Objectives/Goals: To contextualize existing knowledge in translational science, women’s health research can benefit from artificial intelligence (AI) methods for obtaining insights out of massive scientific literature. This study explored the feasibility of making an AI model answer diverse questions based on a set of our institution-specific publications. Methods/Study Population: We developed an automated pipeline to retrieve PubMed publications on women’s health that acknowledged the Mayo Clinic Clinical and Translational Science Award of the past two cycles. The publications need to be indexed with MeSH terms (as a major topic) including “Women’s Health”, “Obstetrics”, “Gynecology”, “Female Urogenital Diseases and Pregnancy Complications”, and “Perinatal Care”, etc. In addition, we enforced the species to be “Human” and the demographics to include “Female.” The metadata of every article including title, abstract, publication year, journal, and author affiliation were saved into a single JSON file. The JSON file was then used as the input context along with a few test prompts submitted to the gemini-2.5-flash model. All the experiments were conducted in a Google Colab environment. Results/Anticipated Results: A total of 237 articles on women’s health were retrieved for the experiment. The AI model correctly summarized the top 5 journals but did not obtain the correct article counts until explicitly requested to heed accuracy. When asked to list three prominent departments and their focus areas, the model gave sensible answers aligned with our knowledge. When asked to identify 10 articles that best represent the T4 stage, the model returned appropriate answers that demonstrated an understanding of the implied emphasis on population health and widespread implementation. Lastly, the model provided well-organized advice with annotated experts and strategies, when asked to recommend collaborators for doing AI research in women’s health. Discussion/Significance of Impact: Based on a literature dataset, the general-purpose AI model delivered impressive results when asked to fulfill tasks that involved information extraction, summarization, reference to external knowledge, and strategy consultation. The ongoing work aims to enhance scalability and accessibility of the tool while adding more rigorous evaluations.

Objectives/Goals: Cerebral palsy (CP) results in impaired selective motor control (SMC), or the ability to move joints independently. This study leverages computer vision algorithms to create a scalable, low-cost, and accessible tool that automatically quantifies SMC behavior in infants born preterm, allowing for earlier screening and intervention for CP. Methods/Study Population: In single-camera videos of 600 15-week-old infants, all of whom have standardized clinical SMC scores available, human pose estimation algorithms will be used to derive 3D joint trajectories and kinematic metrics. These features will be used to train and cross-validate a machine learning model aimed at automating clinical scoring of SMC from single-camera video recordings of infants. Then, this model will be used to automatically score the SMC in longitudinal single-camera videos of 47 preterm infants (collected every two weeks from birth until 16 weeks of age). The SMC score trajectories between infants who eventually received a CP diagnosis and those who did not will be compared to establish an inflection point at which SMC development begins to diverge, establishing an optimal timepoint for intervention. Results/Anticipated Results: Preliminary results show that 3D angles of 22 joints across the infant’s body can be estimated in each frame of a 1 to 3 minute single-camera video. We anticipate that the error between the 3D angle estimated from a custom fitted kinematic model and the 2D angle calculated using human pose estimation measures will be less than 5 degrees for all joints measured. With a dataset of over 1,500 videos from over 600 infants, we expect that the trained machine learning model will achieve a sensitivity and specificity of over 90% in predicting the clinical SMC score. Finally, we expect that the automated, longitudinal SMC score trajectories will begin to significantly deviate in infants with a later CP diagnosis, from those without, at around term age (40 weeks gestational age), based on recent clinical findings. Discussion/Significance of Impact: We provide the first automated, quantitative analysis of SMC from video recordings that can be used within real-world settings such as the NICU, at home, or at daycare. This work will result in a cost-effective, precise, smartphone-accessible tool that empowers earlier, more equitable diagnosis and treatment planning.

Objectives/Goals: To evaluate the translational reliability, reproducibility, diagnostic performance, and subgroup equity of multimodal artificial intelligence (AI) models for dermatology triage across multiple model platforms. Methods/Study Population: Limited access to dermatology expertise delays diagnosis and care, motivating development of multimodal AI systems that integrate clinical images with patient data for triage. We assembled 200 biopsy-confirmed PAD-UFES-20 lesions (melanoma, keratinocyte carcinoma, benign) with paired images and metadata, prioritizing demographic balance. Six multimodal AI models (GPT-5, GPT-5-mini; Gemini 2.5 Pro, Gemini 2.5 Flash; Claude Sonnet-4, Claude Opus-4) analyzed these lesions with identical prompts predicting diagnostic probabilities, triage (urgent vs routine), and rationale. Outcomes included sensitivity, specificity, AUROC, F1, and subgroup equity. Model rationales were reviewed for interpretability, and subset re-prompting tested reproducibility for translational robustness. Results/Anticipated Results: Across six models, sensitivity range was 0.89–1.00, specificity 0.21–0.65, AUROC 0.77–0.87, and F1 scores 0.72–0.81. GPT-5 achieved the most balanced performance (0.92 sensitivity, 0.65 specificity, AUROC 0.87, F1 0.81), while Gemini 2.5 Pro and Flash reached perfect sensitivity but low specificity (0.21–0.25). Claude Sonnet-4 showed near-perfect sensitivity (0.99) but over-called benign cases (0.24 specificity), while Opus-4 had the lowest sensitivity (0.89). Urgent triage aligned with dermatologist biopsy patterns (87–97%), and sensitivity was consistent across sex and skin type (p ≥ 0.29). Subset re-prompting produced similar results, supporting reproducibility. Model rationales reflected dermatologic reasoning, supporting interpretability, and translational readiness. Discussion/Significance of Impact: Multimodal AI models showed balanced diagnostic performance for dermatology triage, with platform-specific trade-offs between sensitivity and specificity. Subgroup equity, interpretable rationales, and subset reproducibility define key elements for reliable translation into dermatology workflows and prospective validation.

Objectives/Goals: This scoping review aims to synthesize current literature on post-deployment monitoring of AI-enabled digital health solutions within clinical practice. Findings identify existing approaches and gaps that inform guidance for post-deployment monitoring in clinical practice. Methods/Study Population: We conducted a scoping review in accordance with PRISMA-ScR guidelines to characterize the current landscape of post-deployment monitoring in healthcare systems. A PubMed search targeted peer-reviewed articles in English published between 2015 and mid-2025, including text or MeSH terms on 1) health system/hospital; 2) artificial intelligence; 3) post-deployment; and 4) evaluation/monitoring. We performed a thematic analysis to identify common challenges, gaps, and opportunities in AI oversight. Additionally, we reviewed guidelines addressing post-deployment AI monitoring. All analyses were conducted using Rayyan.ai and Microsoft Word. Results/Anticipated Results: Among the six studies included after the full-text review, five provide recommendations to ensure transparency, safety, and model performance. These recommendations encompassed monitoring model performance and real-time case report, post-market surveillance, adverse event reporting, end-user training, data standardization and documentation, and interdisciplinary collaboration. One study reports a framework for post-deployment impact grading. Currently, no guidelines addressing post-deployment of AI monitoring in health systems exist. Our findings highlight the urgent need for structured post-deployment processes to ensure AI in healthcare systems is safe, effective, and trustworthy. Discussion/Significance of Impact: The absence of post-deployment guidelines raises concerns. This review underscores the need for interdisciplinary collaboration to establish a post-deployment monitoring process with scientific rigor, scalability, and sustainability that aligns with operational realities.

Objectives/Goals: To identify and characterize real-world AI error points in vascular imaging, and then develop and use curie8, a large language model (LLM)-based data pipeline, to curate a broad, multimodal dataset for COMPASS – an AI stress-testing framework to define safety boundaries for clinical deployment. Methods/Study Population: We analyzed a pilot dataset of CT pulmonary angiography (CTPA) studies processed by a commercial AI tool for pulmonary embolism detection. AI errors were categorized using a radiologist-defined clinical and technical taxonomy to identify factors which affect AI performance. In parallel, curie8 was developed as an institutional LLM pipeline to query the radiology report text database and automatically link corresponding DICOM medical images and metadata. Together, the pilot dataset and taxonomy will inform the design of a full curie8-curated stress-test CTPA real-world dataset across the health system’s nine imaging sites (Figure 1). 2025-10-20 ACTS Fig and Table.pptx [https://somumaryland-my.sharepoint.com/:p:/g/personal/fdoo_som_umaryland_edu/EXPxiwh3cxhHpnWYLglMxK4BaByWvFivJq5p9ho5Csm4GA?e=ZPCrAC] Results/Anticipated Results: Pilot analyses (n=5,923) showed high overall performance of the commercial AI PE detection tool (sensitivity 89.6%, specificity 98.9%, PPV 88.6%, NPV 99.0%, accuracy 98.1%). AI errors were a small fraction, but had characteristic features – such as small embolus false negatives and artifact-driven false positives – influenced by patient and scan characteristics (Table 1) and technical factors (Table 2). These findings will guide thecurie8 pipeline in building a full multimodal vascular imaging dataset for benchmarking AI models under real-world stressors. Expected deliverables include a harmonized stress-test framework for vascular imaging AI safety evaluation. 2025-10-20 ACTS Fig and Table.pptx [https://somumaryland-my.sharepoint.com/:p:/g/personal/fdoo_som_umaryland_edu/EXPxiwh3cxhHpnWYLglMxK4BaByWvFivJq5p9ho5Csm4GA?e=ZPCrAC] Discussion/Significance of Impact: COMPASS combines institutional LLM-driven data curation (curie8) with structured AI error analysis to build standardized stress-testing resources. This framework advances reproducible, regulatory-aligned evaluation of AI safety and reliability in clinical imaging settings.

Objectives/Goals: Endometriotic lesions visualized during diagnostic laparoscopy are heterogenous in their appearance. Our research aims to 1) assess the interobserver variability of surgeons in their description of endometriotic lesions and 2) assess their accuracy for prediction of invasion into underlying tissues. Methods/Study Population: We will recruit at least 3 gynecologic surgeons to review a dataset of a planned minimum of 120 coded abdominopelvic endometriotic lesions that were biopsied during diagnostic laparoscopy and that are linked to pathology results. Surgeons will independently classify each lesion by visual appearance according to standard and established criteria and predict the presence of deep tissue invasion for each lesion using a binary yes/no classification system with pathology as the reference standard. We will analyze interobserver variability using Cohen’s k statistic. For the accuracy of the prediction of lesion invasion, we will use sensitivity, specificity, and +/- predictive values and +/- likelihood ratios. Results/Anticipated Results: We expect that our results for aim 1 will be consistent with prior studies that demonstrate variability in the description of the visual appearance of endometriotic lesions by gynecologic surgeons. We also expect that for aim 2, there will be a discrepancy between surgeon predicted lesion depth invasion and findings on pathologic analysis, highlighting the limitations of visual inspection and interpretation of endometriotic lesion behavior. Discussion/Significance of Impact: Improving our understanding as to the visual perception of endometriotic lesions will allow for a better understanding of the variability of lesion types and will inform ongoing studies for the development of a machine learning system for the detection and classification of endometriosis.

Objectives/Goals: Synthetic data holds potential for inclusion in medical product development pipelines. Therefore, we explored the current regulatory and practice landscape to identify best practices used to ensure, and communicate to key stakeholders, synthetic data quality, relevance, and reliability in regulatory settings. Methods/Study Population: We identified areas in which synthetic data created using generative AI holds the most value for health researchers. We reviewed regulatory documents, published literature, and expert insights to examine how regulators currently use, define, apply, and govern synthetic data created using generative AI. Next, we identified data management tools, best practices, ethical considerations, and regulatory developments that are necessary for generating fit-for-use synthetic datasets. Using this information, we developed a risk-based credibility assessment framework that aligns with current governmental standards that can be useful for users of synthetic data derived from generative AI applications in regulatory settings. Results/Anticipated Results: Synthetic data created using generative AI holds value for health researchers across four key areas: Acting as a 1) privacy-enhancing technology, 2) data science ’sandbox’ for training and exploration, 3) mechanism to navigate legalities around data sharing and/or use, and 4) method to augment underrepresented subgroups in datasets. However, synthetic data raise ethical and legal concerns, particularly regarding privacy, consent, stakeholder engagement, and ownership. Regulators and health technology assessment bodies, including FDA, EMA, MHRA, and Canada’s Drug Agency, are exploring synthetic data to supplement medical datasets, validate external control arms, enhance model performance, and inform regulatory decision-making. Discussion/Significance of Impact: Our work underscores the need to continue cultivating transparent, ethical, and fit-for-use approaches to synthetic data generation using generative AI. Moving forward, effective synthetic data use and development requires a culture of learning and transparency among regulators, end users, and those involved in data generation and exchange.

Objectives/Goals: The objective of this study was to evaluate the performance of multimodal machine learning (ML) models trained to predict differentiated thyroid cancer (DTC) recurrence using clinical data combined with novel natural language processing (NLP) derived features extracted from patient cytopathology and surgical pathology reports. Methods/Study Population: This was a retrospective study of adult thyroid cancer patients treated at an academic medical center. Patients were classified as having cancer recurrence or no recurrence. NLP features were extracted from cytopathology and surgical pathology reports using Term Frequency–Inverse Document Frequency (TF-IDF), latent Dirichlet allocation (LDA), and a zero-shot large language model (LLM) classification. 5 multimodal ML models were trained to predict cancer recurrence utilizing a combination of NLP and LLM features and clinical variables. Model performance was evaluated using area under the receiver operating characteristic curve (ROC-AUC) and precision recall area under the curve (PR-AUC). The top performing model was optimized with a 5-fold cross-validation. Feature importance was calculated. Results/Anticipated Results: 480 patients with differentiated thyroid cancer diagnosed on surgical pathology were included in this study. The baseline model (clinical variables only) had a F1-score of 0.52 and an AUC of 0.53. The optimized gradient boosting model utilizing all features (EMR, LDA, TF-IDF, and LLM) had a F1-score of 0.87 and an AUC of 0.86. Topic words and themes from the patient cytopathology and surgical pathology reports were generated using LDA. Topic themes in cytopathology reports include malignancy, lymph node evaluation, and molecular testing. Topic themes in surgical pathology reports include histologic subtype, orientation of nodule, and intraoperative biopsy. The LDA themes of malignancy and histologic subtype ranked the highest in terms of feature importance. Discussion/Significance of Impact: Multimodal models utilizing novel NLP features derived from unstructured pathology reports may enable improved prediction of recurrence in patients with DTC. Our optimized model demonstrated that 4 of the top 6 highest features were LDA topics. Topic modeling may be a valuable tool to extract relevant information from unstructured clinical notes.

Objectives/Goals: To evaluate the feasibility, acceptability, and usability of DAPHNE, an AI-powered chatbot for health-related social needs screening and referral in pediatric primary care, and to assess feasibility of study procedures to inform a future large-scale trial. Methods/Study Population: This pilot randomized controlled trial will enroll 60 caregivers of children ≤2 years from Nationwide Children’s Hospital pediatric primary care clinics. Participants will be randomized 1:1 to DAPHNE or standard care. The intervention provides private, natural language-based HRSN screening and personalized resource referrals via a secure, HIPAA-compliant mobile/web chatbot. Primary outcomes include feasibility (enrollment ≥70%, retention ≥70%), acceptability (WEQ ≥80%), usability (SUS ≥68), and technical performance (accuracy F1 ≥0.7, latency <3 seconds). Secondary outcomes include caregiver stress, self-efficacy, satisfaction, quality of life, provider workflow integration, study burden, and EHR-derived care utilization. Results/Anticipated Results: We anticipate that feasibility thresholds for enrollment, retention, and procedural adherence will be met. We expect caregivers to rate the chatbot above usability and acceptability thresholds, with high engagement and accurate resource navigation across devices. Provider surveys and interviews are anticipated to show positive perceptions of workflow integration. EHR data will capture referral patterns, visit adherence, and utilization trends. Findings will generate pilot data to refine the chatbot, optimize trial design, and inform comparator selection and clinical outcomes for a subsequent R01-level efficacy study. Discussion/Significance of Impact: This study will provide critical feasibility data on AI-driven HRSN screening in pediatrics. If effective, DAPHNE could reduce stigma, improve caregiver access to community resources, and inform scalable strategies for integrating conversational AI into pediatric primary care.

Objectives/Goals: This project aims to develop a multimodal vision-language deep learning model that can integrate cell microscopy images and sequencing data in retinal cells to help identify novel pathways involved in diseases such as glaucoma and age-related macular degeneration. Methods/Study Population: Retinal cells, derived from human induced pluripotent stem cells, will be cultured in monolayers and divided into control and stressed groups. The cells will be exposed to oxidative and inflammatory stimuli, then sequenced and imaged at various timepoints. Confocal microscopy will capture morphology and marker staining, while single cell RNA sequencing will profile gene expression. A pretrained dual-encoder vision-language model will be fine-tuned with the multimodal data to embed images and omics data into a shared space via self-supervised learning. The model will then be further adapted for downstream tasks, such as predicting transcriptomic changes from cell imaging and vice versa. Results/Anticipated Results: We anticipate that the fine-tuned model will be able accurately align assign stress conditions of the cell from imaging features and transcriptome profiles, as well as predict cell morphology from gene expressions and vice versa. Additionally, we expect it to be able to generate candidate gene networks and predict gene perturbations related to cellular stress response. The performance of the model will be evaluated as a combination of machine learning metrics such as classification accuracy and area under the curve (AUC) and biological metrics such as single-cell clustering quality. Discussion/Significance of Impact: Unlike most other clinical AI applications that utilize single-modal data, this study pioneers a multimodal vision-language model capable of integrating imaging and sequencing data in ophthalmology. The model may uncover new retinal disease pathways and serve as a foundation for diverse downstream tools for the research community.

Objectives/Goals: Artificial Intelligence(AI)and automation are rapidly transforming healthcare, yet their integration into clinical workflows often falls short due to technical, ethical, and organizational challenges. A standardized, quality-driven approach can aid the AI implementation process and ensure value delivery. Methods/Study Population: Trust emerges as the central hurdle, encompassing both patient and provider confidence in AI systems. Patients raise concerns over safety, transparency, and the physician-patient relationship, while providers express apprehension towards algorithmic opacity, data quality, and legal ambiguity. To address these concerns, the Understand, Transform, Sustain (UTS) framework offers a behavior-based, systems-level approach to AI deployment. Developed by Mayo Clinic’s Quality Academy, UTS integrates process improvement (PI) principles across three phases, emphasizing stakeholder engagement, transparency, and patient safety throughout the AI lifecycle. Results/Anticipated Results: The Understand phase identifies inefficiencies by mapping workflows, collecting data, and recognizing areas for improvement, ensuring appropriate priorities are addressed. In the Transform phase, interventions are designed, implemented, and tested through improvement cycles and feedback loops. Data to build algorithms are carefully evaluated to avoid biases, and AI output is assessed for opacity risk to maintain transparency and explain ability. The Sustain phase monitors outcomes and standardizes practices for long-term value. Data audits and automated extraction tools are applied for fidelity and harmonization, promoting scalability and collaboration among organizations. Discussion/Significance of Impact: While keeping human intelligence central to AI integration, UTS ensures sustained efficiency and trust – critical for successful healthcare transformation. This framework positions itself as a catalyst for responsible innovation, aligning technological advancement with clinical priorities and patient protection.

Objectives/Goals: To identify drivers of publication impact and output in surgical AI across countries and assess whether cross-border collaboration, especially between developed and developing nations, is associated with higher citations and knowledge diffusion over 2000–2025. Methods/Study Population: We queried PubMed EDirect for surgical AI articles (2000–2025) and linked Crossref citations. Papers were grouped by collaboration type: single developed, single developing, developed–developed, developed–developing, and developing–developing. Success metrics included citations/paper, total papers, and collaboration rates. Country-level covariates from World Bank (infrastructure, economics, R&D, technology, education) were integrated. Analyses: chi-square(counts), Kruskal–Wallis (citation distributions), ANOVA (means), and Pearson correlations (predictors of citations and production). Results/Anticipated Results: Developed–developing collaborations had the highest citation impact (21.3 citations/paper) vs developed–developed (21.1), single developed (20.1), developing–developing (13.2), and single developing (8.7) (all p<0.001). In developing nations, citation impact correlated most with healthcare infrastructure – hospital beds (r=0.896), physicians/1,000 (r=0.891), and health expenditure (r=0.883) (all p<0.001). Paper production correlated most with economics and R&D—GDP (r=0.991), researchers in R&D (r=0.997), and R&D spend (r=0.989) (all p<0.001). Technology access (mobile subscriptions, r=0.874) and tertiary enrollment (r=0.871) also tracked with citation impact. Discussion/Significance of Impact: Cross-border collaboration, especially developed–developing, maximizes impact and advances equitable knowledge transfer in surgical AI. Policies should fund international partnerships, strengthen infrastructure in developing countries, and build networks to improve global diffusion of innovation.

Objectives/Goals: The purpose of this project is to generate a model identifying environmental and behavioral predictors on STI diagnosis. Using the fitted model, we examined how the effects of these predictor variables vary as a function of age. Understanding age-dependent correlates of STIs can allow us to be more precise when identifying intervention strategies. Methods/Study Population: We examined surveys and EHRs using the All of Us Researcher Workbench. We created two cohorts (STI diagnosis vs. No-STI diagnosis) matched on self-reported demographic variables. We also extracted survey-derived participant environmental indicators such as alcohol use and annual income. Most participants (60%) were classified as White, majority (94%) had health insurance, 53% were employed for wages or self-employed, and 82% went to or graduated from college (N=5873). We conducted a random forest analysis to predict STI diagnosis given the identified variables. To explain machine learning predictions for complex models, like random forests, we used SHAP values. Using this technique, we further explored the relationship between the variables identified as significant to the model and age at STI diagnosis. Results/Anticipated Results: Results of the random forest showed that the internal estimate of prediction was 51.64% accurate, indicating a low accuracy. The mean decreased Gini values for the top 5 contributors to the model were as follows: annual income (276.74943), limited social contact (196.94722), education (166.86742), safety dynamics in the neighborhood (142.62170), and community safety concerns (141.37972). The results of the SHAP analysis suggested that the relationship between low-earning participants and STI diagnosis has a negative relationship with age at diagnosis; in other words, the relationship between income and STI diagnosis decreases with age. Discussion/Significance of Impact: Our environment plays an important role in our health. This project demonstrates that participant-reported environmental variables were associated with STI outcomes in the model. Furthermore, some of these effects are found to be age-dependent. Future research could explore how modifying behavioral or contextual factors may influence STI risk.

Objectives/Goals: This study aims to assess perceptions of children with insulin-dependent diabetes and their guardians on the perceived benefits and challenges on the use of a diabetes-specific generative pretrained transformer to assist with diabetes care in insulin administration. Methods/Study Population: The study involved one-on-one semi-structured interviews with both adolescents aged 14-19 with insulin-dependent diabetes and their guardians. An interview guide was developed utilizing the constructs from the health information technology acceptance model. We have completed 1 interview with a caregiver with a child with diabetes. A total of 10 interviews are planned. All interviews have been or will be recorded utilizing a video conferencing platform and transcribed verbatim. The interview content will then be organized and coded using NVivo software. Thematic analysis was conducted in accordance with analysis guidelines. Results/Anticipated Results: From initial interviews, 3 key themes have arisen regarding the use of artificial intelligence for diabetes care. First, use of large language models has been utilized in their daily life; however, there is hesitancy to utilize these resources for medical advice. A second important theme that surfaced several times was the importance of trust in the source of the artificial intelligence that she would be using since it was for medical care. Another theme was the importance of ease of use of a large language model. This theme was illustrated by being able to share with family and friends since children are often involved with multiple caregivers. It would also have to have an easy-to-use interface, such as pre-programmed personal settings and recognizable picture icons to make her more likely to use it. Discussion/Significance of Impact: Users emphasized trust in AI sources for diabetes care, noting that clear visuals and personalization improve usability. Further interviews may reveal broader perspectives on adoption, challenges, and barriers among adolescents and caregivers with insulin-dependent diabetes.

Objectives/Goals: To quantify between-study heterogeneity in TP53 expression across DLBCL studies using frequentist and Bayesian meta-analytic frameworks; estimate the association between TP53 status and cell-of-origin (GCB vs. non-GCB); and identify methodological and biological sources of variation, using study-level data. Methods/Study Population: Four retrospective study-level datasets (n=457) from France, Canada, Brazil, and the USA were included. TP53 status was primarily assessed by immunohistochemistry (IHC) with study-specific thresholds (20–50% positive cells). Cell-of-origin classification followed the Hans or Choi algorithms as reported by each study. We fit frequentist random effects models with REML (R metafor) and Bayesian hierarchical models (R rstan) Results/Anticipated Results: Frequentist random effects meta-analysis indicated moderate-to-high heterogeneity (I² = 66.7%, p = 0.038) and a pooled odds ratio for GCB among TP53-positive vs. TP53-negative cases of 0.953 (95% CI: 0.682–1.330). The Bayesian hierarchical model estimated P(GCB | TP53+) with a posterior mean of 0.71 (95% CrI: 0.38–0.93); study-level posterior means ranged from 0.37 to 0.53. Both approaches confirm substantial between-study variation attributable to methodological factors (IHC thresholds, COO algorithms) and biological differences Discussion/Significance of Impact: This analysis clarifies the heterogeneity of TP53 expression in lymphoma and shows how methodological and biological variability influence results. Bayesian modeling enhances reliability in small, heterogeneous biomarker syntheses, supporting standardization of TP53 assessment and improving reproducibility in DLBCL.

Objectives/Goals: This is a scoping review of the use of generative AI (GenAI) for qualitative analysis in the health sciences. The primary objectives are to summarize the methods used for qualitative analysis based on the approach (e.g., inductive vs. deductive), metrics for examining GenAI’s performance by approach type, and best practices for improving performance. Methods/Study Population: We searched six databases (PubMed, EMBASE, CINAHL, Scopus, Web of Science, and PsycINFO) using a comprehensive search string tailored to each database. We identified 5,853 unique results, of which 223 were identified as potentially relevant after review of abstracts. We will conduct a full manuscript review for each potentially relevant result to confirm inclusion, including use of GenAI for qualitative analysis of physical or mental health text data where a full-length paper is available. Two study team members will review each manuscript to confirm inclusion prior to data extraction and coding. We will also enter each manuscript into Yale Clarity, a secure GenAI platform powered by multiple large language models (e.g., ChatGPT, Claude), to examine GenAI’s capacity to facilitate manuscript screening and coding. Results/Anticipated Results: We will code included articles for: the research topic area; type of qualitative data analyzed (interview, focus group, medical chart, text-based digital or social media content); type of study (original or secondary data, review); type of qualitative analysis (thematic analysis, content analysis, grounded theory); type of approach (inductive, deductive); steps in analytic process that generative AI was used (initial open coding and sense making, final coding based on established codebook, identification of overarching themes and narratives); GenAI platform(s) and large language model(s) used; types of GenAI prompts used to facilitate analyses; processes and measures/metrics to evaluate the accuracy and quality of GenAI’s results; outcomes and methods used to increase GenAI’s accuracy and/or quality. Discussion/Significance of Impact: Qualitative research is essential to understand patient perspectives and increase treatment effectiveness and accessibility for all. To actualize GenAI’s potential to facilitate rapid and large-scale qualitative analysis, we first need to understand its strengths, weaknesses, and best practices for maximizing GenAI’s accuracy and quality.

Objectives/Goals: To benchmark the extent to which variability in automated WMH quantification influences stroke and dementia risk stratification by characterizing differences in volume estimates and spatial concordance across three segmentation tools (BIANCA, SAMSEG, and nnU-Net). Methods/Study Population: This retrospective, cross-sectional study will analyze de-identified clinical brain MRI scans from patients aged ≥50 years, including those with and without prior stroke, obtained from institutional imaging repositories. Eligible scans include co-registered T1-weighted and FLAIR sequences. Scans with poor image quality and scans from patients with multiple sclerosis, other demyelinating disorders, or brain tumors will be excluded. WMHs will be segmented using three tools: BIANCA (supervised k-NN classifier), SAMSEG (unsupervised Bayesian model), and nnU-Net (self-configuring deep learning). Agreement in WMH burden will be assessed with intraclass correlation coefficients (ICC[3,1]) and spatial concordance with Dice similarity coefficients (DSC). Bland-Altman analyses will evaluate bias. Results/Anticipated Results: We will analyze 600 brain MRI scans. All three tools are expected to detect similar WMH spatial distributions, though total WMH volume estimates may differ. nnU-Net may yield higher WMH burden estimates than BIANCA or SAMSEG. WMH burden will be neuroradiologist-adjudicated to establish the gold standard, against which concordance for each tool and inter-tool agreement will be assessed using ICC[3,1] and DSC. We anticipate moderate-to-high agreement, but tool-specific variability could influence reproducibility and interpretation of WMH burden for future WMH-related research studies. This study will quantify the extent of variability among tools in clinical brain MRI scans and evaluate whether these differences could meaningfully alter future stroke risk classification thresholds. Discussion/Significance of Impact: This study will benchmark automated segmentation tools for WMH-related research by standardizing WMH quantification to improve reproducibility and reliability for stroke and dementia risk stratification. Establishing tool-specific benchmarks may also guide integration of automated WMH measures into clinical decision-support systems.

Objectives/Goals: To validate 20 culturally tailored clinical vignettes for a translational study assessing ChatGPT’s diagnostic support in adolescent and young adult mental health. The validation ensures methodological, ethical, and cultural rigor to examine ChatGPT’s potential in collaborative clinical decision-making. Methods/Study Population: Twelve licensed clinical psychologists evaluated 50 culturally adapted vignettes using a nine-item instrument assessing diagnostic agreement (Fleiss κ), content validity (I-CVI, S-CVI/Ave), comorbidity sensitivity, difficulty, ethical sensitivity, and perceived risk. Qualitative feedback refined clarity and realism. Vignettes meeting κ ≥ 0.60, S-CVI ≥ 0.80, and ethics ≥ 3 were retained, ensuring one easy, two moderate, and one difficult case per condition. The resulting 20 validated cases will serve as standardized stimuli to assess ChatGPT’s diagnostic accuracy and its ability to collaborate effectively with clinical psychologists. Results/Anticipated Results: High inter-rater reliability (κ ≥ 0.60) and strong content validity (S-CVI ≥ 0.85) are expected. Balanced difficulty and accurate risk detection will confirm calibration. Qualitative feedback will enhance realism and cultural appropriateness. Validated vignettes will enable rigorous assessment of ChatGPT’s diagnostic accuracy and clinical decision-support potential within a translational research framework. Vignette cases will be presented to three groups: 1) ChatGPT app, 2) Clinical mental health care providers, and 3) Clinical mental health care providers collaborating with ChatGPT. Each group will provide a diagnosis and confidence level to compare accuracy, sensitivity, and collaborative value across human-AI and human-only conditions. Discussion/Significance of Impact: This validation provides a rigorous foundation for evaluating ChatGPT’s translational role in mental health diagnostics. Psychometrically and ethically sound vignettes strengthen evidence-based integration of AI tools into precision mental health care and promote responsible human-AI collaboration.

Objectives/Goals: In patients with paraesophageal hernia (PEH), evaluate associations between CT radiomic features and clinical data with (1) undergoing elective repair and (2) volvulus/gastric outlet obstruction and develop an ML pipeline for automated feature extraction and prognosis estimation. Methods/Study Population: Retrospective case–control of adults (>18) at Penn Medicine (2017–2025) with hiatal hernia ICD-9/10 codes and CT chest or CT abdomen/pelvis from clinical care. Radiology reports, H&P, and operative notes will be used to identify those with PEH. We will use traditional statistics and classical ML methods (linear/logistic regression, decision trees, random forest) to test associations between radiomic features (hiatal defect diameter (HDD), hernia sac volume (HSV), herniated gastric volume (HGV)) and outcomes. We will explore CNN-based models – nnU-Net (e.g., TotalSegmentator) and ResNet-based models (e.g., Merlin) – and foundation models (MedSAM2, VISTA-3D) to build the pipeline. Results/Anticipated Results: We hypothesize that CT-derived radiomic features will be associated with (1) undergoing elective repair and (2) developing acute volvulus/GOO. Greater HDD, HSV, and HGV are expected to be associated with increased odds of both outcomes. The ML pipeline is expected to provide organ segmentation and reproducible automated feature extraction using state-of-the-art architectures, and to capture synergistic effects between radiographic and clinical variables to inform patient-level prognosis. Discussion/Significance of Impact: In asymptomatic/minimally symptomatic patients, PEH care often defaults to watchful waiting; candidacy and timing for elective repair are uncertain. Our ML pipeline integrates clinical and CT data to standardize preop radiomics, estimate prognosis, and inform if/when to operate.

Objectives/Goals: To identify real-world type 2 diabetes (T2D) treatment pathways and evaluate whether AI models trained on electronic health record (EHR) medication histories can improve prediction of glycemic outcomes in routine clinical practice. Methods/Study Population: We conducted a retrospective cohort study using de-identified EHR data from the TriNetX Research Network (>3 M adults with T2D, 2019–2024). Medication, laboratory, and demographic data were standardized in PostgreSQL on the University of Utah CHPC. Patients with T1D, pregnancy, or incomplete follow-up were excluded. Longitudinal medication sequences were clustered using hierarchical agglomerative methods to identify treatment trajectories. Transformer-based models incorporating full prescription histories were trained to predict 6-month HbA1c, and performance (RMSE) was assessed overall and across demographic groups. Results/Anticipated Results: Among >3 M adults with T2D, we identified 40 distinct treatment clusters representing mono-, dual-, complex-, GLP-1–based, and variant pathways. GLP-1/GIP agents were typically used as 2nd–3rd-line therapy after metformin or SGLT2i and produced the largest BMI and HbA1c improvements. Baseline XGBoost models using EHR features predicted 6-month HbA1c with RMSE ≈ 0.72. The last HbA1c value was the most predictive feature, though additional labs and demographics modestly improved accuracy. Ongoing work is developing transformer models to capture longitudinal patterns and medication history effects. Discussion/Significance of Impact: National EHR networks can support development of trustworthy, fairness-aware AI to model diabetes pharmacotherapy. Insights reveal disparities in advanced therapy uptake and inform equitable decision-support design.