Bringing together national leaders from academia, NIH, industry, and healthcare to advance trustworthy AI in learning health systems.
Note: Agenda is subject to change.
| Time | Program |
|---|---|
7:30–8:30 AM | Registration, Networking Breakfast & Exhibits |
8:45–9:00 AM | Welcome & Opening Remarks |
9:00–10:15 AM | Session I: Building the AI-Enabled Learning Health System |
| Carlos Bustamante, Vice Dean of Research, OU College of Medicine - Enabling Precision Health at Scale for All | |
| Paul Brooks, Joness Chair in Data Science and Analytics, OU School of Industrial and Systems Engineering - Automated Discovery in Biomedical Data | |
| Erik Holbrook, Chief Clinical Informatics Fellow, Mass General Brigham, Primary Care Physician, Brigham and Women’s Hospital - Clinical Documentation with AI | |
| Aaron Cohen, Flatiron Health, Inc., Adjunct Assistant Professor of Medicine, NYU Langone School of Medicine - From the EHR to Digital Twins: Foundational Concepts for Advancing Cancer Research | |
| 9:50–10:15 AM | Session II: Implementing the AI-Enabled Learning Health System, Fireside Chat & Panel Discussion Moderator: Ian Dunn (Dean, OU College of Medicine) |
10:15–10:30 AM | Break |
10:30–12:00 PM | Session III: National Perspectives on AI Infrastructure and Research NIH leaders discuss the national strategy, infrastructure, and opportunities shaping the future of AI-enabled biomedical research. |
| Susan Gregurick, Director of the Office of Data Science Strategy (ODSS) - NIH AI Strategy and Opportunities for Investigators | |
| Sean Mooney, Director, Center for Information Technology - Enterprise AI at NIH | |
12:00–1:00 PM | Networking Lunch, Posters & Industry Demonstrations |
1:00–2:00 PM | Session IV: Cancer Research in the Age of AI How AI is transforming cancer discovery, clinical research, and precision oncology Greg Sawyer |
2:00–2:30 PM | Session V: AI and the Future of Cancer Research and Care Academic, clinical, and industry leaders discuss the opportunities, challenges, and future of AI in oncology. Moderator: Wei Chen (Chair, OU School of Biomedical Engineering) |
2:30–2:45 PM | Break |
| 2:45–3:45 PM | Session V: AI and the Future of Cancer Research and Care |
| Sebastian Nickel & Ankur Kapoor, Siemens Healthineers - AI and Agentic AI in Healthcare Imaging | |
| Richard Kennedy, Professor, OU College of Medicine - The Age-Friendly Learning Health System | |
| David Bard, Chief Research Informatics Officer, OU College of Medicine - Portable AI Prediction Models Using OMOP and the OHDSI Ecosystem | |
| David Kendrick, Chair in Medical Informatics, OU School of Community Medicine, Associate Vice Provost for Strategic Planning, OU Health Campus - AI, Health Information Exchange, and the Future of Learning Health Systems | |
3:45–4:30 PM | Session VII. Leading the AI Transformation of Academic Health Systems Executive leaders explore institutional priorities for advancing AI across research, education, and clinical care. Moderator: Gary Raskob (Senior Vice President and Provost, The University of Oklahoma Health Campus) |
4:30–4:45 PM | Closing Remarks |
Learn more about the speakers HERE
Ideal for organizations seeking meaningful engagement with attendees throughout the symposium.
Designed for organiztions seeking deeper engagement with istitutional leaders and strategic collaborators.
| Name | Title | Abstract |
|---|---|---|
Riya Manandhar
|
PREDICTING OVARIAN CANCERDISEASE TRAJECTORIES THROUGH LONGITUDINAL RAMAN SPECTRAL SIGNATURES OFASCITES SAMPLES
| Introduction: Ovarian cancer remains a leading cause of gynecological cancer-related mortality, largely attributable to late-stage diagnosis, intra-tumoral heterogeneity, and therapeutic resistance. Malignant ascites constitutes a dynamically evolving tumor microenvironment whose biochemical architecture and longitudinal treatment response remain incompletely characterized, limiting effective monitoring strategies for disease progression and recurrence prediction. Methods: Raman spectroscopy was applied to cellular and fluid fractions of longitudinal ascites samples collected from ovarian cancer patients across multiple treatment stages. A multi-layered analytical pipeline integrating unsupervised dimensionality reduction, network-based co-variation analysis, supervised regression and classification, and temporal probabilistic modeling was implemented to characterize, track, and predict treatment-induced biochemical changes at both cohort and individual patient levels. Results: Pre-treatment ascites exhibited intrinsic, patient-specific biochemical heterogeneity organized into discrete, hub-coordinated network modules spanning protein, lipid, and nucleic acid pathways, demonstrating that the ascitic microenvironment is molecularly structured rather than random. Longitudinal profiling identified two divergent treatment-induced remodeling trajectories: a heterogenic trajectory characterized by biochemical diversification, and a homogenic trajectory marked by convergence toward a treatment-adapted phenotype. Support Vector Regression achieved near-perfect treatment-stage classification (AUC up to 1.00) and predicted an unseen treatment stage with 81.5% spectral similarity. Discrimination between initial and recurrent samples exceeded 83% accuracy across two independent classifier architectures. Hidden Markov Models captured individual-level temporally structured biochemical state trajectories with Pearson correlations up to 0.97. Translational Impact: These findings establish malignant ascites as a dynamically evolving, predictably structured environment whose treatment-induced remodeling is detectable, quantifiable, and predictable through label-free Raman spectroscopy combined with computational modeling, supporting the development of personalized spectroscopic biomarkers for longitudinal monitoring, recurrence surveillance, and precision oncology applications. |
Sam Moore
| From Data to Decisions: Machine Learning, Athlete Availability, and ComprehensiveMonitoring in NCAA Women's Sports | Athlete Availability (AA; %) is the capacity for unrestricted participation in training and competition and is often considered the most important factor for team and individual success. However, with the rise of athlete monitoring applications, the data-overwhelm currently reported by coaches and practitioners has severely limited implementation of data-driven decision-making. Additionally, prior data show sex-divergent outcomes regarding injury, return-to-play, and subjective and objective readiness responses. This study is novel in its collaborative academic-athletic approach, using a machine learning variable selection model for analysis across a wider breadth of applied data streams, providing greater utility of findings to interdisciplinary practitioners working with female athletes. This study used the elastic net regression to identify influential predictors of AA related to subjective wellness, training load, force profile, body composition, and objective recovery, in a multiteam cohort of elite female athletes. Goodness of fit was described with root mean squared error (RMSE). RMSEs for combined, lacrosse, and soccer models were 17.8, 8.9, and 17.8%, respectively. Influential predictors of training load, recovery, and wellness variables were selected in all models, with differing impact between teams. Single-team analyses demonstrated inconsistent predictors; most notable contrasts included no force profile variables selected for lacrosse AA, and no body composition variables selected for soccer AA. Differences in RMSEs and predictors between teams suggest single-team models as a stronger approach to understanding key outcomes for AA. These findings allow for the refinement of testing batteries, honing of meaningful predictors to prioritize data streams, and future modeling algorithms to identify athletes at risk of low AA prior to injury. This approach, grounded in cross-campus collaboration, offers a dynamic infrastructure for researchers and practitioners at institutions of varied resource levels for meaningful innovation to improve female athlete health and performance, while also making more effective use of data already being collected. |
Qinghao Zhang
| Deep-learning-Assisted Photoacoustic and Ultrasound Evaluation for Pre-transplant Human Liver Graft Quality and Transplant Suitability | End-stage liver disease (ESLD) is one of the leading causes of death worldwide. Currently, the only curative option for patients with ESLD is liver transplantation. However, the demand for donor livers far exceeds the available supply, partly because many potentially viable livers are discarded following biopsy evaluation. While biopsy is the gold standard for assessing liver histological features related to graft quality and transplant suitability, it often leads to high discard rates due to its susceptibility to sampling errors and limited spatial coverage. Besides, biopsy is invasive, time-consuming, and unavailable in clinical facilities with limited resources. Here, we present an AI-assisted photoacoustic/ultrasound (PA/US) imaging framework for quantitative assessment of human donor liver graft quality and transplant suitablity at the whole-organ scale. With multimodal volumetric PA/US images as the input, our deep-learning (DL) model accurately predicted the risk level of fibrosis and steatosis, which indicate the graft quality and transplant suitability, when comparing with true pathological scores. DL also identified the imaging modes (PAI wavelength and B-mode USI) that correlated the most with prediction accuracy, without relying on ill-posed spectral unmixing. Our method was evaluated in six discarded human donor livers comprising sixty spatially matched regions of interest. Our study will pave the way for a new standard of care in organ graft quality and transplant suitability that is fast, noninvasive, and spatially thorough to prevent unnecessary organ discards in liver transplantation. |
Doga Demirel
|
Behavior-Tree–ConstrainedLarge Language Model Agents Achieve Complete Protocol Adherence as AI TeamMembers in Emergency Surgical Simulation
| Introduction Team-based simulation of operating room emergencies depends upon trained confederates or scripted virtual characters. Confederates increase scheduling and personnel costs; scripted characters cannot respond naturally to unanticipated learner decisions. Large language models (LLMs) enable adaptive interaction, but may drift from assigned clinical roles and omit required procedural protocols. We developed and evaluated context-aware AI surgical team members that pair natural language interaction with explicit clinical guardrails in a virtual operating room. Methods We implemented an AI circulating nurse and anesthesiologist that interact with a human surgeon in an intra-abdominal bleeding scenario. Each agent pairs a locally deployed LLM with a behavior-tree(BT) controller. The LLM receives a role-specific prompt, dialogue history, patient physiology, simulation events, the current procedural goal, and a restricted action set; it generates dialogue and selects tool-mediated actions that update the environment. BT constrains scenario progression, assigns role responsibilities, and blocks omission of required safety steps. We ran 50 autonomous trials with BT constraints and 50 without. Outcomes were completion of the surgical time-out, conversion to open surgery, and abdominal packing; all actions were time-stamped. Groups were compared by Fisher exact test. Results Constrained agents completed the required workflow in 50/50 runs (100%). Unconstrained agents completed the surgical time-out in 0/50 (0%), converted to open surgery in 35/50 (70%), and completed abdominal packing in 40/50 (80%); all comparisons favored the constrained architecture (p ≤ 0.001). Constrained agents retained variability in dialogue and mitigation sequencing, indicating that adherence was not obtained by collapsing to a fixed script. Time-stamped logs supported retrospective non-technical skills assessment. Conclusions Coupling BT with LLM-generated interaction produced AI surgical team members that adhered to required clinical protocols in every trial while retaining flexible dialogue and decision pathways. Unconstrained LLM agents were unreliable, failing the surgical time-out entirely. This architecture supports responsive AI team members in team-based surgical simulation and reduces dependence on human confederates. |
James Cheng
|
Quantifying Benefits of a Combined REDCap-on-FHIR +EMERSE Semi-Automated ChartAbstraction Procedure for Clinical Studies
| Introduction Fast Healthcare Interoperability Resources (FHIR) standards can improve data quality and reduce manual effort in research workflows that rely on electronic health record (EHR) data. The University of Oklahoma implemented two FHIR-enabled tools—REDCap Clinical Data Interoperability Services (CDIS) and EMERSE—to improve the accessibility and efficiency of EHR data use in clinical research. We evaluated the impact of FHIR-based automation on abstraction workload, accuracy, and implementation feasibility. Methods We conducted a crossover study of 50 patient charts comparing traditional manual abstraction with a semi-automated REDCap CDIS workflow. Two research assistants (RAs) abstracted randomly assigned charts using one method per chart. For manual abstraction, structured EHR data were reviewed and entered into case report forms (CRFs). In the semi-automated workflow, REDCap CDIS used FHIR-based APIs to import structured data, including demographics, diagnoses, and treatment information, directly into REDCap. Results FHIR-enabled automation substantially reduced CRF completion time. Mean completion time decreased from **5:24 minutes manually to 1:38 minutes with CDIS**. RA1 improved from 6:03 to 1:23, while RA2 improved from 3:51 to 1:56. These results demonstrate meaningful reductions in manual workload and more efficient access to structured clinical data. Conclusion REDCap CDIS provides a scalable, standards-based approach to reducing EHR abstraction burden and supporting data quality. EMERSE complements this workflow by enabling retrieval of unstructured clinical notes. Together, these tools offer a practical framework for FHIR-enabled research automation, with potential applications in longitudinal studies and multi-site collaborations. |
Arash Ahmadi
|
Rubric- GuidedReinforcement Learning Improves Cardiovascular Question Answering in aLocally Deployable Medical Language Model
| Large language models could support cardiovascular question answering for patients and clinicians, but deployment in health systems is constrained by privacy, cost, and reliability: cloud-scale models transmit clinical text to third parties, and supervised fine-tuning compresses multi-dimensional clinical quality into a single imitation target. We post-trained an open 14-billion-parameter model (Qwen3-14B) with Group Relative Policy Optimization (GRPO) against physician-style rubrics. Training prompts were filtered to heart-related queries from the RaR-Medicine corpus, and an LLM judge scored each rubric criterion independently. We introduce two variance-aware reward functions, Hybrid and Complexity-aware, that preserve partial credit and scale with rubric size. These rewards replace the sparse binary aggregation that destabilizes GRPO. The model was trained and served entirely on one workstation GPU (NVIDIA RTX 6000 PRO), which supports fully local, privacy-preserving inference. On a held-out heart-related HealthBench subset (n=500 prompts, physician-written rubrics), accuracy improved from 0.362 to 0.502 and F1 from 0.532 to 0.668 (+38.7% and +25.7% relative). For comparison, GPT-OSS-120B scored 0.508 accuracy on the same subset with roughly eight times more parameters. The two rubric-aggregation strategies from the original Rubrics-as-Rewards framework improved the same base model far less, by +9.4% and +13.8% relative accuracy, and paired McNemar tests confirmed our rewards outperformed both (p<0.001). An independent practicing clinician reviewed 56 held-out responses across nine cardiac strata: 60.7% were clinically acceptable, 76.8% acceptable or with only minor issues, and 1.8% unsafe. Patient communication scored highest (4.05/5) and context awareness lowest (2.64/5), and the most common failure mode was failure to solicit clinical history. Variance-aware rubric rewards make reinforcement learning practical for focused clinical domains and deliver competitive cardiovascular question answering on hardware available in an academic lab. Clinician-rated acceptability exceeded the automated benchmark score. |
Paul Okafor
|
Predicting Age-Related Blood–Brain BarrierDisruption in p16-3MR Mice Using a Block-Wise Missing Data Framework
| Aging is a major risk factor for neurological disorders and is associated with vascular dysfunction, blood-brain barrier (BBB) disruption, and cellular senescence. Modeling vascular aging requires integrating heterogeneous biological and physiological data, which is challenging when datasets are small and contain substantial structured missingness. In this study, we developed a computational framework to predict vascular biological age in p16-3MR transgenic mice using a block-wise missing-data approach designed to accommodate heterogeneous patterns of data availability. The framework integrates neurovascular phenotypes, including BBB permeability, capillary density, and neurovascular coupling, with senescence-related cellular measures. We evaluated the approach across different age-group configurations to assess its ability to distinguish vascular aging patterns across the lifespan. The framework demonstrated promising predictive performance while retaining incomplete records without case-wise deletion or conventional imputation. |
Airi Shinamura
|
Predicting Labor Duration and Understanding the Causal Effects of Labor-ManagementInterventions
| Labor management is critical for hospital operations, particularly for coordinating delivery timing in fetal anomaly cases and reducing overnight deliveries when staffing resources are limited. Although labor progression has been extensively studied, opportunities remain to improve personalized prediction of time to delivery and to better understand how labor-management interventions may influence delivery timing. We studied a range of maternal, clinical, and labor-related characteristics and applied machine learning approaches to predict labor duration. Model interpretation techniques were used to identify influential predictors, and an additional analytical framework was used to explore the potential effects of labor-management interventions occurring over the course of labor. |
Geneva Daniel
|
A Comparative Pilot of Manual vs. LLM-AugmentedAbstraction for Neonatal Echocardiogram Reports
| Extracting data from clinical imaging reports is labor-intensive and error-prone. Large language models offer a promising alternative to supplement manual and rule-based extraction. This study compared manual data extraction, rules-based (regular expression/RegEx), and two LLMs (DeepSeek R21-3V and GPT-OSS 20B) using deidentified free-text echocardiogram reports. The study found that LLMs can reliably extract structured numeric fields once outliers are identified and handled. Interpretive fields are harder to extract consistently due to over-inference from surrounding context. Findings support a near-term use and evaluation of a hybrid workflow where LLMs pre-fill days and humans review flagged fields. |
Arnold Kanagwa
|
Predicting UnplannedHospital Readmissions
| Unplanned hospital readmissions within 30 days are widely recognized as indicators of poor care quality. They increase operational costs, strain hospital capacity, and negatively impact reputation, accreditation, and financial performance particularly under programs like Medicare that impose penalties. These readmissions often reflect gaps in treatment, discharge planning, and follow-up care, making them a critical focus for improving patient outcomes and hospital performance. As of 2023, OU Medical Center reported a hospital-wide 30-day unplanned readmission rate of approximately 15.7%, slightly above the national average of 14.56% (https://hmpmetrics.com/hospital/summary/370093). Between 2021 and 2024, associated penalties totaled over $600,000, underscoring the financial and operational impact of this issue. To address this challenge, the project aimed to develop and deploy a machine learning model to identify patients at the highest risk of readmission within 30 days post-discharge. A classification model was successfully built, trained, and integrated into SQL Server for real-time prediction. The model targets patients aged 18 and older admitted to OU Medical Center between June 1, 2023, and April 30, 2025, with encounters from May 1, 2025, to July 10, 2025 used for deployment testing. Among the models evaluated, the XGBoost classifier delivered the best performance, achieving an AUC score of 0.95 on the test dataset. This high level of accuracy positions the model as a valuable tool for reducing readmissions, improving care quality, and mitigating financial penalties. |
Jalal Saidi
|
Evaluating Retrieval-Augmented Generation and Prompting Strategies Across Open Large Language Models for Medical Question Answering
| Retrieval-augmented generation (RAG) is increas- ingly used to ground large language models (LLMs) in external evidence, a capability that is especially relevant to medical ques- tion answering where factuality, recency, and traceability are im- portant. This study evaluates RAG and prompting configurations across nine open-model settings drawn from the Gemma, Llama, and Qwen families. The available results cover 286 evaluation items and compare a no-RAG baseline and the best RAG configuration under both standard choice prompting and explanation-oriented prompting. Across all nine model settings, the best RAG configuration improved choice accuracy over the no-RAG baseline. Without ex- planation prompts, gains over the no-RAG baseline ranged from 9.09 to 17.49 percentage points; with explanation prompts, gains ranged from 12.24 to 36.36 points. Dense retrieval and a Hybrid- 2 configuration alternated as the strongest retrieval strategies across model sizes. Explanation prompting did not consistently improve final-choice accuracy, suggesting an interaction among model capacity, retrieval strategy, few-shot demonstrations, and reasoning-oriented prompting. These findings motivate system- atic evaluation of retrieval and prompting choices rather than assuming a single RAG recipe generalizes across model families. |
Olajumoke Oladapo
|
Tumor Immune Landscape in Mouse Spleen following Localized Ablative Immunotherapy inPancreatic Cancer Model
| Tumor Immune Landscape in Mouse Spleen following Localized Ablative Immunotherapy in Pancreatic Cancer Model Olajumoke B. Oladapo1, Suryaveer Kapoor1, Marjan Ghanbariabdomaleki1, Trisha I. Valerio1, Coline L. Furrer1, Abigael P. Wiliams1, Wei R chen1,2 and Marmar Moussa1,3* 1Stephenson School of Biomedical Engineering, University of Oklahoma, Norman, Oklahoma 73019, USA 2Stephenson Cancer Center, The University of Oklahoma Health Campus, Oklahoma City, OK 73104, USA 3School of Computer Science, University of Oklahoma, Norman, Oklahoma 73019, USA *Correspondence: marmar.moussa@ou.edu Abstract Pancreatic ductal adenocarcinoma (PDAC) is highly resistant to immunotherapy due to its immunosuppressive tumor microenvironment (TME), characterized by dense fibrosis and limited effector T-cell infiltration [1,2]. Photothermal therapy (PTT) combined with the immunoadjuvant glycated chitosan (GC) represents a promising strategy to enhance anti-tumor immunity [3]. Here, we evaluated the immunological effects of PTT+GC in a metastatic pancreatic cancer mouse model using single-cell RNA sequencing (scRNA-seq) of tumor and spleen tissues. C57BL/6J mice bearing subcutaneous Panc02H7 tumors were treated with PTT (980 nm, 0.85 W/cm2, 10 min) followed by intratumoral administration of 100 µL GC. Tumors and spleens were harvested 13 days post-treatment, dissociated into single-cell suspensions, and analyzed using the 10x Genomics 5′ Gene Expression assay. scRNA-seq analyses identified diverse immune and stromal populations with distinct treatment-associated responses. Two B-cell populations exhibited antigen-presentation features, with one showing a more activated phenotype marked by Cd40 and Icosl. T cells displayed effector/cytotoxic and TCR-signaling features (Nkg7, Ccl5, Zap70) alongside Il7r/Tcf7-associated differentiation. Myeloid cells exhibited antigen-processing/phagolysosomal features (Cd68, Lgals3, Ctss, Tyrobp) and a C1q-expressing population (C1qa-c), while tumor-associated macrophages showed increased Apoe. Stress-associated genes (Hspa8, Dnaja1, Dnajb1, Hsph1, Ddit4, Fkbp4) were globally upregulated. Random Forest analysis of module-scored programs identified stress response as the strongest predictor of treatment, alongside myeloid/neutrophil, antigen-presentation, B-cell receptor, helper T-cell, cytotoxic, activation, and degranulation programs. GO enrichment supported myeloid/leukocyte activation, neutrophil responses, and immune effector functions. PTT+GC induces coordinated innate and adaptive immune activity alongside a prominent cellular stress response, characterized by myeloid/neutrophil activation, antigen-presentation programs, and T-cell effector activity. These findings provide mechanistic insight into the immune response associated with PTT+GC and support further investigation of this approach as an immunomodulatory strategy for PDAC, including its potential integration with immunotherapies. Funding This work was supported by the National Science Foundation [NSF-2341725, NSF-2443386]; National Institutes of Health [NIH-K25CA270079]; and the University of Oklahoma Big Idea Challenge 2.0 award (BIC2.0). Research reported in this publication was also supported in part by the National Institute of General Medical Sciences of the National Institutes of Health under Award Number P20GM162339. References [1] C. J. Halbrook, C. A. Lyssiotis, M. P. Di Magliano, and A. Maitra, “Pancreatic cancer: Advances and challenges,” Cell, vol. 186, no. 8, pp. 1729–1754, 2023. [2] P. Farhangnia, H. Khorramdelazad, H. Nickho, and A.-A. Delbandi, “Current and future immunotherapeutic approaches in pancreatic cancer treatment,” Journal of Hematology & Oncology, vol. 17, no. 1, p. 40, 2024. [3] F. Zhou, J. Yang, Y. Zhang, M. Liu, M. L. Lang, M. Li, and W. R. Chen, “Local phototherapy synergizes with immunoadjuvant for treatment of pancreatic cancer through induced immunogenic tumor vaccine,” Clinical Cancer Research, vol. 24, no. 21, pp |
Faith Morrison
|
Early-, Late-, and Postpartum PreeclampsiaDetection Using Machine Learning and Electronic Healthcare Data
| Preeclampsia affects 2–8% of pregnancies and can lead to serious maternal complications, yet early-onset, late-onset, and postpartum preeclampsia are often analyzed as a single condition despite their clinically distinct timing and risk profiles. In this study, we will develop machine learning models to the electronic health record dataset to evaluate whether separating these preeclampsia subtypes improves prediction and interpretation of their unique risk factors. Patient records are divided according to the clinically relevant time windows for each subtype and represented using both longitudinal and static data. Machine learning models are used to distinguish patients with and without preeclampsia and explore potential differences in risk factors across preeclampsia subtypes. Model performance and feature importance are evaluated using standard predictive and interpretability measures. |
Reza Babaei
|
ViTalMind: Multi-Agent System for Evidence-Grounded Health Claim Verification &Coaching
| Background. Health information on social and video platforms is abundant but frequently contradictory, unverified, and stripped of context, leaving people unable to tell rigorous evidence from marketing or anecdote. Objective. We present ViTal Mind, a health-intelligence platform that uses large language model (LLM) agents to verify health claims, decode food labels, and deliver personalized, evidence-attributed coaching. Methods. ViTal Mind orchestrates eight specialized agents — including claim extraction, adversarial fact-checking, personalized coaching, meal planning, and two retrieval-augmented (RAG) council agents — coordinated through a FastAPI backend and a React frontend. User-submitted video transcripts are decomposed into discrete claims, classified by type, and scored against a seven-tier evidence-quality hierarchy from meta-analysis to anecdote. Each claim is further screened through nine structured lenses — including dose-response gap, cherry-picking, conflict of interest, and population specificity — to surface context often absent from the source. An Expert Council agent retrieves persona-grounded responses from independently indexed vector stores built on published work from named health researchers, while a School Council agent contrasts competing nutritional philosophies. All LLM inputs and outputs pass through a security layer performing prompt-injection sanitization, output filtering, and field-level encryption; every health-advice response carries a medical disclaimer. Results. The system is deployed in production with a dual-model LLM layer and supports claim verification, ingredient-quality label decoding, biomarker tracking, and AI-assisted meal planning within a single authenticated session. Conclusion. By pairing multi-agent orchestration with an explicit evidence-quality and bias-detection framework, ViTal Mind offers a reproducible architecture for transparent, attributable AI health guidance. |
Youla Ali
|
BEAS: An Unsupervised Machine Learning Approach for Scalable Batch Effect Assessment Scoring
| Introduction: Integrating multiple single-cell RNA sequencing datasets is essential for building large-scale cell atlases, yet it introduces technical batch variation from differences in sequencing platforms, reagents, or protocols. Integration methods, from classical dimensionality reduction to machine-learning models, embed cells into shared spaces, but assessing success remains difficult. Existing metrics such as kBET, LISI, and ASW operate at the neighborhood level and show dataset-dependent performance. The best-performing, kBET, assumes every cell population is represented across all batches and heavily penalizes integration when a population is unique to one dataset. Methods: We developed BEAS, a score that assesses integration quality at the population level. For each neighborhood in an integrated dataset, we computed per-dataset centroids as the mean expression vector across contributing cells, aligned them to a common gene space, and calculated pairwise cosine similarities. Each neighborhood's similarity was weighted by the product of contributing cell counts and adjusted by the proportion of datasets represented, penalizing single-dataset-dominated neighborhoods and rewarding broad cross-dataset representation. Per-neighborhood scores were aggregated into one interpretable score. We validated BEAS on semi-simulated datasets with known outcomes and compared it against kBET. Results: Across five integration scenarios spanning high, medium, and low expected quality, BEAS tracked the expected outcomes, scoring high-quality integrations near 0.9 and low ones near 0.4. Unlike kBET, BEAS correctly reflected cases where a population was unique to one dataset. For the lung scenario, kBET scored 0.95 (indicating a high batch effect) while BEAS scored 0.934, correctly matching the expected high integration quality. Translational Impact: BEAS assesses integration quality directly from the output embedding, making it applicable to any approach, including machine-learning models. Method-agnostic evaluation matters as integration pipelines increasingly build cell atlases, which underpin mapping of diseased tissue and identifying disease-relevant cell populations. BEAS is available within a public platform. |
Brandt Wiskur
|
The AI Openness inBehavioral Health Metric: A Pilot Study
| Background: Artificial intelligence (AI) is increasingly integrated into behavioral health, academic medicine, education, and healthcare administration; however, implementation remains uneven. Factors influencing AI receptivity in academic health settings remain incompletely understood, and validated measures of AI readiness are limited, despite their importance for guiding responsible adoption and implementation. Methods: We conducted a pilot psychometric evaluation of the AI Openness in Behavioral Health Metric among 105 respondents. The survey assessed four AI-related domains: Understanding & Use, AI Healthcare Trust, AI-Related Concerns, and Implementation Confidence. The survey included three abbreviated Big Five personality domains (Agreeableness, Neuroticism, and Openness) to evaluate potential personality-related influences on AI attitudes. We calculated mean scale scores for each domain. We assessed internal consistency using Cronbach’s α and McDonald’s ω; descriptive statistics and Pearson correlations characterized scale performance and associations. Results: Understanding & Use (M=4.42, SD=1.28; α=.811, ω=.813) and AI Healthcare Trust (M=4.39, SD=1.49; α=.832, ω=.835) demonstrated good internal consistency. Implementation Confidence showed modest reliability (M=4.78, SD=1.11; α=.605, ω=.599), while AI-Related Concerns showed lower consistency (M=5.43, SD=1.13; α=.432, ω=.605), suggesting potentially distinct concern dimensions. Neuroticism demonstrated acceptable reliability (α=.77, ω=.78), whereas Agreeableness (α=.55, ω=.67) and Openness (α=.45, ω=.46) were less reliable. Conclusions: Findings support the preliminary psychometric viability of Understanding & Use and AI Healthcare Trust while identifying opportunities to refine AI-Related Concerns and Implementation Confidence. Future iterations should revise and expand these domains before validation in broader health-professional populations. Personality measures showed limited associations with AI Healthcare Trust and may provide limited explanatory value. As AI becomes more embedded in medical practice and education, identifying the factors that shape trust, readiness, and implementation will be increasingly important for guiding responsible and effective adoption. |
Registration is now closed.
