2025 Poster Award Recipients
Best Overall Poster
#50: A Generalized Framework for Alchemical Machine-Learned Coarse-Grained Interaction Models in Polymer-Grafted Nanoparticle Self-Assembly
Melody Zhang, PhD Candidate, College of Engineering
Outstanding Research Innovation
#15: A Workflow for Predicting High Temporal and Spatial Resolution Rainfall Data Using Machine Learning Tools
Marwah Al Ismail, PhD Candidate, College of Literature, Science, and the Arts
High-Impact Application
#41: Leveraging Modern Machine Learning to Improve Early Warning Systems and Reduce Chronic Absenteeism in Early Childhood
Tiffany Wu, PhD Student, College of Literature, Science, and the Arts
Excellence in Communication
#44: Cultural Aspects of General Artificial Intelligence
Maria Fields, PhD Student, College of Engineering
Best Reproducibility
#33: Scalable Geometric Defect Detection in Manufacturing Using Synthetic 3D Point Cloud Data
Mei (Alice) Ruo-Syuan, Graduate Student Research Assistant, College of Engineering
Audience Favorite
#7: Overcoming Moderation Barriers: Applying LLMs to Sensitive Narrative Datasets
Anay Halwasiya, Graduate Student, College of Literature, Science, and the Arts
Student Posters
Joshua Ashkinaze, PhD Candidate, School of Information
Ceren Budak, Associate Professor of Information, School of Information
Stephanie Preston, Professor of Psychology, College of Literature, Science, and the Arts
Personal experiences powerfully shape attitudes and policy support. But individuals remain constrained by their limited lived experiences. This creates barriers to empathetic understanding and support for policies benefiting others—particularly regarding distant risks, unfamiliar social experiences, and abstract scientific benefits. We propose the As-If Machine (AIM), a novel multi-agent retrieval-augmented generation system designed to bridge psychological distance through personalized narrative simulations. Through a MIDAS PODS grant, we are scaling up evaluations this semester.
AIM leverages narrative transportation theory, mental simulation, and imagined contact hypothesis to help users vividly imagine alternative life scenarios. The system employs four specialized AI agents: an Initializer Agent that generates personalized scenario onsets, a Predictor Agent that forecasts scenario consequences, a Writer Agent that creates cohesive narrative arcs, and an Autocompleter Agent that facilitates interactive co-writing. See the diagram above.
The system is designed as a flexible intervention engine, allowing researchers to create custom experiments through simple YAML configuration files without programming expertise. This significantly reduces barriers for designing AI interventions across disciplines.
We will test AIM’s effectiveness across three critical application areas: (1) motivating proactive engagement with long-term risks like democratic backsliding and pandemic preparedness, (2) building empathy across social divides including healthcare access and economic mobility barriers, and (3) rebuilding public trust in science by making research benefits tangible and illustrating disinvestment costs.
Preliminary randomized controlled trials show promising results. In our democratic backsliding pilot (n=20), AIM reduced psychological distance and increased policy support. Our chronic pain study (n=23) demonstrated increased empathy and support for accommodative policies. This interdisciplinary project bridges psychological theories of persuasion with AI capabilities, offering a scalable solution to fundamental challenges in empathy, risk perception, and science communication.
Eleanor Lin, PhD Student, College of Engineering
David Jurgens, Associate Professor of Information, School of Information
State-of-the-art large language models (LLMs) code-switch (i.e., mix languages), but how and why is still poorly understood–especially cognitive differences from humans. We address this gap by introducing a cognitive framework for characterizing code-switching in LLM reasoning. We start from reasoning examples sourced from diverse models, languages, domains, and tasks. Fusing top-down theory-driven and bottom-up data-driven approaches, we then develop a taxonomy of code-switched reasoning behaviors (see Figure 1). Our taxonomy reveals that LLM and human code-switching behaviors in LLMs partially align. Additionally, more naturalistic, human-like code-switching may boost model performance, particularly for languages from the long tail of training data distributions. Our work serves as a first, necessary step toward uncovering parallels between LLM and human code-switching. With further testing, LLMs could potentially serve as proxies for human multilingual cognition. Additionally, our approach can develop future reasoning taxonomies informed by cognitive science and education.
John Mobley IV, Graduate Student, Gerald R. Ford School of Public Policy
Jeanne Sanders, Research Lab Specialist, Gerald R. Ford School of Public Policy
Isabel Miller, Research Investigator, Gerald R. Ford School of Public Policy
Paul Jensen, Associate Professor of Biomedical Engineering, College of Engineering
Karin Jensen, Assistant Professor of Biomedical Engineering, College of Engineering
Large-language models (LLMs) such as ChatGPT can generate text that appears human-generated, yet important differences remain between LLM- and human-generated content. Therefore, we must explore the differences between LLM-generated and human-generated content. Many LLMs are trained on a corpus of data that contain underlying biases and inequities. Because of these underlying issues, exploring differences in LLM- and human-generated data is especially important in topics with high potential for impacting equity and wellness, such as mental health and well-being (MHWB). Topics of MHWB are particularly important to explore in disciplines such as engineering that have cultures of high stress and diminished MHWB.
We systematically compared human interview transcripts with LLM-generated interview transcripts. Our dataset included interview transcripts with engineering faculty and students (n = 38) about topics relating to well-being and mental health in engineering.
We found most content to be similar between human- and LLM-generated interviews, with the highest similarity on questions that were definitional (e.g., describing wellness) and explored common experience (e.g., recognizing stress). Significant differences emerged for systems-level questions (e.g., proposed changes to reduce stress in academic environments). Introducing demographic variation in the prompt increased variation in the LLM-generated output; however, this also introduced clear examples of stereotyping that were not present in interviews with humans.
These results indicate that LLMs such as ChatGPT may be useful when brainstorming ideas, such as during interview protocol creation. However, these results also urge caution when using generative AI to create content, since they indicate a likelihood of perpetuating stereotypes, such as people in positions of power being unaware of others’ experiences. This caution is particularly encouraged with creating content relating to critical conversations to advance equity, such as topics on MHWB in engineering.
Suvrorup Mukherjee, Graduate Student, College of Literature, Science, and the Arts
Abeyankar Giridharan, Graduate Student, College of Literature, Science, and the Arts
Chang Li, Graduate Student Research Assistant, College of Engineering
Xinhe Wu, Graduate Student, College of Literature, Science, and the Arts
Modeling extreme volatility and sentiment-driven fluctuations inherent in cryptocurrency markets often falls short, as traditional financial models fail to capture these nuances. Existing research in capturing volatility in financial assets has shown promising results using GARCH variants and Bretó’s state-space framework. We propose an enhanced Partially Observed Markov Process model that incorporates market sentiment to better forecast Bitcoin’s price movements. Specifically, we extend Bretó’s framework in two significant ways. Firstly, we integrate the Fear and Greed Index (FGI) as a direct input to account for shifts in trading psychology. Secondly, we replace the standard Gaussian noise assumption with a Student’s t-distribution to more accurately model the frequent price spikes and heavy-tailed returns characteristic of cryptocurrencies. This sentiment-augmented model was trained and evaluated using daily Bitcoin prices and FGI data from January 2020 to April 2025. The model demonstrated superior performance, significantly outperforming benchmark models, including GARCH, Heston’s stochastic volatility model, and the original Bretó model, in both log-likelihood and filter stability. Bimodal convergence in likelihood and estimated parameter trajectories encapsulates tradeoffs between the shock scale and persistence and speaks to the model’s elasticity and sophistication. A key finding is the asymmetric effect of investor sentiment, with fear exerting a stronger influence on volatility than greed. A close comparability between observed and simulated Bitcoin returns using estimated parameters supports the robustness of this framework. The successful integration of sentiment data offers a significant advancement in financial modeling of digital assets, with practical implications for risk management, derivative pricing, and the development of more sophisticated trading strategies in the cryptocurrency space.
Tianming Liu, Ph.D. Candidate, College of Engineering
Large language models (LLMs) offer significant potential by serving as human proxies to advance travel demand modeling, but their behavioral misalignment with human travelers remains a critical obstacle. Furthermore, existing alignment methods are often impractical or inefficient when applied to the sparse datasets typically available for travel choices, limiting the adoption of these powerful new tools. We introduce a novel framework to align LLMs with travel choice behavior. Our method first infers a set of traveler personas from empirical data, then estimates a persona loading function that uses learned embeddings to select the appropriate persona for an individual based on their socio-demographics. Validated on the Swissmetro mode choice dataset, our approach significantly outperforms established benchmarks in predicting both aggregate and individual choice outcomes. Our research offers a more adaptable, interpretable, and resource-efficient pathway to robust LLM-based travel behavior simulation, paving the way to integrate LLMs into transportation modeling practice in the future.
Ali Arslan, Graduate Student Research Assistant, University of Michigan – Flint
Passwords remain the most widely used form of authentication, valued for practicality but vulnerable to leaks and advanced guessing attacks. Strengthening password security is an urgent challenge for practitioners and researchers. In this work-in-progress study, we present a reinforcement learning–based framework that generates realistic, diverse, non repetitive passwords, enabling stronger strength metrics and safer synthetic datasets for research.
Our contributions are threefold. First, we employ a decoder-only Transformer to learn the semantic and structural patterns of human-created passwords, building on prior findings that such patterns are predictable. Second, we introduce temperature sampling to control output diversity. The temperature parameter (T) alters the probability distribution: higher values flatten it to increase randomness, while lower values sharpen it to emphasize more likely choices. Third, we extend this method with a Deep Q-Network (DQN) agent that dynamically adjusts T when redundancy exceeds threshold values, ensuring diversity while maintaining realism. This adaptive temperature control is an original feature absent from existing password generators.
The significance lies in its dual impact. For end-users, it enables more accurate real-time password strength evaluation than current tools such as zxcvbn, encouraging stronger choices. For researchers, it provides synthetic datasets that reflect authentic password distributions without relying on leaked data, reducing misuse risks and broadening testing resources.
Our model is trained on the Have I Been Pwned (HIBP) dataset of 15 billion real passwords, used here for the first time in this context, and benchmarked against PassGPT and PassGAN. Preliminary results indicate improvements in realism, diversity, and guessability. Ongoing work focuses on hyperparameter tuning and algorithmic refinements to enhance efficiency and scalability.
This framework represents a promising step toward next-generation password security research and evaluation.
Anay Halwasiya, Graduate Student, College of Literature, Science, and the Arts
Large language models (LLMs) such as GPT offer new opportunities for analyzing unstructured narratives at scale, but their deployment to analyze information about sensitive topics presents unique challenges that demand careful attention to ethical AI practices and the design of human-centered applications. In our study of 40,000 narratives about adolescent risk behaviors, including bullying, self-harm, and substance use, we encountered two recurring problems: first, that the narratives themselves often contained sensitive details which triggered terms of service (ToS) violations even for otherwise neutral tasks (e.g., identifying “who is this about”), and second, that prompts explicitly requesting classification of sensitive topics were more likely to result in refusals or errors. We experimented with multiple prompting and batching approaches, finding that large, context-rich prompts substantially reduced model refusals, lowering error rates from a high of 13% to under 0.05%, compared to shorter, isolated queries. We further observed that few-shot prompting, constructed using a manually coded subset of the dataset, was highly effective in improving reliability and accuracy for classifications where zero-shot prompting had low accuracy. Patterns of ToS violations also emerged, with certain survey-style items, particularly questions q14_15 through q14_23, which cover cutting and nonsuicidal self-injury, suicidality, mass harm and weapons, adult and gang involvement, identity-based violence, and family- or school-affiliated threatening behaviors, disproportionately triggering refusals, suggesting sensitivity linked to specific content categories. Despite these challenges, GPT consistently produced structured, high-quality outputs once the prompting strategy was refined. Our findings highlight both the constraints imposed by moderation systems and practical methods to overcome them, offering guidance for researchers seeking to deploy LLMs on sensitive or large-scale narrative datasets.
Minseo Park, PhD Candidate, College of Engineering
Our research targets DeFi’s structural dependence on over-collateralized loans by introducing an on-chain-only credit rating that infers a wallet’s creditworthiness exclusively from its blockchain transaction history. Current U.S. legislative momentum in the cryptocurrency and stablecoin realm – most notably FIT21 and the GENIUS Act – is laying the regulatory groundwork for on-chain financial services. As this foundation matures, lenders and protocols need native, privacy-preserving credit signals that do not rely on off-chain identity. Our approach structures enabling features from wallet flows and positions (e.g. net-inflow regularity, recurrence, volatility; leverage and health-factor behavior; liquidation frequency and time-under-debt; collateral-buffer dynamics; protocol diversity and tenure) and trains a survival-boosted ensemble to estimate horizon-specific default probabilities using protocol-defined liquidation/bad-debt events as labels. Credit scores are calibrated to numeric and letter grades and accompanied by SHAP-based reasoning codes that render ratings auditable while keeping all inputs on-chain. An end-to-end pipeline continuously ingests transactions, updates features in rolling windows and serves ratings via APIs. Preliminary back tests across multiple market regimes show robust rank-order risk separation relative to heuristic baselines with stable out-of-sample behavior. By converting transparent, tamper-evident wallet histories into standardized, interpretable ratings—with no off-chain data or PII – this work provides a plug-in signal for risk-based limits and pricing and, importantly, demonstrates that purely on-chain data can unlock prudent, unsecured lending, thus improving capital efficiency and expanding access for users without traditional credit files.
Shaoying Zheng, Graduate Student, School of Information
Accessing specific building footprints and land parcels from continent-scale datasets presents a significant computational challenge. We have developed an efficient, cloud-native workflow to query and extract these geospatial features for any user-defined area of interest.
Our method leverages the H3 spatial indexing system to intelligently query data stored in a partitioned database on Azure. Crucially, we utilize the GeoParquet file format; this cloud-optimized format includes spatial metadata that allows our query engine to bypass irrelevant data chunks, which is fundamental to the workflow’s efficiency. An initial bounding box identifies the relevant data partitions, which are then queried in memory using DuckDB’s spatial functions. This on-the-fly approach minimizes I/O and processing overhead, providing a scalable and rapid solution for researchers to extract precise geospatial data from massive repositories.
Marwah Al Ismail, PhD Candidate, College of Literature, Science, and the Arts
Marin Clark, Professor of Earth and Environmental Sciences, College of Literature, Science, and the Arts
Ries Plescher, Masters Student, University of Oregon
Madeline Hille, BGC Engineering Inc.
Deepak Chamlagain, Associate Professor of Geology, Tribhuvan University
Predicting landslides triggered by extreme rainfall events (EREs) is hampered by the scarcity of accurate, high-resolution rainfall data in remote, steep mountain areas. Satellite-driven rainfall data offers a possible solution with high spatial and temporal resolution, but their accuracy is often compromised in high-relief terrain, where most landslides happen. To overcome this challenge, we developed a workflow that uses an Ensemble Bagged Trees model to accurately estimate rainfall at locations without gauge stations, extending a previously validated scaling method for a pilot project in central Nepal. The scaling technique uses daily gauge measurements to adjust satellite data (i.e., NASA’s GPM IMERGE precipitation measurements) at half-hour intervals, based on a daily scaling factor. We consider 19 predictors related to rainfall amounts, satellite parameters, location, and time during the monsoon season from 49 gauge stations over 6 years. To validate this method, we split our stations into 75% training data and 25% testing data. Our results indicate that 8 predictors—mainly related to the daily unscaled satellite rainfall, location, time, and day within the monsoon—produce the best performance. This method yields similar annual statistics for the storms and extreme events generated by the predicted scaled rainfall data compared to those generated by the gauge-measured data. In addition, this model enables us to predict the timing and duration of storms with an average precision and recall of 80% for both. However, because the model struggles to predict the highest scaling factors, there is a systematic underprediction of the highest-intensity storms (i.e., EREs). Overall, this innovative workflow is a promising step toward generating reliable rainfall data with high spatial coverage, enabling the forecasting of landslide locations and timings for early warning systems.
Sophia Cheng, Graduate Student, School of Information
Single-cell RNA sequencing (scRNA-seq) has played a pivotal role in advancing the understanding of biology by enabling researchers to measure gene expression at the resolution of individual cells. This project is motivated by the causal ambiguity of identifying thresholds for quality control (QC) metrics in
the scRNA-seq pre-processing workflow. Current approaches aim to minimize noise from ambient RNA contamination, at the risk of eliminating meaningful biological signals. The goal of this study is to perform a comparative scRNA-seq analysis of cells classified as biological signals (“real”) versus those labeled as technical artifacts (“noise”), with the aim of evaluating whether current QC processes systematically exclude potentially informative cellular states.
We1 focus our analysis on human breast tissue for its heterogeneous cellular composition and biological complexity, guided by the reference article “A single-cell RNA expression atlas of normal, preoneoplastic and tumorigenic states in the human breast”. We reverse engineered the supplementary data to classify cells as “real” or “noise” and validated this framework by reproducing the authors’ t-SNE analysis on the “real” cells. Applying the same approach to the “noise” cells, we cross-checked patterns with UMAP to address potential loss of global structure. Finally, we examined pathway-associated gene expression to assess functional processes underlying observed similarities.
Our main finding suggests potential biological signals in the “noise” cells that may represent damaged, dormant, and/or multifunctional cells. In the context of cancer research, these potential signals could offer valuable insights. For example, quantifying the proportion of “noise” cells could serve as a novel diagnostic indicator.
1 “We” and “our” are used for narrative flow, and not an indication of shared work.
Lingxiao Li, PhD Candidate, School of Information
Text-conditioned molecular generation seeks to translate natural-language descriptions into novel chemical structures, opening an intuitive channel for scientists to specify functional groups, and physicochemical constraints without handcrafted rules. Diffusion-based models—particulary latent diffusion models (LDMs)—have recently shown promise by performing stochastic search in a continuous latent space that compactly captures molecular semantics. Yet current LDM approaches suffer from weak cross-modal alignment, coarse one-shot conditioning, and evaluation metrics that overlook semantic fidelity. We propose Chain-of-Generation (CoG), a multi-stage latent diffusion framework that decomposes an input prompt into a sequence of semantic sub-prompts. CoG progressively incorporates these segments to define intermediate goals, guiding the denoising trajectory toward molecules that satisfy increasingly rich linguistic constraints. A post-alignment learning phase strengthens the correspondence between textual and molecular latent spaces, while an enhanced denoising network exploits this alignment during generation. To assess fidelity and interpretability, we introduce deep qualitative trajectory analysis, tracing how individual prompt components manifest as structural motifs across diffusion steps. Extensive experiments on benchmark and real-world tasks demonstrate that CoG yields higher semantic alignment, diversity, and controllability than one-shot baselines, producing molecules that more faithfully reflect complex, compositional prompts while offering transparent insight into the generation process.
Saipranav Avula, Graduate Student, College of Literature, Science, and the Arts
The Great Lakes, the world’s largest freshwater system, provides drinking water to over 38 million people and is a vital resource for irrigation, shipping, hydroelectric power, and recreation. Given its ecological and economic importance, designing observing networks is critical for monitoring dynamics in the Great Lakes. However, due to limited resources monitoring, it is crucial to efficiently use available observing platforms such as buoys and research vessels.
We present a machine learning framework that applies convolutional Gaussian neural processes (ConvGNPs) to the problem of Great Lakes observing system design. Specifically, we address the question: Where should the next generation of temperature sensors be placed to most efficiently reduce uncertainty in surface temperature variability?
Our approach fuses datasets—including remotely sensed surface temperature, bathymetry, land coverage, and historical ice concentrations—to train ConvGNPs which capture spatiotemporal covariance structures across variables. We systematically evaluated training strategies, including data chunking for efficiency and sampling restrictions to avoid land points. Specifically, incorporating ice concentration data as contextual inputs improved predictive performance and reduced training loss.
Using posterior means and variances from ConvGNP predictions, we implemented a greedy active learning algorithm to iteratively acquire locations with maximum uncertainty. We compared acquisition functions to identify optimal placements for new sensors and assessed their impact on predictive coverage. Results indicate that the existing observational network aligns with high-information regions, but targeted expansion would further strengthen coverage of surface temperature variability.
Xingran Chen, PhD Candidate, School of Public Health
Tyler McCormick, Professor of Statistics, University of Washington
Bhramar Mukherjee, Professor of Biostatistics, Yale University
Zhanke Wu, Associate Professor of Biostatistics, School of Public Health
Pre-trained machine learning (ML) predictions have been increasingly used to com plement incomplete data to enable downstream scientific inquiries, but their naive integration risks biased inferences. Recently, multiple methods have been developed to provide valid inference with ML imputations regardless of prediction quality and to enhance efficiency relative to complete-case analyses. However, existing approaches are often limited to missing outcomes under a missing-completely-at-random (MCAR) assumption, failing to handle general missingness patterns under the more realistic missing-at-random (MAR) assumption. This paper develops a novel method which delivers valid statistical inference framework for general Z-estimation problems using ML imputations under the MAR assumption and for general missingness patterns. The core technical idea is to stratify observations by distinct missingness patterns and construct an estimator by appropriately weighting and aggregating pattern-specific in formation through a masking-and-imputation procedure on the complete cases. We provide theoretical guarantees of asymptotic normality of the proposed estimator and efficiency dominance over weighted complete-case analyses. Practically, the method affords simple implementations by leveraging existing weighted complete-case analysis software. Extensive simulations are carried out to validate theoretical results. Real data examples are provided to further illustrate the practical utility of the proposed method. The paper concludes with a brief discussion on practical implications, limita tions, and potential future directions.
Ninad Kamath, Graduate Student Research Assistant, College of Engineering
Differentiating histopathologically similar disease states, such as chronic pancreatitis and incipient pancreatic ductal adenocarcinoma, represents a formidable diagnostic challenge in late-stage presentations. This clinical bottleneck contributes to elevated mortality and constrains therapeutic options. While spatially resolved multiplex immunoprofiling holds promise for discovering intricate characteristics, typical study cohorts rarely achieve the scale required for adequate classification power, limiting the discovery of robust biomarkers.
Our framework introduces a novel approach to extract meaningful information from cellular coordinates by quantifying the complex architectural features within each cell’s local neighborhood. The core innovation lies in re-conceptualizing the static pathology slide as a dynamic sequence. We posit that pathological transformation, particularly in solid tumors, is a radially dependent process that evolves outward from an origin.
To capture this progression, our pipeline partitions the tissue into a series of density based quantiles. Within each successive zone, we quantify the evolving patterns of inter-cellular spatial relationships with spatial correlation coefficients, generating a statistical snapshot of the tissue’s architecture within the specific quantile. By ordering these snapshots sequentially from the core to the periphery, we construct a “radial pseudo-trajectory.” This is a quantitative data series that maps the progressive structural remodeling of the tissue, providing a unique signature of the disease’s spatial evolution.
This radial pseudo-trajectory becomes the direct input for Hidden Markov Models, which show strong classification performance based on their radial signatures. These architectures are exceptionally suited to learn the long-range dependencies and subtle transitions embedded within the trajectory. Our pipeline, which will be made into a Python toolkit, offers a computationally efficient method to distill complex spatial biology into a strong diagnostic signal, effectively converting a static image into a dynamic narrative for discovering insights into different pathologies.
Seongju Jang, PhD Student, College of Engineering
Francis Baek, Assistant Professor of Civil and Environmental Engineering, Georgia Institute of Technology
SangHyun Lee, Professor of Civil and Environmental Engineering, College of Engineering
For a mobile construction robot to execute tasks on site, proper positioning at task-relevant locations is essential. Many studies have focused on autonomous navigation to predefined coordinates, with limited work on autonomously identifying the task-required destination and positioning accordingly. This limitation keeps human workers tied to robot positioning for task execution. Large Multimodal Model (LMM)-based robot navigation suggests the possibility of robots autonomously understanding tasks, identifying, and moving to the required locations. Several studies have integrated LMMs with Building Information Modeling (BIM) to leverage predefined building information for task-relevant positioning, but tasks involving elements absent from BIM (e.g., workers, equipment) remain challenging. Other studies leverage visual information with LMMs, allowing robots to align with objects not defined in BIM; however, without BIM data, long-range navigation in complex construction sites is still difficult. In this research, we propose a Multi-LMM-Agent framework that adaptively uses BIM and visual information according to the situation, enabling the robot to autonomously position itself at task-relevant locations in indoor construction environments. We employ three agents. The Manager Agent receives a task from the user and breaks tasks down into subtasks and distributes them to the other agents. The BIM Agent uses drawing information to perform long-range navigation. The Vision Agent identifies task targets not included in the BIM and carries out fine-grained positioning accordingly. To implement such a system on a real robot, we built a Docker-based backend that enables real-time feedback among the robot operating system, AI agents’ decision-making, and visual information acquisition. We implemented the framework on the quadruped robot, Unitree Go2, which successfully positioned itself at task-required locations in the real world with a high success rate. This study contributes to enabling construction robots to autonomously move to required locations and perform tasks on-site without continuous human involvement in low-level positioning control.
Zeynep Deniz Lal, PhD Candidate, College of Engineering
Brian Johnson, Lead Research Engineer, College of Engineering
Faezeh Shanehsazzadeh, Research Fellow, College of Engineering
Sanaz Habibi, Assistant Research Scientist, College of Engineering
James Ashton-Miller, Research Professor Emeritus, College of Engineering
Frederick Korley, Professor of Emergency Medicine, Medical School
Mark Burns, Professor Emeritus of Chemical Engineering, College of Engineering
Traumatic brain injury (TBI) is a major cause of morbidity and mortality, where diagnostic delays can limit opportunities for life-saving intervention, especially within the critical “golden hour” post-injury. Current imaging-based assessments are often impractical in urgent, resource-limited, or field settings, creating a need for rapid, portable diagnostic alternatives. Although FDA-approved blood-based biomarkers (GFAP, UCH-L1, NfL) are promising, timely and sensitive point-of-care detection remains a challenge.
We present a portable, credit card-sized microfluidic device using bead-based sandwich immunoassays in a variable-height channel to rapidly and quantitatively measure TBI biomarkers from just 20 μL of whole blood in under 10 minutes. The assay produces fluorescent signals proportional to biomarker levels, detected and analyzed via a high-resolution imaging system. To overcome the limitations of bulk analysis, we apply advanced machine learning techniques to analyze binding events at the single-bead level. Datasets including single and multiplexed spiked buffer samples are processed with image analysis algorithms to extract features like bead intensity, count, and spatial localization. Random Forest models
incorporating feature engineering and nested cross-validation were developed to ensure robust biomarker detection in complex biological matrices. Additionally, we are employing data augmentation strategies to maximize information from available datasets and improve generalizability, especially with limited, heterogeneous data.
Our integrated microfluidic and machine learning platform demonstrates high predictive accuracy for quantifying TBI biomarkers in complex samples, achieving coefficients of determination (R²) up to 0.9960 on test data. The system delivers outputs to support clinical distinction of TBI severity, with ongoing work to integrate automated pipelines and further validate on clinical samples.
Pairing cutting-edge microfluidics with ML-driven analytics, this platform holds promise for emergency, military, outpatient, and home care settings. Through innovative application of data science, machine learning, and portable diagnostics, our approach offers the potential to transform early TBI detection and management, reduce unnecessary imaging and costs, and improve patient outcomes on a broad scale.
John Sohn, PhD Candidate, College of Engineering
Falls are a devastating health issue, causing more than 46,000 deaths and 10.9 million injuries annually in the U.S. Recently, mobile assistive systems have gained attention as a way to proactively prevent falls by providing end users with just-in-time interventions such as sending real-time alerts to avoid hazards. For such systems to be successfully implemented, however, reliable assessment of perceived fall risk is essential, as users are more likely to trust and adopt a system that aligns with their own experiences of risk. Current fall risk assessment methods primarily focus on measuring gait or postural instability, highlighting a lack of methods for assessing perceived fall risk. To address this gap, we propose an unsupervised method for assessing perceived fall risk by directly analyzing body movement and physiological arousal using wearable inertial measurement units (IMUs) and electrodermal activity (EDA) sensors (Figure 1). Grounded in earlier findings that EDA serves as a reliable indicator of perceived risk, the core concept of our method is to refine abnormal physical movement scores (from IMU) with arousal scores (from EDA) to better align with individuals’ perceived experiences of risk. Physical instability is quantified through a graph-based approach that measures deviations across body axes, while anomalous arousal is captured using a convex optimization method that precisely extracts the arousal component from raw EDA signals. These scores are then integrated to generate a fall hazard index, which serves as a proxy for assessing perceived fall risk. For validation, IMU and EDA data were collected from 28 participants during daily walking. The results demonstrate that the proposed method can effectively assess perceived fall risk, achieving an unweighted average recall of 88.8%. Ultimately, this study provides a computational model that can be seamlessly integrated into mobile assistive systems to proactively prevent falls and enhance individuals’ health and safety.
Jingyi Qiu, PhD Student, School of Information
Hong Chen, Graduate Student, School of Information
Zongyi Li, Postdoctoral Associate, Massachusetts Institute of Technology
The rise of AI has sparked growing concerns over “hype” in machine learning papers, yet a reliable way to quantify rhetorical style, independent of substantive content, has remained a challenge. This paper introduces a novel framework to disentangle and measure the rhetorical style of a research paper from its substantive content. Instead of relying on keyword dictionaries or direct ratings, we introduce an agent-based framework in which multiple LLM rhetorical personas generate al ternative abstracts from the same substantive content, an LLM judge compares them through pairwise evaluations, and outcomes are aggregated using a Bradley–Terry preference model. This produces a continuous rhetorical score that locates each paper’s rhetorical style along a spectrum from conservative to visionary. Applying this instrument to 8,485 ICLR submissions randomly sampled from 2017–2025, we provide the first large-scale quantification of scientific rhetoric. Our findings show that rhetorical style is predictive of downstream attention: visionary framing significantly correlates with future citations and media coverage. These results demonstrate that how research is presented influences its visibility independent of technical merit. More broadly, our work introduces a robust and scalable methodol ogy for measuring subtle textual properties and enables new directions in AI for metascience.
Kapotaksha Das, PhD Student, University of Michigan – Dearborn
Mohamed Abouelenien, Associate Professor of Computer and Information Science, University of Michigan – Dearborn
Michael Cole, Clinical Associate Professor of Emergency Medicine, Medical School
James Cooke, Program Director of Learning Health Science, Medical School
Vitaliy Popov, Assistant Professor of Learning Health Sciences, Medical School
Effective care for critically ill patients requires frequent, high-fidelity training with meaningful feedback on essential clinical skills. However, current medical simulation training requires constant human observation for assessment and feedback, limiting its consistency, efficiency, and scalability. To automate this assessment, we introduce an open-source large language model (LLM) pipeline embedded in a multi-user VR cardiac arrest simulation, which processes simulation transcripts in near real-time, providing opportunities for precision feedback, practice, and/or remediation. We demonstrate that LLMs, when used with well-defined codes, moderate transcript contexts, and single-task prompts, can approach expert-level agreement for diagnostic and intervention reasoning. However, they face similar challenges to humans in identifying team-based non-technical skills. This pipeline provides explainable rationales that support error analysis and iterative schema improvement compared to traditional classifiers. Our results highlight key methodological considerations for deploying LLM-powered behavioral coding in high-stakes domains, establishing a foundation for future human-centered feedback systems.
Siqi Liang, PhD Student, School of Information
Bao Hoang, Professorial Assistant, Michigan State University
Yijiang Pang, PhD Student, Michigan State University
Hiroko Dodge, Professor of Neurology, Harvard University
Jiayu Zhou, Associate Professor of Information, School of Information
Mild Cognitive Impairment (MCI) is an early stage of dementia characterized by cognitive decline and behavioral changes. Early detection is crucial for timely interventions, improved clinical trial cohort selection, and the development of targeted therapies. Linguistic markers have recently emerged as a non-invasive, cost-eRective method for MCI detection. This study analyzes linguistic markers from conversations between participants and healthcare professionals to distinguish MCI from cognitively normal (NL) individuals. The dynamics of multiple conversations of a subject carry fine-granular linguistic change over time and expect to greatly enhance detection accuracy. However, individual variations in speaking styles pose challenges for learning cognitive characteristics from temporal sequences of conversations. To address this, we propose a temporal harmonization method to mitigate distributional diRerences in linguistic features across subjects, improving model generalization. Our results show that machine learning models leveraging subject-invariant harmonized temporal features greatly improve the prediction performance of MCI detection from multiple conversations.
Jiankun Wang, PhD Student, School of Information
Sumyeong Ahn, Postdoctoral Research Fellow, Michigan State University
Taykhoom Dalal, PhD Candidate, Cornell University
Xiaodan Zhang, Data Scientist, Michigan State University
Weishen Pan, Postdoctoral Associate, Cornell University
Qiannan Zhang, Postdoctoral Associate, Cornell University
Junyuan Hong, Postdoctoral Fellow, The University of Texas at Austin
Bin Chen, Associate Professor, Michigan State University
Hiroko Dodge, Professor of Neurology, Harvard University
Fei Wang, Associate Dean, Cornell University
Jiayu Zhou, Associate Professor of Information, School of Information
Early prediction of Alzheimer’s disease and related dementias (ADRD) from electronic health records (EHRs) is challenging yet crucial. We propose a hybrid, confidence-driven clinical decision-support system that combines traditional supervised learning models (SLs) with large language models (LLMs). By dynamically invoking the LLM’s reasoning on cases where the SL ensemble is uncertain, the framework significantly improves ADRD-risk prediction, enabling earlier intervention and more efficient trial recruitment.
For each patient, structured EHR snapshots—vitals, laboratory values, ICD-10 diagnoses, RxNorm medications, and CPT procedures—are automatically converted into de-identified natural-language summaries via a local LLaMA-2- 7B. When the ensemble SL confidence falls below a tunable threshold (e.g., σ = 0.6), these summaries—together with similarity-retrieved demonstrations—are fed to LLaMA-2-70B for in-context learning; high-confidence cases are resolved by the SLs alone. All computation runs on an on-premise GPU cluster, ensuring no protected health information leaves the secure environment.
We evaluate the approach on real-world data from Oregon Health & Science University (OHSU), encompassing more than 2.5 million patients and 20 million encounters. Across six benchmark tasks (spanning multiple computable phenotypes and prediction windows), the proposed pipeline yields a 8 % mean F1-score improvement over different baselines, underscoring its effectiveness. Ablation studies validate the contribution of SL-guided denoising for in context examples, and reveal that neither model scaling nor domain-specific fine-tuning consistently boosts performance—highlighting open questions for future work.
By selectively harnessing the strengths of SLs in clear-cut cases and LLMs in more complex scenarios, the proposed system delivers more reliable early ADRD screening without additional data collection or privacy risk, and is readily extensible to other EHR-based prediction tasks.
Haobo Zhang, PhD Student, School of Information
Alzheimer’s disease (AD) is a complex neurodegenerative disorder with data scattered across diverse research cohorts, modalities, and literature, posing a major barrier to biomarker discovery and translational progress. This proposal introduces the Multimodal Biomedical Knowledge Graph (MMBKG) — a first-of-its-kind resource that integrates heterogeneous AD data, including neuroimaging, genomics, clinical phenotypes, and literature-derived knowledge, into a unified, analyzable framework. Our approach addresses the fragmentation of existing AD datasets (e.g., ADNI, NACC, UK Biobank, All of Us) by harmonizing them using a hybrid OMOP+FHIR model and embedding their patient-level features into a graph-based structure. Through a novel “cohort grounding” strategy, we dynamically link abstract biomedical knowledge to empirical data across cohorts, enabling hypothesis validation and cross-cohort modeling. We will evaluate MMBKG’s effectiveness in enhancing early AD detection, cognitive outcome prediction, and biomarker identification through deployment on large-scale EHR datasets. The proposed work represents a paradigm shift in AD research, enabling rich, generalizable insights through integrated modeling of multi-modal, real-world and research data. The MMBKG framework will serve as a reusable foundation for knowledge-driven precision medicine in neurodegenerative disease and beyond.
Mei (Alice) Ruo-Syuan, Graduate Student Research Assistant, College of Engineering
Chenhui Shao, Associate Professor of Mechanical Engineering, College of Engineering
Geometric integrity is central to the functionality, reliability, and safety of manufactured products, making geometry-based qualification a core quality control activity. Although modern 3D metrology enables fine-scale inspection of dimensional accuracy, surface quality, and shape conformity, its adoption is limited by point cloud properties: high dimensionality, unstructured nature, large file sizes with interoperability constraints, and possible sparsity in defective regions. Converting raw scans into analysis-ready data requires substantial metrology and manufacturing expertise, creating barriers to scalable adoption of 3D metrology in automated quality inspection.
We developed a novel synthetic data generation pipeline that eliminates the dependence on largely labeled real datasets while enabling automated 3D geometric defect detection. Our methodology couples parametric computer-aided design (CAD) modeling with domain knowledge to generate comprehensive training data reflecting real-world manufacturing conditions. Focusing on 12 gear part families, we created four quality classes per family (nominal, tooth wear, tooth-root breakage, and pitting) and incorporated manufacturing variability. Case studies approximate shop-floor conditions by investigating (1) process-induced defects recurring in characteristic locations, (2) degradation-driven defects varying in size and location, and (3) metrological factors including resolution limits and measurement noise representative of industrial 3D scanners.
We developed a deep learning model adapted from the PointNet++ architecture for end-to-end 3D point cloud analysis, yielding results showing (1) 100% accuracy in part design classification and 85% accuracy in four-class defect detection, (2) that increased scanning resolution doesn’t always improve defect detection performance, as accuracy dropped from 76% to 57% when resolution increased from 0.681 mm to 0.787 mm, and (3) metrological factors influence quality control outcomes, with model performance dropping 6% as noise standard deviation increased from 0.8 mm to 1.0 mm. Additionally, we collected 3D measurements of various real-world parts to evaluate this approach and envisioned this will ultimately enable scalable geometric defect detection in manufacturing applications.
Li-Wei Shih, PhD Student, College of Engineering
Chenhui Shao, Associate Professor of Mechanical Engineering, College of Engineering
Vision-based human action recognition has significant potential in smart manufacturing applications such as human–robot collaborative assembly and safety surveillance, yet most existing studies rely on simplified setups and small datasets, which raises concerns about robustness and generalizability in production. This study considers procedure-level action recognition, using ultrasonic welding as an example that spans setup, execution, and post-process checks. We collect a new dataset with substantial variation in action sequences, camera viewpoints, and lighting conditions to reflect realistic factory environments, and we adapt a Video Swin Transformer-based model with lightweight temporal modules for scalable deployment. The model attains 94.7% accuracy under ideal conditions and shows modest performance drops under non-ideal conditions such as lower resolution, harsh lighting, and off-axis views. To address these drops and to enable fast, cost-effective deployment when training data are limited, we treat the full ideal-condition dataset as the source domain and each challenge case as a limited-data target, applying transfer learning to easier shifts like resolution and lighting and using domain adaptation for the more difficult cross-view case. This work contributes a procedure-level benchmark and an adaptable HAR pipeline that enable sample-efficient adaptation and reliable deployment in complex manufacturing environments.
Taofeeq Togunwa, PhD Student, Medical School
Akash Shanmugam, Barwon Health
Yi Geng, Bendigo General Hospital
Si Jing, I-Med Radiology Network
Introductory Summary: Ultrasound (US) reporting has become increasingly complex, demanding, and time consuming for radiologists, particularly in low-and-middle-income countries (LMICs), where reporting delays (>72 hours) hinder clinical decision-making. Multimodal artificial intelligence (AI) may support radiologists by generating preliminary radiological reports
after initial image assessment by these models, but the practical impact of such human-AI collaboration is unclear. This study evaluates GPT-4o-generated preliminary ultrasound reports for diagnostic accuracy and reporting efficiency.
Methods: We randomly selected 121 US cases from Radiopaedia, covering abdominal, breast, thyroid, scrotal, musculoskeletal, obstetric, and gynecological pathologies, categorized as diagnostically “certain” or “almost certain.” Clinical histories and US images were used to prompt GPT-4o, generating preliminary reports. A Royal Australian and New Zealand College of Radiologists (RANZCR) board-certified radiologist scored reports for accuracy of findings and appropriateness of management suggestions. Revision time to achieve clinically adequate reports was compared with published manual reporting benchmarks*.
Main Results: GPT-4o-generated reports achieved a mean accuracy of 52% (SE=3%), indicating moderate clinical utility but inconsistent capture of full radiological details. Across most US modalities, report revision time decreased significantly (p<0.01), except for leg vein ultrasound (p=0.5). Abdominal Doppler US showed the greatest efficiency gain, saving an average of 238 seconds per report (p<0.001).
Impact: GPT-4o demonstrates potential as a collaborative partner for radiologists, reducing reporting delays while maintaining human oversight. Completeness and consistency remain limited, highlighting the need for specialized, ultrasound-focused multimodal AI models. By combining AI assistance with expert evaluation, this approach could improve workflow efficiency and decision-making in LMIC radiology settings, especially where human radiology expertise is limited. Future studies should assess real-world applicability across diverse healthcare environments.
Atharva Kashyap, PhD Student, College of Engineering
Physically Assistive Robots help people with physical disabilities perform various activities of daily living. Largely, these robots overfit particular populations or groups of people, prompting a need for personalization. In this work, we present a work-in-progress framework to evaluate the performance of Large Language Models (LLM) in their ability to reason about physical disabilities, especially in context of supporting personalization. We compiled a dataset that includes five physical disabilities, generated several anatomically accurate physical functions of the body, and curated brief real medical histories for people with a particular physical disability. Using the disability dataset, real patient histories, and body functions, we designed prompts. We expect to use zero-shot prompting techniques to evaluate LLM performance in this context. Our next step is to conduct an LLM evaluation study against ground truth compiled through a data collection study with doctors.
Tanay Panja, Undergraduate Student, U-M
Maple (Acer) pollen is a major aeroallergen across the United States, contributing to seasonal allergic rhinitis and asthma exacerbations worldwide. Accurate pollen forecasting is vital for public health prepared ness and clinical response, requiring specialized models for common pollen-dispersing species. Yet existing approaches are largely limited to sparse, generalized observational networks. Furthermore, short-term pre diction is an understudied aspect in observational studies.
We present a novel spatiotemporal machine learning framework that integrates pollen observations with comprehensive environmental data and autoregressive features to predict daily Acer pollen concentrations. Meteorological predictors—including solar radiation, precipitation, temperature extremes, vapor pressure, and snow water equivalent—were lagged at 1-week, 1-month, and 3-month intervals to capture delayed effects. Additional engineered features incorporated cyclical seasonality and nonlinear interactions.
Five classification models, logistic regression, random forest, XGBoost, LightGBM, and a multi-layer perceptron, were benchmarked against persistence and climatology using stratified cross-validation. The multi-layer perceptron achieved the best performance, with 68.2% accuracy and a weighted F1-score of 0.684, representing a 16 percentage point improvement over environmental-only models and substantially outperforming climatology. Lagged features consistently ranked among the strongest predictors, confirming the value of autoregressive modeling. We applied Ordinary Kriging for spatial interpolation, generating continuous daily pollen surfaces across the continental United States.
The novelty of this study lies in combining autoregressive machine learning with multiscale meteorological predictors and geostatistical interpolation, moving beyond static climatology. This integrated approach captures short-term variability and seasonal dynamics, delivering more accurate predictions than traditional climatology-based models.
Our results show that machine learning can transform aeroallergen forecasting with actionable, high resolution predictions. By advancing beyond climatology, this framework provides a foundation for public health tools to mitigate seasonal allergies and asthma.
Keywords: climatology, spatiotemporal modeling, pollen forecasting, environmental health
Funding: This work was supported by MIDAS through the Propelling Original Data Science (PODS) grant award.
Tiffany Wu, PhD Student, College of Literature, Science, and the Arts
Chronic absenteeism is a critical issue that has been linked to many long-term academic and social-emotional consequences for students. The current study focuses on improving a key system already in place in many school districts—early warning systems (EWSs)—in order to decrease chronic absenteeism in students’ earliest schooling years. Using a demographically diverse population of students followed from PreK to third grade in Boston Public Schools (N=6,698), we demonstrate how and why two machine learning (ML) algorithms—the Synthetic Minority Oversampling Technique (SMOTE) and Extreme Gradient Boosting (XGBoost)—can improve EWS accuracy in proactively identifying students who are at risk of becoming chronically absent, outperforming both traditional logistic regression and other ML models. The best-performing XGBoost model with SMOTE was approximately 54 percentage points more accurate (in terms of recall rate) than the logistic regression model closest to those used in current EWSs. By improving early risk identification, these models can help schools get supports to the students who need them most before patterns of disengagement become harder to reverse.
Importantly, we find that models using only behavioral indicators like attendance rate, and excluding student demographic characteristics such as race and socioeconomic status, maintained comparable predictive performance. This suggests that districts can build effective EWSs that avoid relying on sensitive demographic variables, reducing the risk of reinforcing existing disparities while still targeting students most in need of support. SHAP (SHapley Additive exPlanations) analysis further revealed that recent attendance patterns were consistently the strongest predictors of chronic absenteeism, reinforcing the feasibility of building transparent, actionable models centered on behavioral data. Lastly, our analyses introduce varying probability thresholds and the incorporation of different years of data, showing the potential of these models to cater to school districts aiming to leverage ML predictions while adhering to budgetary or intervention constraints.
Angana Borah, PhD Candidate, College of Engineering
Existing challenges in misinformation exposure and susceptibility vary across demographic groups, as some populations are more vulnerable to misinformation than others. Large language models (LLMs) introduce new dimensions to these challenges through their ability to generate persuasive content at scale and reinforcing existing biases. This study investigates the bidirectional persuasion dynamics between LLMs and humans when exposed to misinformative content. We analyze human-to-LLM influence using human-stance datasets and assess LLM-to-human influence by generating LLM-based persuasive arguments. Additionally, we use a multi-agent LLM framework to analyze the spread of misinformation under persuasion among demographic-oriented LLM agents. Our findings show that demographic factors influence susceptibility to misinformation in LLMs, closely reflecting the demographic-based patterns seen in human susceptibility. We also find that, similar to human demographic groups, multi-agent LLMs exhibit echo chamber behavior. This research explores the interplay between humans and LLMs, highlighting demographic differences in the context of misinformation and offering insights for future interventions.
Maria Fields, PhD Student, College of Engineering
Pedro Ávila Villarreal, Undergraduate Student, University of Monterrey
Yili Liu, Professor of Industrial and Operations Engineering, College of Engineering
Generative Artificial Intelligence (GenAI) has rapidly become prevalent in our society with many applications from chatbots to media creation. With the potential to affect all industries and many sectors of society, GenAI is projected to add trillions of dollars to the global economy and affect billions of people’s lives. However, there are substantial issues with GenAI that our society needs to address. It is our mission to decipher the dynamic relationship between humans and GenAI to help design and develop Human-Centered GenAI.
Cultural attitudes and societal norms greatly influence the acceptance and utilization of new technologies such as Generative Artificial Intelligence (GenAI). This study compares opinions toward GenAI between three large Spanish-speaking countries (Mexico, Spain, Argentina) on three continents and four large English-speaking countries (the US, UK, India, and Nigeria) on four continents. One hundred forty-two text sources were collected via the internet, containing text materials discussing GenAI in these countries. Thematic Analysis of these texts was used to identify their major themes, and the results revealed the top 8 themes for the Spanish-speaking countries and the English-speaking countries, respectively. The results showed similarities and differences both between Spanish- and English- speaking countries and within each language group. This research provides concrete evidence that it is important to consider cultural aspects of GenAI as one of the human factors issues that need urgent attention.
Daniel Menacho Ordoñez, Graduate Student, College of Engineering
Ricardo A. Gonzales, PhD Candidate, University of Oxford
Matthew K. Burrage, Clinical Research Fellow, University of Oxford
Ibrahim Altun, University of Queensland
Betty Raman, Associate Professor of Cardiovascular Medicine, University of Oxford
Rina Ariga, Junior Research Fellow, University of Oxford
Rohan S. Wijesurendra, Senior Clinical Research Fellow, University of Oxford
Masliza Mahmod, Head of Clinicals Trials Group, University of Oxford
Eylem Levelt, Professor of Cardiology, University of Melbourne
Wei-Ming Huang, Assistant Professor of Radiology, MacKay Memorial Hospital
Chun-Ho Yun, Radiology Director, MacKay Memorial Hospital
Vanessa M. Ferreira, Deputy Clinical Director, University of Oxford
Qiang Zhang, Associate Professor of AI in Cardiovascular Imaging, University of Oxford
Stefan K. Piechnik, Professor of Biomedical Imaging, University of Oxford
Motivation: Cardiovascular magnetic resonance stress perfusion imaging is a promising sequence for diagnosing and managing ischemic heart disease [1]. It detects myocardial ischemia and predicts long-term prognosis in coronary artery disease [2,3]. However, its acquisition requires gadolinium-based contrast agents (GBCA), which are contraindicated in some patients. Alternative approaches include synthetic images [4] and stress T1 mapping without GBCA, though this latter suffers from co-registration issues and pixel-wise visualization. We propose a transformer-based model for automatic co-registration of rest/stress T1 maps to improve the visualization and stress T1 reactivity.
Materials and Methods: This retrospective multi-center study included 4,840 T1 maps from 561 subjects acquired on 1.5T and 3.0T Siemens MRI scanners. The database comprised of 103 healthy volunteers and 458 patients with cardiovascular conditions. We implemented a Swin Transformer [5] as the core of the VoxelMorph framework [6] for non-rigid registration of stress and rest T1 maps, enabling pixel-wise percentage differences in T1 values (dT1). Myocardial stress T1 reactivity from these dT1 maps was compared against manually derived values. Evaluation metrics used were intra-class correlation (ICC) and mixed-model ANOVA.
Results: The deep-learning registration model demonstrated excellent performance in replicating manual myocardial stress T1 reactivity derived from dT1 maps. Agreement with manual measurements was strong (ICC = 0.97, mean difference −0.06 ± 0.74%). Moreover, the model showed no significant dependence (p > 0.1) on acquisition method (magnetic field strength, slice position) or patient features (pharmacological agents, gender).
Conclusion: We developed a novel approach to enhance pixel-wise visualization and reactivity quantification for stress T1 maps, replicating manual performance. Our method leverages the Swin-Transformer architecture to achieve more reliable co-registration of stress/rest T1 maps across diverse acquisition conditions and patient demographics.
David Hanauer, Clinical Associate Professor of Learning Health Science, Medical School
Simon Shavit, Undergraduate Student, College of Literature, Science, and the Arts
Introduction
Researchers and clinicians often turn to electronic health record (EHR) data for details on dietary intake and concerns for allergic reactions. In this work we created a knowledge resource to match foods from around the world to their common ingredients and potential allergens.
Methods
We used Wikipedia (2024-12-01 English download) to develop a list of foods from around the world. Wikipedia food pages contain a metadata field main_ingredient that was used to identify foods. Using the food list obtained from Wikipedia, we created a catalog of ingredients for each food using prompts sent to the large language model (LLM) OpenAI ChatGPT 3.5 API (CGPT). For each food name and associated ingredients, we prompted CGPT for a list of allergens in each food based on the eight major food allergens described by the Food Allergen Labeling and Consumer Protection Act: milk, eggs, fish, shellfish, tree nuts, peanuts, wheat, and soy. Prompts and knowledge resources are available at: https://github.com/dhanauer/food-allergens.
Results
Using Wikipedia, we identified 6,693 foods with a main_ingredient field. From this list, we found 7,030 distinct ingredients. Using CGPT for the same foods, a list of 4,083 ingredients was obtained, suggesting more consistency and less variability. An UpSet plot of allergens in the foods is shown in the Figure. We noted issues with both the Wikipedia output and CGPT output: for example, both Wikipedia and CGPT did not always list ingredients in a consistent manner, and some were quite free-form.
Discussion
We found Wikipedia to be most useful for creating an overall list of foods, but the LLM was better at providing a consistent list of ingredients. A combined approach proved to be most accurate.
Chandler Sachs, PhD Candidate, School for Environment and Sustainability
Abstract: Critical minerals play a central role in technologies of the 21st century, with nations across the globe racing to access these raw materials that underpin renewable energy technologies, aerospace systems, and AI infrastructure. Demand for lithium, one critical mineral vital for decarbonization, is projected to more than triple by 2030 under baseline scenarios. Despite the centrality of lithium and other minerals critical to the energy transition, we still lack spatially explicit, up-to-date data on where mining occurs, how it evolves, and what its impacts are. Existing mining datasets are fragmented, manually digitized, or derived from low-resolution imagery, limiting their utility for science and for policy. Recent advances in geospatial artificial intelligence (GeoAI) have allowed for improvements in identification and monitoring of complex land use changes, like those generated by mining. Employing these advancements, we have developed SAGE (Scalable AI for Geospatial Extraction), a minimally-supervised segmentation framework that combines remote sensing science with Meta’s Segment Anything Model (SAM), a state-of-the-art foundation vision model for zero- and few-shot image segmentation. By adapting SAM to extractive landscapes, SAGE overcomes the limitations of manual delineation and supervised classification, enabling precise mining detection at scale. The framework will integrate SAM with multispectral satellite data, using tailored spectral indices to enhance performance. Deployed on the UM’s Great Lakes high-performance computing cluster, the pipeline will process global imagery to produce the first comprehensive, analysis-ready geodatabase of lithium mines. This type of global, automated segmentation of lithium mining using foundation models has never been attempted, and its success would represent a groundbreaking advance in remote sensing and geospatial AI. The resulting dataset will enable new insights into environmental monitoring, resource governance, and critical mineral supply chains. SAGE will be transferable to other critical minerals, laying the groundwork for continuous, global-scale monitoring of extraction activities over time and space.
Melody Zhang, PhD Candidate, College of Engineering
Shih-Kuang (Alex) Lee, Graduate Student Research Assistant, College of Engineering
Sharon C. Glotzer, Professor of Chemical Engineering, Professor of Materials Science and Engineering, College of Engineering
Rebecca K. Lindsey, Assistant Professor of Materials Science and Engineering, College of Engineering
Molecular simulations are essential for exploring and optimizing complex colloidal nanoparticle systems, yet they often require a tradeoff between accuracy and computational cost. All-atom simulations provide detailed resolution of interparticle interactions but are prohibitively expensive for large systems. Coarse-grained (CG) approaches overcome this limitation by grouping atoms into an individual interaction site, retaining key chemical features while enabling efficient simulations. However, the high tunability of polymer-grafted nanoparticles (PGNs), namely design parameters such as polymer length and grafting density, makes it challenging to identify optimal conditions for targeted self-assembly. Inverse design strategies address this problem by searching for PGN configurations that yield desired structures, but their success depends critically on accurate descriptions of PGN interactions with respect to inter-PGN distance and the tunable design parameter. Machine-learned interaction models (ML-IAMs) can learn these intricate and higher-dimensional interaction, making them promising tools for describing interaction landscape for inverse design methods.
Here, we present a generalized framework that incorporates a physics-informed “alchemical” dimension into ML-IAMs, enabling continuous mapping between PGN interaction landscapes and tunable molecular design parameters. We integrate steered molecular dynamics (SMD) with the forward–reverse (FR) method1 into Zhou et al.’s grid sampling strategy2, allowing efficient generation of potential of mean force (PMF) data for CG PGNs with varying polymer lengths. We use ChIMES3, a physics-informed ML-IAM, to accurately reproduce PMFs and validate equilibrium structural and thermodynamic properties. We further extend ChIMES into an alchemical model, termed X-ChIMES, which unifies the PMF landscape across PGN design variables. Finally, by embedding X-ChIMES into the Digital Alchemy framework4,5, we demonstrate inverse design of PGNs toward target crystal structures and reveal how optimal alchemical parameters shift with self-assembly targets. Our framework provides a broadly applicable strategy for integrating machine learning, coarse-graining, and alchemical design into the discovery of self-assembling nanomaterials.
Faculty Posters
Yujing Yang, Postdoctoral Research Fellow, College of Engineering
With rapid advances in three-dimensional (3D) sensing technologies, point cloud data have become increasingly accessible for process monitoring and anomaly detection in additive manufacturing (AM). However, extracting defect-pertinent information from point clouds remains a critical challenge. The state-of-the-art approaches mainly operate in the spatial domain, i.e., analyzing the raw x, y, and z coordinates of points. They are limited in the ability to fully reveal defect-induced variations, especially in the presence of noises, and distinguish them from benign patterns. To fill this gap, we develop a new framework that shifts point cloud-based AM defect detection from the spatial to the spectral domain. Point clouds are modeled as graphs, which are then processed by spectral filters derived from diffusion wavelets. As such, both global and local surface irregularities on point clouds can be highlighted across multiple scales, forming a high dimensional tensor. A new online learning scheme is further designed that integrates online tensor decomposition with incremental one-class learning to enable efficient layerwise monitoring. Evaluations on simulation and real-world case studies demonstrate the effectiveness and computational efficiency of the approach in capturing a wide range of defect types. This work highlights how the synergy between rich 3D data and advanced AI
Fiona Molloy, Postdoctoral Research Fellow, Medical School
Aman Taxali, Research Statistician, Medical School
Mike Angstadt, Statistician Staff Specialist, Medical School
Katherine Toda-Thorne, Clinical Research Technician, Medical School
Katherine L. McCurry, Assistant Professor of Psychiatry, Medical School
Alexander Weigard, Assistant Professor of Psychiatry, Medical School
Omid Kardan, Assistant Professor of Psychiatry, Medical School
Camille Lehrmann, Clinical Research Technician, Medical School
Joshua Vens, ABCD Research Assistant, Medical School
Cleanthis Michael, Graduate Student Research Assistant, College of Literature, Science, and the Arts
Mary M. Heitzeg, Professor of Psychiatry, Medical School
Chandra Sripada, Professor of Psychiatry, Medical School
Sleep is critical for social, academic, and occupational outcomes, yet most adolescents receive less than the recommended amount of sleep. Thus, there is a need to understand how reduced sleep affects the developing brain, which in turn can inform early risk identification and novel interventions. There is substantial evidence that sleep duration has complex interlinkages with brain connectivity, but the directionality of this relationship and associated longitudinal changes remain unclear: does reduced sleep cause connectivity changes or are connectivity changes the cause of reduced sleep? Our novel “little data informs big data” approach combines the Adolescent Brain Cognitive Development (ABCD) Study, a longitudinal observational study of 10,878 youth, and a second sample of 76 participants scanned before and after a sleep deprivation causal manipulation. First, in the ABCD dataset, we identified a robust and generalizable neurosignature of reduced sleep using multivariate predictive modeling and multimodal sleep assessment. Second, in an independent sample of ABCD participants, we demonstrate that greater reductions in sleep duration across two years are significantly related to greater expression of this neurosignature. Third, in the sleep deprivation dataset, we show that expression of the ABCD reduced sleep neurosignature is significantly increased within individuals following sleep deprivation, and that neurosignatures of reduced sleep from the two samples exhibit significant spatial correspondence. This suggests that the connectivity patterns captured in the neurosignature are a consequence of reduced sleep. These results clarify links between sleep and the developing brain and provide the strongest evidence to date that sleep produces characteristic brain functional connectivity changes across adolescence.
Greg Rybarczyk, Professor of Geography, University of Michigan – Flint
Kristine Zhou
Xiao Li, Assistant Professor of Land Surveying and Geo-Informatics, The Hong Kong Polytechnic University
Hritik Ghorpade, Research Assistant, University of Michigan – Dearborn
The adoption of micromobility devices, particularly e-bikes and e-scooters, represents a significant shift in urban transportation. Despite the growing adoption of these modalities, significant knowledge gaps persist in understanding the environmental contexts and temporal patterns of crashes and injuries in the United States. This study addresses these limitations through a comprehensive ten-year analysis (2014-2023) of e-bike and e-scooter-related injuries and hospitalizations using the National Electronic Injury Surveillance System (NEISS) database. Our methodology combines natural language processing of injury narratives with binary logistic regression to examine how environmental factors, socio-demographic variables, and infrastructure characteristics influence injury and hospitalization rates. The regression analysis revealed that e-scooter riders faced dramatically elevated injury risk from pedestrian collisions (OR=6.465, p=0.023) and younger age groups showed increased injury likelihood (under 18: OR=1.878, p=0.004). Conversely, e-bike riders under 18 experienced 75% lower injury odds (OR=0.247, p<0.001), though riding in recreational settings doubled injury risk (OR=2.449, p=0.022). For hospitalizations, substance involvement more than doubled admission risk for e-scooter riders (OR=2.477, p<0.001), while e-bike collisions with motor vehicles increased hospitalization odds 2.6-fold (OR=2.589, p<0.001), yet younger patients paradoxically showed reduced admission likelihood when hospitalized (under 18:OR=0.211 for e-scooters, OR=0.228 for e-bikes, both p<0.001). Through longitudinal analysis of injury circumstances and outcomes across diverse geographic contexts, the outcome of this research aims to inform evidence-based safety interventions and urban design policies that can enhance micromobility safety.
Morgan Sielaff, Research Area Specialist, Medical School
Despite rapid growth of artificial intelligence (AI) in healthcare, patients lack clear, evidence-based information on its impact on their care. We aimed to identify key information for a health AI tool label and explore community perspectives on ethical and trustworthy AI adoption. In 2024, we conducted five virtual community deliberations across Michigan, engaging 159 participants. Participants completed 20-minute pre- and post-surveys to assess changes in trust, knowledge, and attitudes about AI in healthcare. The 6-hour deliberation sessions featured expert presentations and two small-group discussions. Participants were recruited from community organizations across Michigan (Friends of Parkside, ACCESS). In the first small-group discussion, participants were optimistic about AI’s potential but concerned about safety, bias, and lack of transparency. Participants emphasized the need for transparency, oversight, and education. Additionally, participants were interested in governance that would allow for community involvement in AI development and healthcare practice policies. When participants completed the labeling activity, we found that participants deemed privacy and security, health equity, and safety and effectiveness of AI tools as priorities for inclusion on a health AI tool label. The post-deliberation survey results showed that 94% of participants agreed that “health systems should tell patients how they use AI tools in their healthcare.” These findings underscore the importance of patient-informed solutions that promote transparency, oversight, and engagement. Community members acknowledge both the potential benefits and risks of AI, emphasizing the need for clear, accessible information about how AI tools are used. There is strong public demand for transparent communication, engagement, and oversight to ensure that adoption of AI is both trustworthy and patient-centered. A labeling system could help address ethical concerns and support trustworthy and equitable integration of AI into healthcare. These insights may be useful in guiding policy and practice to advance safe, effective and equitable integration of AI in healthcare.
Doug Rosin, Research Associate, Institute for Social Research
Researchers at the Census Bureau and the University of Michigan are collaborating to digitize and link respondent information from the 1960-1990 U.S. Censuses. While the majority of this data is handwritten records stored on microfilm, the “group quarters” (GQ) residents in the 1990 Decennial Census were stored as condensed-font printouts on microfiche. This poster describes our eRorts to digitize the GQ records for individuals living in institutional or shared living situations, including dormitories, prisons, hospitals, etc. The data, covering approximately 6.6 million people living in 142,000 GQs, is structured in a regular format of machine-printed individual person records that are grouped by the GQ in which the person lives. The data was originally stored as 65,000 microfiche images that we digitized into TIF files. We extracted data from these TIFs using Tesseract, an OCR engine that uses a long short-term memory neural network to identify and output text in images. The main problem we faced in digitizing this data is that there is almost a complete lack of character and line spacing in the images; in many cases, the characters are set so closely that they overlap. Due to the highly condensed print, Tesseract has diRiculty recognizing the text, producing more errors than accurate output. Computer vision-based image preprocessing methods alone did not address the problem. Our poster presents a novel method we developed, using pixel density to detect areas where whitespace exists or would exist if the text was not condensed. Using those measurements, we cropped the image into sections, rows, and columns, and padded, preprocessed, and OCRed the cropped images. We also further trained the Tesseract models by providing additional dictionaries and pattern files. We conclude the poster with data showing the impact of various combinations of preprocessing, pixel density cropping, and model customization on OCR output accuracy.
Matthias Wilms, Assistant Professor of Radiology, Medical School
The human brain undergoes dynamic structural changes throughout the lifespan, some of which can be associated with neurodegenerative conditions such as Alzheimer’s disease. Longitudinal Magnetic Resonance Imaging (MRI) and other neuroimaging data are valuable for analyzing those trajectories of change and to determine if the change pattern is associated with typical or atypical aging. However, the analysis of longitudinal imaging data with machine learning tools is highly challenging given their discrete nature with diJerent spatial and temporal image sampling patterns within individuals and across populations. This leads to computational problems for most traditional deep learning methods that cannot represent the underlying continuous biological process and, therefore, resort to diJerent resampling schemes to unify the images, potentially resulting in a loss of information. We, therefore, propose to model and analyze this data in a spatially and temporally continuous fashion. To achieve this, we present a novel fully data-driven method for representing aging trajectories across the entire brain by modelling subject-specific longitudinal T1-weighted MRI data as continuous functions using Implicit Neural Representations (INRs). Specifically, we introduce a novel INR architecture that partially disentangles spatial and temporal trajectory information, and we subsequently design a new and eJicient classifier that directly operates on the INRs’ parameter space to classify brain aging trajectories. To evaluate our method in a controlled data environment, we develop a biologically grounded trajectory simulation scheme and generate longitudinal T1-weighted 3D MRI data for 450 healthy and Alzheimer’s-like subjects at regularly and irregularly sampled timepoints. In the more realistic irregular sampling experiment, our INR-based method achieves 81.3% accuracy for the brain aging trajectory classification task, outperforming a standard deep learning baseline model 73.7%. Our method provides a new foundation for machine learning-based analyses of the ever-growing amount of longitudinal neuroimaging data.
Robert Tomlinson, Research Computer Specialist, Life Sciences Institute
Yujia Hu, Research Associate, Cleveland Clinic
Bing Ye, Research Professor, Life Sciences Institute
Behavior is a critical function of the brain. As such, behavioral analysis is widely applied in neuroscience, psychology, evolutionary biology, ecology, medicine, and many other research fields. However, behavior analysis has been constrained by labor-intensive traditional methods or over-simplified automated methods. LabGym, an open-source AI-enabled software tool, has reimagined the process of behavior analysis through automated computer vision-driven methods that transform behavior into a quantifiable and repeatable dimension of science.
LabGym identifies behaviors holistically, which in turn uniquely enables categorization of user-defined behaviors and the quantifications of behavior-related metrics.LabGym further extends the boundary of behavior analysis by enabling social behaviors to be studied at the level of each individual. Its efficacy has been demonstrated across diverse animal species, from soft-bodied worms to primates.
LabGym’s initial efforts represent a solution to a longstanding pain point within the neuroscience community. However, since LabGym offers a flexible and intuitive platform for users to efficiently analyze behavior, it has surpassed 59,000 GitHub downloads, and has garnered widespread adoption within additional fields such as ecology, evolutionary biology, environmental science, and clinical research. Notably, it is not skeuomorphic – rather than merely accelerating existing workflows, it opens the door to new types of experiments and insights that were once out of reach.
Xiaofeng Liu, Schmidt AI in Science Fellow, Michigan Institute for Data & AI in Society
Yunsu Park, Undergraduate Student, College of Literature, Science, and the Arts
Yuyue Zhu, Graduate Student, New York University
Yi Hong, Assistant Research Scientist, School for Environment and Sustainability
The hydrological simulation of large, transboundary water systems such as the Laurentian Great Lakes remains challenging. Although deep learning has advanced hydrologic forecasting, prior efforts have been fragmented, and no unified basin-wide model for daily streamflow has been established. We address this gap by developing a single Entity-Aware Long Short-Term Memory (EA-LSTM) model trained without basin-specific calibration. We compile a cross-border dataset integrating daily meteorological forcings, static catchment attributes, and observed streamflow for 975 sub-basins across the United States and Canada (1980–2023) and train using a temporal split. The unified EA-LSTM attains a median Nash–Sutcliffe Efficiency (NSE) of 0.685 and a median Kling–Gupta Efficiency (KGE) of 0.678 in validation, substantially exceeding a standard LSTM (median NSE 0.567, KGE 0.555) and the operational NOAA National Water Model on a common subset (median NSE 0.209, KGE 0.440). Although the skill is reduced in very small and highly urbanized catchments and during extreme events, the performance is robust across diverse hydroclimatic settings. These results demonstrate that a single, calibration-free deep learning model can provide accurate, scalable streamflow prediction across an international basin, offering a practical path toward unified forecasting for the Great Lakes and a transferable framework for other large, data-sparse watersheds.
Minoh Jeong, Postdoctoral Research Fellow, College of Engineering
Seonho Kim, PhD Student, The Ohio State University
Alfred Hero, Professor of Electrical Engineering and Computer Science, College of Engineering
Deterministic embeddings learned by contrastive learning (CL) methods such as SimCLR and SupCon achieve state‐of‐the‐art performance but lack a principled mechanism for uncertainty quantification. We propose Variational Contrastive Learning (VCL), a decoder‐free framework that maximizes the evidence lower bound (ELBO) by interpreting the InfoNCE loss as a surrogate reconstruction term and adding a KL divergence regularizer to a uniform prior on the qθ(z| x) unit hypersphere. We model the approximate posterior as a projected normal distribution, enabling the sampling of probabilistic embeddings. Our two instantiations— qθ(z| x) VSimCLR and VSupCon—replace deterministic embeddings with samples from and incorporate a normalized KL term into the loss. Experiments on multiple benchmarks demonstrate that VCL mitigates dimensional collapse, enhances mutual information with class labels, and matches or outperforms deterministic baselines in classification accuracy, all the while providing meaningful uncertainty estimates through the posterior model. VCL thus equips contrastive learning with a probabilistic foundation, serving as a new basis for contrastive approaches.
Long-Jing Hsu, Schmidt AI in Science Postdoctoral Fellow, Michigan Institute for Data & AI in Society
Jung-Chun Liu, Graduate Student Research Assistant, College of Engineering
Atharva Kashyap, Graduate Student Research Assistant, College of Engineering
Conversational robots, powered by large language models (LLMs) and artificial intelligence (AI), play an increasingly important role in shaping humanity’s future. However, little research has considered the resilience of human–AI collaboration. Fostering resilience in the age of AI is especially important given concerns about LLM errors, human over-reliance, and the potential substitution of human connection. Beyond merely preventing these adverse outcomes, it is also crucial to help humans build resilience in the face of such societal changes.
Drawing from a systematic review of research on robots and resilience (using the ACM Digital Library and Scopus databases), including screening 22 papers and an in-depth analysis of resilience literature, we identified key components necessary for designing AI systems that enhance, rather than undermine, human resilience. Our findings reveal how robots can be a positive inner voice that helps users discover and strengthen their capacity to cope with challenges. Examples include robots portrayed as companions, tools such as coaches encouraging positive self-reflection, and mediators that support professionals in supporting a constructive inner voice.
Following this review, we propose a robot with algorithms powered by large language models (LLMs) to analyze language patterns and keywords that indicate both the severity of adversity and a person’s level of resilience. The identified resilience level could then guide the robot in tailoring its interaction- ranging from minimal support to targeted interventions- by drawing on coping strategies and reflection on personal strengths (see figure below).
This review and proposed framework represent an early step toward linking resilience research with AI collaboration in fostering human resilience.
Douglas Craig, Statistician Staff Specialist, Medical School
Summary: We transform de-identified Emergency Department (ED) free-text notes into a compact, regular grammar, ED-ese, and evaluate whether grammar-constrained language modeling can predict next clinical steps and outcomes while remaining human interpretable.
Methods: We designed a minimal token inventory and finite-state grammar that segments each encounter into phases (triage, workup, treatment, disposition). A lightweight NLP compiler maps notes to ED-ese phrases. We then train two sequence models: (1) a strong non-neural baseline (n-gram/weighted finite-state model) and (2) a small GPT-style transformer fine-tuned on ED-ese with grammar-masked loss and grammar-constrained decoding. Evaluation focuses on next-token prediction, sequence-ending ranking for disposition (AUROC/PR), and calibration. We also mine recurrent “story motifs” (frequent episodes archetypes) to support case-based retrieval.
Data: Initial scope is ~50,000 de-identified chest-pain encounters from ED notes; splits are patient-level with site-aware holdouts.
Status & Results: Early stage. Seed grammar, token lexicons, and compiler are implemented; corpus processing and baseline modeling are underway. We present the design, example ED-ese narratives, coverage/error analyses, and preliminary predictive metrics.
Impact: Compiling heterogeneous notes into finite-state “stories” enables fast, auditable prediction and cohort discovery without heavy ontology mapping. Grammar-constrained LLMs return probability distributions over valid next steps, creating interpretable decision support signals and a reusable substrate for retrieval and quality-improvement studies. The approach is syndrome-agnostic and designed to generalize beyond chest pain to broader ED and inpatient workflows. Ultimately, we hope to show how a curated corpus of ED-ese transformed encounters could also serve as a foundation for training and education.
Amir Moosavi, Michigan Data Science Fellow, Michigan Institute for Data & AI in Society
Purpose: Increased utilization of marginal kidneys can expand the donor pool; however, concerns about outcomes and center-level variability have limited widespread adoption. We aimed to investigate national center-level trends in the marginal kidney post-transplant outcomes, hypothesizing that there is a significant difference between centers.
Methods: We performed a retrospective analysis of deceased-donor kidney transplants in adult recipients (≥18 years) from the Organ Procurement and Transplantation Network’s STAR file between January 1, 2010, and December 31, 2023. Transplants with Kidney Donor Profile Index (KDPI ) >85% were classified as marginal, and those with KDPI ≤85% as standard. Only centers performing ≥10 transplants annually (2019-2023) were included. Donor-recipient demographics (e.g., age, sex, dialysis status) and center characteristics were analyzed alongside four outcome metrics. We studied the 1-year graft failure as the primary endpoint, while delayed graft function, acute rejection, and length of stay were treated as secondary endpoints (metrics selected based on data availability). For each metric, a logistic regression model with a random intercept was fit using backward elimination (significance: 10%). Confidence intervals for center coefficients were extracted from the models and subsequently used to classify centers into outcome groups (“worse,” “normal,” or “better”) relative to a metric-specific reference transplant center.
Results: In our analysis, 33 centers (31.4 %) showed worse 1-year graft failure outcomes, while ten centers (9.5 %) exhibited worse delayed graft function outcomes. A majority of 64 centers (60.9%) demonstrated worse performance in at least one of the metrics. In the primary metric, centers exhibited substantial differences that were largely independent of their results on secondary metrics. When categorized by the primary metric outcomes, centers showed marked variability across all secondary endpoints (see the Figure), underscoring significant between center differences. Between metrics, variability was also evident (e.g., among 21 centers with worse outcomes in the primary metric, ten maintained better outcomes in the delayed graft function).
Conclusions: Our analysis indicates that transplant centers demonstrate significant variability (potentially systematic) in post-transplant outcomes of marginal kidneys. We identified that the majority of centers can benefit from targeted quality improvements for at least one of the metrics, while standardizing practices across centers can potentially reduce between-center variabilities.
Greta Branford, Clinical Assistant Professor, Associate Chief Medical Information Officer, Medical School
Introduction
Ambient listening is a technology that securely records patient-clinician conversations during visits and processes the recording to create structured clinical documentation. Generated notes are then available for the clinician to review, edit, and finalize within the EHR system. From an organizational perspective, there are several proposed benefits of this technology in patient care settings, including reducing overall administrative burden, improving provider efficiency, enhancing patient relationships, and streamlining EHR integration. However, the impact of this technology on clinician work-related well-being, is relatively less understood – yet equally important – given overall declines in physician and advanced practice provider (APP) well-being in recent years (e.g., Aiken et al., 2023).
Method & Results
Accordingly, we evaluated whether and to what extent DAX Copilot, an AI-powered tool from Microsoft that automates clinical documentation, is associated with pre-post changes in various provider well-being metrics. Overall, across participating providers (N = 196), repeated measures t-tests indicated that tool use was associated with statistically significant decreases in self-reported work-related burnout, time spent charting after hours, interference of work with family life, cognitive load, and increased job satisfaction. Trivial changes were observed for intentions to leave. We also conducted provider group difference analyses for key outcomes: positive improvements were generally observed across all groups and outcomes, with women and primary care providers experiencing the most significant benefits. In contrast, surgical providers and non-physicians (e.g., APPs) reported fewer improvements as compared to other provider groups.
Impact
This preliminary study found that innovative ambient AI was associated with improvements in several key provider well-being metrics, but with varying outcomes by gender identity and clinical specialty. These results suggest that ambient AI may be a potential mechanism for improving the experience of work for clinicians by decreasing the burden of clinical documentation.
Hojun Son, Research Fellow, College of Engineering
The quality and structure of data are essential in shaping AI capabilities, yet advancing knowledge and discovery requires more than simply increasing data volume. Latent features, difficult to capture at the raw data level, can be learned by AI models that aim to maximize performance without understanding their contextual meaning. We analyze the consequences of such data–AI interactions through domain-adaptive object detection (DAOD), where a lack of context understanding can cause performance degradation.
Object detection plays a crucial role in interpreting sensor data by recognizing objects at specific time instances. However, models can suffer severe performance drops when applied to novel domains. In this talk, we investigate how latent associations between foreground (FG) and background (BG) features influence object detection, with a focus on DAOD. We address three critical questions: (a) are FG–BG associations encoded during training, (b) do they causally affect detection performance, and (c) how do they impact domain adaptation?
To examine FG–BG associations, we analyze class-wise and feature-wise performance degradation using background masking and feature perturbation, quantified as drop rates in accuracy. We further probe causal effects using do-calculus guided by class activation mapping (CAM) and introduce a novel metric, the domain association gradient, defined as the ratio of drop rate to maximum mean discrepancy (MMD).
Validated across multiple models and datasets, our findings underscore the importance of explicitly addressing context bias in DAOD. By analyzing how FG–BG associations affect generalization, this work illuminates the complex interplay between data and AI and demonstrates how methodological innovations in studying data–AI relationships can inform robust, generalizable AI systems.
Chenlan Wang, Research Fellow, School of Public Health
This work investigates group formation dynamics in contexts characterized by resource pooling, spatial cohesion, and agent heterogeneity, with a particular focus on the collaborations among cross-sector partnerships (CSPs) comprising public, private, and nonprofit organizations. Employing both a game-theoretic model and an agent-based simulation, we model organizations as strategic agents who optimize group associations or collaborations to maximize collective utility while accounting for spatial and resource based constraints. We formally prove the existence of stable group equilibria using an individually stable equilibrium (ISE) concept, ensuring that no agent can unilaterally improve their position by joining another group.
Through extensive agent-based simulations, we analyze how group size and composition are shaped by variations in individual resource availability, geographic dispersion, and resource heterogeneity. Results show that limited individual resources and high spatial constraints restrict group formation to spatially proximate actors, often yielding smaller and less diverse groups. In contrast, when resources are abundant, groups readily form across larger spatial ranges, with increased likelihood of sectoral diversity. We further demonstrate that higher resource heterogeneity and proximity collectively foster larger, more diverse partnerships, reflecting a synergy driven by resource complementarity.
Our findings identify key trade-offs governing group formation: as the potential for resource pooling increases, fewer but larger and more heterogeneous groups emerge, whereas increased geographic dispersion tends to fragment groups. The model and results provide guidance for designing robust cross-sector collaborations and heterogeneous multi-agent systems, with practical implications for fields such as disaster response, collaborative sensing, smart infrastructure, and human–robot teaming. By elucidating the interplay among spatial, organizational, and resource factors, this work advances understanding of the mechanisms underpinning stable group formation and effective partnerships in complex environments.
Joseph Osumeje, Schmidt Science African Faculty Fellow, Michigan Institute for Data & AI in Society
Data-AI synergy is transforming scientific discovery, enabling breakthroughs through the integration of advanced data science methodologies and artificial intelligence (AI). This vision explores recent foundations and frontiers in domain-specific AI models and novel data integration strategies. Multi-agent AI systems, such as Gemini 2.0-based virtual researchers, automate and augment reasoning processes for hypothesis generation, experimental design, and iterative innovation, accelerating both the speed and novelty of discoveries across biomedical, climate, and materials research. Meta-analytic evidence demonstrates that human-AI collaboration significantly enhances creative and synthesis tasks, although high-frequency decision-making workflows require further process optimization. Advances in data science infrastructures, including open data sharing, cloud-based collaboration, and interdisciplinary digital platforms, allow research teams worldwide to leverage heterogeneous, high-dimensional datasets for both hypothesis-driven and data-driven inquiry. Pilot applications in climate modeling, drug repurposing, and antimicrobial resistance illustrate how AI-enabled systems streamline validation, uncover mechanistic insights, and extend analytical capabilities previously inaccessible to human researchers. Despite these advances, critical challenges remain such as algorithmic bias, human AI communication limitations, and insufficient frameworks for quality assessment and reproducibility. Addressing these issues through collaborative networks and trusted evaluation programs is essential. The research roadmap emphasizes expanding experimental designs for human-AI teams, developing advanced evaluation metrics, and establishing standardized, open repositories for benchmarking. By fostering collaboration across institutions, disciplines, and geographies, the synergy of data science and AI can democratize scientific insight, stimulate interdisciplinary innovation, and tackle society’s grand challenges. In conclusion, this vision envisions a new era where collaborative intelligence, methodological innovation, and robust data infrastructures amplify the depth, rigor, and reach of scientific discovery.
Ruta Sharangpani, Research Area Specialist, Medical School
We will use smartwatches and brief surveys to create a digital, mental health “signature” for adolescents, and track these signatures over time. By comparing the signatures to patterns from thousands of previous emergency department (ED) visits for adolescent mental health, we hope to spot warning signs of depression. Our goal is to give patients, families and providers alerts to avoid unnecessary ED visits.
Adolescent ED visits for depression increased over the last decade, with ~ 4.5 million individuals experiencing at least one major depressive episode in the previous year. Leveraging information in unstructured medical notes and using smartwatch data can lead to unique insights in creating precise and personalized mental health digital signatures to predict worsening symptoms and develop interventions that may preclude ED visits.
We created personal physiological digital profiles (PPDP), a data driven signature of an individual’s baseline physical state. We also developed a pipeline using large language models (LLMs) to extract semantic features from unstructured medical notes.
In this study, we will develop a proof-of-concept personalized mental health signatures for adolescents. We will prospectively enroll 50 adolescents with depression, and 50 matched controls in the ED, to collect physiologic parameters, and periodic surveys of mental health symptoms to build PPDPs. Simultaneously, we will use artificial intelligence and LLMs to extract depression specific language elements from unstructured notes of ~250,000 adolescents ED visits over ten years. We will then combine PPDP data and discovered features, to train and validate a machine learning model to predict first ED visit and 30-day ED revisit for depression.
The early detection of deterioration of mental health in adolescents would potentially accurately and safely identify patients who may benefit from digital interventions. This may lead to increase access to care and decrease unnecessary ED visits.
Kennon Stewart, Graduate Student, College of Literature, Science, and the Arts
Machine unlearning work assumes a static, i.i.d training environment that doesn’t truly exist. Modern ML pipelines need to learn, unlearn, and predict continuously on production streams of data. We translate the notion of the batch unlearning scenario to the online setting using notions of regret, sample complexity, and deletion capacity. We further tighten regret bounds to a logarithmic O(ln T), a first for a machine unlearning algorithm. And we swap out an expensive Hessian inversion with online variant of L-BFGS optimization, removing a memory footprint that scales linearly with time. Such changes extend the lifespan of an ML model before expensive retraining, making for a more efficient unlearning process.
Questions? Contact Us.
Message the MIDAS team: [email protected]