    During the development of HART, the researchers encountered challenges in effectively integrating the diffusion model to enhance the autoregressive model. They found that incorporating the diffusion model in the early stages of the autoregressive process resulted in an accumulation of errors. Instead, their final design of applying the diffusion model to predict only residual tokens as the final step significantly improved generation quality.

    Their method, which uses a combination of an autoregressive transformer model with 700 million parameters and a lightweight diffusion model with 37 million parameters, can generate images of the same quality as those created by a diffusion model with 2 billion parameters, but it does so about nine times faster. It uses about 31 percent less computation than state-of-the-art models.

    Moreover, because HART uses an autoregressive model to do the bulk of the work — the same type of model that powers LLMs — it is more compatible for integration with the new class of unified vision-language generative models. In the future, one could interact with a unified vision-language generative model, perhaps by asking it to show the intermediate steps required to assemble a piece of furniture.

    “LLMs are a good interface for all sorts of models, like multimodal models and models that can reason. This is a way to push the intelligence to a new frontier. An efficient image-generation model would unlock a lot of possibilities,” he says.

    In the future, the researchers want to go down this path and build vision-language models on top of the HART architecture. Since HART is scalable and generalizable to multiple modalities, they also want to apply it for video generation and audio prediction tasks.

    This research was funded, in part, by the MIT-IBM Watson AI Lab, the MIT and Amazon Science Hub, the MIT AI Hardware Program, and the National Science Foundation. The GPU infrastructure for training this model was donated by NVIDIA.

  • Most current AI struggles to read clocks and calendars

    Most current AI struggles to read clocks and calendars

    Some of the world’s most advanced AI systems struggle to tell the time and work out dates on calendars, a study suggests.

    While AI models can perform complex tasks such as writing essays and generating art, they have yet to master some skills that humans carry out with ease, researchers say.

    A team from the University of Edinburgh has shown that state-of-the-art AI models are unable to reliably interpret clock-hand positions or correctly answer questions about dates on calendars.

    Unlike simply recognising shapes, understanding analogue clocks and calendars requires a combination of spatial awareness, context and basic maths – something that remains challenging for AI, the team says.

    Overcoming this could enable AI systems to power time-sensitive applications like scheduling assistants, autonomous robots and tools for people with visual impairments, researchers say.

    The team tested if AI systems that process text and images – known as multimodal large language models (MLLMs) – can answer time-related questions by looking at a picture of a clock or a calendar.

    Researchers tested various clock designs, including some with Roman numerals, with and without second hands, and different coloured dials.

    Their findings show that AI systems, at best, got clock-hand positions right less than a quarter of the time. Mistakes were more common when clocks had Roman numerals or stylised clock hands. AI systems also did not perform any better when the second hand was removed, suggesting there are deep-seated issues with hand detection and angle interpretation, the team says.

    The researchers asked AI models to answer a range of calendar-based questions, such as identifying holidays and working out past and future dates. The team found that even the best-performing AI model got date calculations wrong one-fifth of the time.

    The findings are reported in a peer-reviewed paper that will be presented at the Reasoning and Planning for Large Language Models workshop at The Thirteenth International Conference on Learning Representations (ICLR) in Singapore on 28 April 2025.

    Rohit Saxena, of the University of Edinburgh’s School of Informatics, who led the study, said: “Most people can tell the time and use calendars from an early age. Our findings highlight a significant gap in the ability of AI to carry out what are quite basic skills for people. These shortfalls must be addressed if AI systems are to be successfully integrated into time-sensitive, real-world applications, such as scheduling, automation and assistive technologies.”

    Aryo Gema, also of the School of Informatics, said: “AI research today often emphasises complex reasoning tasks, but ironically, many systems still struggle when it comes to simpler, everyday tasks. Our findings suggest it’s high time we addressed these fundamental gaps. Otherwise, integrating AI into real-world, time-sensitive applications might remain stuck at the eleventh hour.”

  • Foundation AI model predicts postoperative risks from clinical notes

    Foundation AI model predicts postoperative risks from clinical notes

    Millions of Americans undergo surgery each year. After surgery, preventing complications like pneumonia, blood clots and infections can be the difference between a successful recovery and a prolonged, painful hospital stay — or worse. More than 10% of surgical patients experience such complications, which can lead to longer stays in the intensive care unit (ICU), higher mortality rates and increased health care costs. Early identification of at-risk patients is crucial, but predicting these risks accurately remains a challenge.

    New advancements in artificial intelligence (AI), particularly large language models (LLMs), now offer a promising solution. A recent study led by Chenyang Lu, the Fullgraf Professor in computer science & engineering in the McKelvey School of Engineering and director of the AI for Health Institute (AIHealth) at Washington University in St. Louis, explores the potential of LLMs to predict postoperative complications by analyzing preoperative assessments and clinical notes. The work, published online Feb. 11 in npj Digital Medicine, shows that specialized LLMs can significantly outperform traditional machine learning methods in forecasting postoperative risks.

    “Surgery carries significant risks and costs, yet clinical notes hold a wealth of valuable insights from the surgical team,” Lu said. “Our large language model, tailored specifically for surgical notes, enables early and accurate prediction of postoperative complications. By identifying risks proactively, clinicians can intervene sooner, improving patient safety and outcomes.”

    Traditional risk prediction models have primarily relied on structured data, such as lab test results, patient demographics, and surgical details like procedure duration or the surgeon’s experience. While this information is undoubtedly valuable, it often lacks the nuance of a patient’s unique clinical narrative, which is captured in the detailed text of clinical notes. These notes contain personalized accounts of the patient’s medical history, current condition, and other factors that influence the likelihood of complications.

    Lu and co-first authors Charles Alba and Bing Xue, both graduate students working with Lu at the time the study was conducted, employed specialized LLMs trained on publicly available medical literature and electronic health records. They then fine-tuned the pretrained model on surgical notes to make better predictions about surgical outcomes. The resulting method — the first of its kind to process surgical notes and use them to make predictions about postoperative outcomes — can go beyond structured data to recognize patterns in the patient’s condition that might otherwise be overlooked.

    Based on nearly 85,000 surgical notes and associated patient outcomes from an academic medical center in the Midwest collected between 2018 and 2021, the team reported that their model performed far better than traditional methods in predicting complications. For every 100 patients who experienced a postoperative complication, the team’s new model correctly predicted 39 more patients who had complications than traditional natural language processing models.

    Beyond the number of patients who could potentially have surgical complications caught early and mitigated, the study also showcases the power of foundation AI models, which are designed to multitask and can be applied to a wide range of problems.

    “Foundation models can be diversified, so they’re generally more useful than specialized models. In this case, where lots of complications are possible, the model needs to be versatile enough to predict many different outcomes,” said Alba, who is also a graduate student in WashU’s Division of Computational & Data Sciences. “We fine-tuned our model for multiple tasks at same time and found that it predicts complications more accurately than models trained specifically to detect individual complications. This makes sense because complications are often correlated, so a unified foundational model benefits from shared knowledge about different outcomes and doesn’t have to be painstakingly tuned for each one.”

    “This versatile model has the potential to be deployed across various clinical settings to predict a wide range of complications,” said Joanna Abraham, associate professor of anesthesiology at WashU Medicine and a member of the Institute for Informatics (I2) at WashU Medicine. “By identifying risks early, it could become an invaluable tool for clinicians, enabling them to take proactive measures and tailor interventions to improve patient outcomes.”

    This study is supported by the Agency for Healthcare Research and Quality within the U.S. Department of Health and Human Services (R01 HS029324-02).

  • A new model accurately predicts the movement of elite athletes to catch the ball in parabolic flight

    A new model accurately predicts the movement of elite athletes to catch the ball in parabolic flight

    How does a tennis player like Carlos Alcaraz decide where to run to return Novak Djokovic’s ball by just looking at the ball’s initial position? These behaviours, so common in elite athletes, are difficult to explain with current computational models, which assume that the players must continuously follow the ball with their eyes. Now, researchers of the University of Barcelona have developed a model that, by combining optical variables with environmental factors such as gravity, accurately predicts how a person will move to catch a moving object just from an initial glance. These results, published in the journal Royal Society Open Science, could have potential applications in fields such as robotics, sports training or even space exploration.

    The paper addresses the outfielder problem, which refers to the baseball player who stands in the outfield to catch the ball after it is hit. It is a classic challenge in physics and the neuroscience of movement, used to explore how humans and animals predict movements in a dynamic environment and how automated systems can be designed to mimic them.

    Joan López-Moliner, professor at the UB’s Faculty of Psychology and member of the Institute of Neurosciences (UBneuro), has led the research and affirms that “faced with this problem, current models are based on guiding locomotion by continuously looking at the ball, while normally the elite athlete can run towards the ball without looking at it.” “Moreover, — he adds — these models do not allow predictions of where the ball will go regarding the observer.” The initial study was part of the doctoral thesis conducted by Borja Aguado, co-author and former member of the group, who, after a stay in Darmstadt (Germany), is now a researcher at the University of Vic.

    The model integrates prior knowledge of the ball’s gravity and physical size into the visual information received in real time. “The model provides live signals that indicate the predicted position of the ball’s fall and the time remaining until it arrives, considering different gravity conditions. This makes it possible to predict precisely how a player will move to catch it, from the very beginning of the flight,” describes López-Moliner, who also coordinates the Vision and Control of Action research group.

    Despite the importance of gravity in anticipating trajectories, this is the first time this factor has been included in such a model. “This omission has overlooked the substantial influence that gravity exerts on the trajectory, which reflects a gap in the way existing models take into account environmental constants,” says the UB professor.

    Moreover, the previous models cannot explain why humans perceive whether a ball is within reach or not to decide whether to start running. “Our model does account for this, as it indicates where the object will go regarding the player,” says the researcher.

    Experiments with virtual reality

    To validate the model, the researchers conducted experiments in an immersive virtual reality environment, in which each participant — wearing goggles and holding a virtual reality device in their hand — had to move to the position where they thought a virtual ball would land. The controlled environment allowed a variety of gravity and ball size conditions to be simulated, showing that the empirical trajectories, movement patterns and temporal responses matched the model’s predictions. “Our model accurately predicts the trajectories observed in the different conditions by the participants. The results underline the importance of integrating environmental constants, such as gravity, to better understand how humans interact with the world around us,” says the researcher.

    Virtual training for athletes and astronauts

    The new model could be a basis for the future development of various practical applications, for example in sports training. “On the one hand, since the model includes several components — such as visual information or gravity — it could be applied in training or virtual simulation platforms. In this way, the degree of sensitivity of a person — such as an elite athlete — to the different components could be seen, or they could be trained to process and use the visually relevant information to optimize performance,” says López-Moliner.

    Moreover, the fact that the model can consider various types of gravity could also have applications in the aerospace sector. “The model can be applied in environments with different gravities and can potentially predict the performance with which a person — for example, the astronaut on the space station — would interact with moving objects,” says the professor.

    Research with artificial neural networks

    The researchers are already working on the next phase: implementing the model in artificial neural networks, computational systems that mimic the functioning of neurons in the human brain. Their goal is to compare the performance of humans and artificial networks. “This would allow us to have a clearer idea of how the computations are implemented at the neural level, as we now have a model at computational scale, but not the neural implementation in an artificial neural network. This knowledge could have clear applications in the field of robotics,” concludes López-Moliner.

  • New AI model measures how fast the brain ages

    New AI model measures how fast the brain ages

    A new artificial intelligence model measures how fast a patient’s brain is aging and could be a powerful new tool for understanding, preventing and treating cognitive decline and dementia, according to USC researchers.

    The first-of-its-kind tool can non-invasively track the pace of brain changes by analyzing magnetic resonance imaging (MRI) scans. Faster brain aging closely correlates with a higher risk of cognitive impairment, said Andrei Irimia, associate professor of gerontology, biomedical engineering, quantitative & computational biology and neuroscience at the USC Leonard Davis School of Gerontology and visiting associate professor of psychological medicine at King’s College London.

    “This is a novel measurement that could change the way we track brain health both in the research lab and in the clinic,” he said. “Knowing how fast one’s brain is aging can be powerful.”

    Irimia is the senior author of the study that describes the new model and its predictive power; the study was published February 24, 2025 in Proceedings of the National Academy of Sciences.

    Biological brain age versus chronological age

    Biological age is distinct from an individual’s chronological age, Irimia said. Two people who are the same age based on their birthdate can have very different biological ages due to how well their body is functioning and how “old” the body’s tissues appear to be at a cellular level.

    Some common measures of biological age use blood samples to measure epigenetic aging and DNA methylation, which influences the roles of genes in the cell. However, measuring biological age from blood samples is a poor strategy for measuring the brain’s age, Irimia explained. The barrier between the brain and the bloodstream prevents blood cells from crossing into the brain, such that a blood sample from one’s arm does not directly reflect methylation and other aging-related processes in the brain. Conversely, taking a sample directly from a patient’s brain is a much more invasive procedure, making it unfeasible to measure DNA methylation and other aspects of brain aging directly from living human brain cells.

    Previous research by Irimia and colleagues highlighted the potential of MRI scans to non-invasively measure the biological age of the brain. The earlier model used AI analysis to compare a patient’s brain anatomy to data compiled from the MRI scans of thousands of people of various ages and cognitive health outcomes.

    However, the cross-sectional nature of analyzing one MRI scan to estimate brain age had major limitations, he said. While the previous model could, for instance, tell if a patient’s brain was ten years “older” than their calendar age, it couldn’t provide info on whether that additional aging occurred earlier or later in their life, nor could it indicate whether brain aging was speeding up.

    A more accurate picture of brain aging

    A newly developed three-dimensional convolutional neural network (3D-CNN) offers a more precise way to measure how the brain ages over time. Created in collaboration with Paul Bogdan, associate professor of electrical and computer engineering and holder of the Jack Munushian Early Career Chair at the USC Viterbi School of Engineering, the model was trained and validated on more than 3,000 MRI scans of cognitively normal adults.

    Unlike traditional cross-sectional approaches, which estimate brain age from one scan at a single time point, this longitudinal method compares baseline and follow-up MRI scans from the same individual. As a result, it more accurately pinpoints neuroanatomic changes tied to accelerated or decelerated aging. The 3D-CNN also generates interpretable “saliency maps,” which indicate the specific brain regions that are most important for determining the pace of aging, Bogdan said.

    When applied to a group of 104 cognitively healthy adults and 140 Alzheimer’s disease patients, the new model’s calculations of brain aging speed closely correlated with changes in cognitive function tests given at both time points.

    “The alignment of these measures with cognitive test results indicates that the framework may serve as an early biomarker of neurocognitive decline,” Bogdan said. “Moreover, it demonstrates its applicability in both cognitively normal individuals and those with cognitive impairment.”

    He added that the model has the potential to better characterize both healthy aging and disease trajectories, and its predictive power could one day be applied to assessing which treatments would be more effective based on individual characteristics.

    “Rates of brain aging are correlated significantly with changes in cognitive function,” Irimia said. “So, if you have a high rate of brain aging, you’re more likely to have a high rate of degradation in cognitive function, including memory, executive speed, executive function, and processing speed. It’s not only an anatomic measure; the changes we see in the anatomy are associated with changes we see in the cognition of these individuals.”

    Looking ahead

    In the study, Irimia and coauthors also note how the new model was able to distinguish different rates of aging across various regions of the brain. Delving into these differences -including how they vary based on genetics, environment, and lifestyle factors — could provide insight into how different pathologies develop in the brain, Irimia said.

    The study also demonstrated that the pace of brain aging in certain regions differed between the sexes, which might shed light onto why men and women face different risks for neurodegenerative disorders, including Alzheimer’s, he added.

    Irimia said he is also excited about the potential for the new model to identify people with faster-than-normal brain aging before they show any symptoms of cognitive impairment. While new drugs targeting Alzheimer’s have been introduced, their efficacy has been less than researchers and doctors have hoped for, potentially because patients might not be starting the drug until there is already a great deal of Alzheimer’s pathology present in the brain, he explained.

    “One thing that my lab is very interested in is estimating risk for Alzheimer’s; we’d like to one day be able to say, ‘Right now, it looks like this person has a 30% risk for Alzheimer’s.’ We’re not there yet, but we’re working on it,” Irimia said. “I think this kind of measure will be very helpful to produce variables that are prognostic and can help to forecast Alzheimer’s risk. That would be really powerful, especially as we start developing potential drugs for prevention.”

  • Scientists optimize biohybrid ray development with machine learning

    Scientists optimize biohybrid ray development with machine learning

    The Harvard John A. Paulson School of Engineering and Applied Sciences (SEAS) and NTT Research, Inc., a division of NTT, announced the publication of research showing an application of machine-learning directed optimization (ML-DO) that efficiently searches for high-performance design configurations in the context of biohybrid robots. Applying a machine learning approach, the researchers created mini biohybrid rays made of cardiomyocytes (heart muscle cells) and rubber with a wingspan of about 10 mm that are approximately two times more efficient at swimming than those recently developed under a conventional biomimetic approach.

    A team led by Harvard SEAS Postdoctoral Fellow John Zimmerman and including NTT Research Medical and Health Informatics Scientist Ryoma Ishii, Harvard SEAS Tarr Family Professor of Bioengineering and Applied Physics Kevin Kit Parker, and members of the Harvard SEAS Disease Biophysics Group led by Parker demonstrated this research in a new paper published in Science Robotics titled, “Bioinspired Design of a Tissue Engineered Ray with Machine Learning.”

    “This research seeks to answer a fundamental question in the development of biohybrid robots, in this case the marine ray: How do we select fin geometries to operate under novel working environments while preserving natural scaling laws in terms of swimming speed and efficiency,” said Ishii, who also works as a visiting scientist for Harvard University. “Our research indicates the application of ML-DO, inspired by protein engineering, offers a more efficient and less computationally intensive path forward in automating the creation of muscular structure-function relationships.”

    Limitations of the Biomimetic Approach

    In biomimetic design, the conventional approach to biohybrids, engineers form functional devices by recreating existing biological structures. That approach, however, has limits. For biohybrid lifeforms that resemble batoid fishes (skates and rays), for example, there is a wide range of natural aspect ratios and fin morphologies. Which ones do you mimic? Also, biomimetics may neglect the natural biomechanical and hydrodynamic forces that govern how fast an organism can swim based on its size and body kinematics, leading to inefficient muscle mass and limited swimming speeds.

    In that light, the motivating question in this study became: How to select fin geometries that operate under novel working environments while preserving natural scaling laws in terms of swimming speed and efficiency?

    The Design Breakthroughs of Machine Learning

    The multi-disciplinary and iterative nature of the problem required computationally intensive modeling, but the team believed that directed optimization by machine learning (ML-DO) would enable an efficient search for fin designs that maximized their relative swimming speeds. They based their hypothesis in part on a trial function that demonstrated an approximately 40 percent improvement of ML-DO over other leading methods in recognizing known high-rank sequences. Testing the assumption involved three steps: 1) developing an algorithm for expressing a multitude of different fin geometries; 2) describing a generalized ML-DO approach for searching within a large discontinuous configuration space; and 3) using this methodology to identify biohybrid fin geometries for high-performance swimming with smooth and orderly flow.

    The ML-DO-driven results included a quantitative exploration of fin structure-function relationships and reconstruction of general trends in open-sea batoid morphology, as well as a winning design: Fins with large aspect ratios and fine tapered tips, which preserved their utility across multiple length-scales of swimming. On that basis, the team built biohybrid mini-rays out of engineered cardiac muscle tissue, which were capable of self-propelled swimming at the millimeter length scale and demonstrated improved swimming efficiencies approximately two times greater than observed in previous biomimetic designs.

    Looking Ahead

    While promising, researchers note that additional work is needed to completely match natural scaling laws. While the devices presented in this study demonstrated greater efficiency than other recent biomimetic designs, they were still slightly less efficient on average than naturally occurring marine lifeforms.

    In the future, researchers expect to continue the development of biohybrid robotics for use cases including remote sensors, probes for dangerous working environments and as therapeutic delivery vehicles. Researchers believe that the ML-DO-informed approach better mimics the selective pressures of evolution, enabling them to better understand how biological tissues are shaped — both in a healthy physiology as well as the maladaptive pathophysiology of disease. Additionally, this research advances scientific understanding of 3D organ biofabrication, such as a biohybrid heart.

    “The Harvard Biophysics Group entered into a joint research agreement with NTT Research two years ago aiming to fundamentally interrogate our understanding of cardiac physiology, with an eye towards advancing the development of biohybrid devices, including biohybrid robotics and biohybrid human heart,” Parker said. “This paper indicates the positive progress our shared research has achieved so far, and I am excited to see what the future of our collaboration has in store.”

    In 2022, NTT Research and Harvard SEAS announced a three-year joint research agreement to engineer a model of the human heart, study fundamental laws of muscular pumps and apply joint findings to the development of a cardiovascular bio digital twin model.