Objectives/Goals: This work aims to identify functional brain networks that differentiate opioid use disorder (OUD) subjects from healthy controls (HC) using machine learning (ML) analysis of resting-state fMRI (rs-fMRI). We investigate the default mode network (DMN), salience network (SN), and executive control network (ECN), as well as demographic features. Methods/Study Population: This work uses high-resolution rs-fMRI data from a National Institute on Drug Abuse study (IRB #HM20023630) with 31 OUD and 45 HC subjects. We extract rs-fMRI blood oxygenation level-dependent (BOLD) features from the DMN, SN, and ECN. The Boruta ML algorithm identifies statistically significant features and brain activity mapping visualizes regions of heightened neural activity for OUD. We conduct fivefold cross-validation classification experiments (OUD vs. HC) to assess the discriminative power of functional network features with and without incorporating demographic features. Demographic features are ranked based on ML classification importance. Follow-up Boruta analysis is performed to study the medial prefrontal cortex (mPFC), posterior cingulate cortex, and temporoparietal junctions in the DMN. Results/Anticipated Results: Boruta ML analysis identifies the DMN as the most salient functional network for differentiating OUD from HC, with 33% of DMN features found significant (p < 0.05), compared to 10% and 0% for the SN and ECN, respectively. The Boruta ML algorithm identifies age and education as the most significant demographic features. Brain activity mapping shows heightened neural activity in the DMN for OUD. The DMN exhibits the greatest discriminative power, with a mean AUC of 69.74%, compared to 47.14% and 54.15% for the SN and ECN, respectively. Fusing DMN BOLD features with the most important demographic features improves the mean AUC to 80.91% and the F1 score to 73.97%. Follow-up Boruta analysis highlights the mPFC as the most important functional hub within the DMN, with 65% significant features. Discussion/Significance of Impact: Our study enhances the understanding of OUD neurobiology, identifying the DMN as the most significant network using ML rs-fMRI BOLD feature analysis. Ethnicity, education, and age rank are the most important demographic features and the mPFC emerges as a key functional hub for OUD. Future research can build on these findings to inform treatment of OUD.

Objectives/Goals: Aspiration causes or aggravates lung diseases. While bedside swallow evaluations are not sensitive/specific, gold standard tests for aspiration are invasive, uncomfortable, expose patients to radiation, and are resource intensive. We propose the development and validation of an AI model that analyzes voice to noninvasively predict aspiration. Methods/Study Population: Retrospectively recorded [i] phonations from 163 unique ENT patients were analyzed for acoustic features including jitter, shimmer, harmonic to noise ratio (HNR), etc. Patients were classified into three groups: aspirators (Penetration-Aspiration Scale, PAS 6–8), probable (PAS 3–5), and non-aspirators (PAS 1–2) based on video fluoroscopic swallow (VFSS) findings. Multivariate analysis evaluated patient demographics, history of head and neck surgery, radiation, neurological illness, obstructive sleep apnea, esophageal disease, body mass index, and vocal cord dysfunction. Supervised machine learning using five folds cross-validated neural additive network modelling (NAM) was performed on the phonations of aspirator versus non-aspirators. The model was then validated using an independent, external database. Results/Anticipated Results: Aspirators were found to have quantifiably worse quality of sound with higher jitter and shimmer but lower harmonics noise ratio. NAM modeling classified aspirators and non-aspirators as distinct groups (aspirator NAM risk score 0.528+0.2478 (mean + std) vs. non-aspirator (control) risk score of 0.252+0.241 (mean + std); p Discussion/Significance of Impact: We report the use of voice as a novel, noninvasive biomarker to detect aspiration risk using machine learning techniques. This tool has the potential to be used for the safe and early detection of aspiration in a variety of clinical settings including intensive care units, wards, outpatient clinics, and remote monitoring.

Objectives/Goals: Current popular scoring systems for evaluating facial nerve function are subjective and imprecise. This study aims to quantify speech and facial motor changes in patients suffering from facial palsy after cerebellopontine angle (CPA) tumor resection to lay the foundation for a scoring algorithm that is higher resolution and more objective. Methods/Study Population: We will obtain audio and video recordings from 20 adult patients prior to and after surgical resection of unilateral CPA tumors between October 2024 and February 2025. We will obtain preoperative recordings within two weeks prior to surgery and postoperative recordings following a preset schedule starting from the day after surgery up to one year. Audio recordings entail patient readings of standardized passages and phonations while video recordings entail patient performance of standardized facial expressions. We will analyze video data for key distance measurements, such as eye opening and wrinkle size, using DynaFace. We will process audio data using VoiceLab to extract metrics such as prominence and tonality. We will perform statistical tests such as t-tests and ANOVA to elucidate changes across time. Results/Anticipated Results: I expect to obtain 9 sets of audio and video recordings from each of the 20 participants. In terms of speech, I expect average speech duration to increase postoperatively. Similarly, I expect to find increases in time spent breathing, number of breaths taken, and mean breathing duration. In terms of facial movement, I expect nasolabial fold length to decrease postoperatively, as well as eye opening size and left-right symmetry at rest. For both audio and video, I expect these changes to revert towards their preoperative baseline as time passes. I also expect average House-Brackmann and Sunnybrook facial grading scores to increase postoperatively and then decrease with time, correlating strongly with the video and audio findings. I will use trajectory analysis and time point matching to handle any missing data. Discussion/Significance of Impact: This study will validate our analysis platform’s ability to automatically quantify measurable changes that occur to speech and facial movement which correlate strongly with existing scoring systems. Future work will synthesize these data streams to move towards establishing biomarkers for facial nerve function that aid clinical decision-making.

Objectives/Goals: Upon diagnosis, patients with acute myeloid leukemia (AML) have significant information needs. Given its recent increase in popularity, patients may use ChatGPT to access information about AML. We will examine the quality, reliability, and readability of information that ChatGPT provides in response to frequently asked questions (FAQs) about AML. Methods/Study Population: From FAQs on the top 3 patient-facing websites about AML, we derived 26 questions, written in lay terms, about AML diagnosis, treatment, prognosis, and functional impact. We queried ChatGPT-4o on 10/14/2024 using a new Google account with no prior history. We asked each question in a separate chat window once, verbatim, and without prompt engineering. After calibration, 5 oncologists independently reviewed ChatGPT responses. We assessed quality via the Global Quality Scale (GQS), scored from 1 (poor) to 5 (excellent) based on flow, topic coverage, and usefulness. For reliability, we assessed whether each response addresses the query and is factually accurate, elaborating on specific inaccuracies. For readability, we assessed Flesch-Kincaid Grade Level, Gunning Fog Index, and Simple Measure of Gobbledygook. Results/Anticipated Results: This will be a descriptive analysis of ChatGPT responses. For quality and reliability assessments, we will report Fleiss’ kappa for inter-rater reliability and expect substantial agreement or greater (≥0.61). Per prior studies in other domains, we hypothesize that ChatGPT responses will have good quality on average (i.e., GQS score near 4). We hypothesize that nearly all responses will address their query and will mostly be accurate; a minority of responses may have partial inaccuracies. Finally, we hypothesize that readability metrics will suggest that a higher educational level (e.g., college-level education) is required for comprehension. Overall, these findings will help elucidate strengths and limitations of ChatGPT for AML and guide discussion of factors patients should be aware of when using ChatGPT. Discussion/Significance of Impact: No prior study has examined the educational quality of ChatGPT for AML. Our study will detail whether patients are receiving trustworthy and meaningful information, identify misinformation, and provide guidance to oncologists when recommending information resources to patients or fielding questions that patients may raise after using ChatGPT.

Objectives/Goals: Translational researchers spend significant amounts of time finding available datasets and other research data resources for their purposes. Objectives of this program are develop and evaluate a multipronged approach to supporting researchers with existing data resources. Methods/Study Population: We established a dedicated service with expertise in data resources to increase awareness, understanding, and utilization of existing data resources. This program assists investigators and trainees discover appropriate data resources, formulate scientific problems in computable formats, advise on state-of-the-art data analytics, data management, build collaborations, mentor data users, and develop a service pipeline for streamlined data resource project management. This is accomplished through these essential functions: (1) Discover, catalog, document, and manage metadata resources, (2) train and present data resources to the research community, (3) provide individual consultations, and (4) explore and assess novel data resources. Results/Anticipated Results: In a phased approach, the data navigation program is performing outreach to the research community and integrating with existing data efforts on campus, presenting and demonstrating existing data resources, established a consultation service, and building core competencies into long-term usage and navigation of resources across campus. Evaluating the program monthly has shown an increase in various metrics for evaluating commitment and engagement including number of requests for access to data resource, consultations, publications and presentations, co-authorship, and proposals. Unawareness and inappropriate use of data resources leads to delays in performing research and potentially unnecessary duplications of efforts. Discussion/Significance of Impact: Our data navigation program has increased use of data resources in research. Next steps are to continue evaluation and further streamline informatics approaches to data discovery, abstraction, formulation, and analysis. Harmonized data resource programs are important translational science approach to foster the next generation of research.

Objectives/Goals: This study examines associations between childhood violence exposure, accelerated biological aging, and adolescent depression using DNA methylation-based epigenetic clocks. Findings aim to identify biomarkers for early detection, guide interventions, and address youth mental health disparities. Methods/Study Population:Data from the Future of Families and Child Wellbeing Study (N = 4,898), a diverse urban U.S. cohort, were analyzed. Childhood violence exposure, assessed using the Parent-Child Conflict Tactics Scale, included measures of physical, emotional, and psychological aggression and neglect. Biological aging at age 15 was evaluated using second-generation epigenetic clocks derived from saliva DNA methylation patterns, while depressive symptoms were measured with the CES-D scale. Multiple linear regression models tested associations between early violence exposure, epigenetic aging, and depressive symptoms, adjusting for socioeconomic status, caregiver mental health, and other key covariates. Results/Anticipated Results: Preliminary results suggest that early violence exposure may be linked to accelerated biological aging and depressive symptoms during adolescence, a critical developmental period. Epigenetic clocks offer an objective method for identifying high-risk youth, complementing mental health evaluations. With further validation and participatory action research, these findings could guide the development of biomarkers for longitudinal testing in school-based screenings and community health programs. These tools aim to be accessible, culturally relevant, and tailored to diverse populations, enhancing early detection, informing personalized interventions, and supporting scalable clinical applications. Discussion/Significance of Impact: This study explores links between early adversity, biological aging, and mental health, advancing understanding of adolescent depression. Epigenetic biomarkers could improve risk detection and guide tailored interventions in schools and community settings, enhancing access and reducing disparities.

Objectives/Goals: Understanding the interconnections among over 20,000 human diseases spanning organ systems could inform more precise diagnosis and treatment of diseases. Here, we examine whether the ability of large language models (LLMs) to learn universal representations of concepts can be leveraged to discover complex relationships across human diseases. Methods/Study Population: To address the challenge of computationally representing thousands diseases spanning multiple organ systems, we used internal representations of concepts by LLMs to encode diseases based on their descriptions from standard disease ontologies (ICD10 and Phecodes). To do this, we leveraged application programming interfaces (APIs) of three LLMs-GPT3.5, Mistral and Voyage to encode disease relationships. We then performed unsupervised clustering of the diseases using their encodings (embeddings) from each LLM to determine whether the resulting clusters reflect disease relationships. To enable deeper exploration of disease relationships, we developed interactive plots that provide a system level view of the relationships between thousands of diseases and their association with specific organ systems. Results/Anticipated Results: We found that unsupervised analysis of disease relationships using the LLM encodings reveal high similarities among diseases based on organ systems they affect. All the LLMs clustered diseases into groups largely defined by the organ systems they affect without being trained to specifically classify diseases into their corresponding organ system classification. An exception to this was tumors in which we observed that most tumors cluster together as a group irrespective of the organs they affect. Interestingly, we found that tumors affecting anatomically related organs show higher similarity to each other than to those affecting distantly related organs. In addition to anatomical relationships between diseases, we found that the LLM embeddings capture genetic relationships between diseases. Discussion/Significance of Impact: Overall, we found that the LLM-derived encodings uphold biologically and clinically significant relationships across organ systems and disease types. These results suggest that LLM encodings could provide a universal framework for representing diseases as computable phenotypes and enable the discovery of complex disease relationships.

Objectives/Goals: The standard care for malignant gliomas includes maximal tumor resection, but challenges arise near functional (speech) areas. Direct cortical stimulation (DCS) identifies functional (nonresectable) cortex. We aim to identify electrophysiologic (via subdural electrode recordings [ECOG]) biomarkers of DCS-positive (functional) areas. Methods/Study Population: Our lab maintains one of the largest datasets of electrophysiology analysis of glioma infiltrated brain cortex in the USA. Recordings of intraoperative brain mapping were analyzed to identify cortical sites that were found to be positive (functional) during DCS. DCS positive and negative (nonfunctional) sites were aligned to corresponding subdural electrodes. Future analysis: We plan to compare the temporal and spectral electrophysiologic variations associated with cortical sites found to be DCS positive versus negative during brain mapping. We plan to train machine learning classifiers that utilize these electrophysiologic biomarkers to discriminate between DCS positive and negative sites. Results/Anticipated Results: In total, our database comprised of 110 resections with brain mapping (DCS) and ECOG, including 4 patients who underwent a second procedure for resection. Eight patients were excluded as their resections were for brain metastases, not glioma. Our final cohort was comprised of 98 glioma resections, including 4 patients who underwent surgery twice for recurrence. During these resections, a total of 1393 sites were mapped via DCS for language function (including picture naming, word reading, and sentence syntax tasks). Of these 1393 sites, 100 sites were found to be DCS positive (7.1% positivity rate). (Currently in the process of conducting analysis comparing electrophysiologic features and biomarkers of DCS positive versus negative sites.) Discussion/Significance of Impact: This research is ongoing. Identifying electrophysiologic biomarkers of critical DCS-positive regions may provide a durable alternative to stimulation mapping. Due to its resource intensity, DCS has access barriers. Future neurosurgeons may use biomarkers from subdural electrode recordings to plan safer cortical resections.

Objectives/Goals: Neurodegenerative diseases involve progressive neuronal loss or dysfunction, often due to accumulated damage and impaired repair mechanisms. Our research evaluates the role of innate immune recognition proteins to provide insights into age-related neurodegeneration and cognitive decline. Methods/Study Population: We will utilize transcriptomic data from the Long-Life Family Study (LLFS), a cohort rich in genetic and phenotypic data related to aging and longevity. Our approach includes assessing a set of innate immune recognition proteins, also known as pattern recognition receptors (PRRs) expression across various age groups, focusing on potential correlations with cognitive performance. By analyzing serum transcriptomic profiles, we aim to map changes in expression and DNA repair genes over time, evaluating their connection to cognitive health and neurodegeneration in aging populations. Results/Anticipated Results: We anticipate that the expression of some PRRs will increase with age and correlate with cognitive decline, suggesting a role in age-related neurodegeneration. We also expect a decrease in DNA repair pathway gene expression in older age groups, contrasting with an increase in genes involved in endogenous DNA detection. These results will reveal how PRRs may function as neuroprotective factors and how their expression changes may relate to the decline in DNA repair processes with age, providing a better understanding of innate immune recognition in cognitive health. Discussion/Significance of Impact: This study will reveal the role of PRRs in aging and neurodegeneration, potentially establishing them as a key player in neuronal protection. Findings may guide future research into therapeutic strategies targeting them for Alzheimer’s and other age-related neurodegenerative diseases.

Objectives/Goals: Global surgical education is largely driven by high-income countries (HICs), with curricula not tailored to the needs of low- and middle-income countries (LMICs). This study assessed country-specific needs for global surgical curricula and used generative AI to develop tailored curricula. Methods/Study Population: A curriculum framework was developed using expert opinion. Using a focused needs assessment survey, we evaluated international medical students’ and trainees’ needs for structured global surgery curricula, covering research, education, data and develop tailored curriculum templates for each country, ensuring alignment with the distinct needs of respective LMIC and HIC respondents. The AI-generated curricula were then compared across countries to identify variations in content and focus areas. Results/Anticipated Results: A total of 145 respondents from 18 countries and 6 continents participated, with 94 from LMICs and 51 from HICs. Four countries [Uganda (n = 31), Nigeria (n = 34), the USA (n = 23), and the UK (n = 23)] had more than 10 respondents, with the creation of a country specific global surgery curriculum. Curricula developed by HIC trainees focused on access to resources and infrastructure, future directions of global surgical research, and the role of medical students and early career development with a decreased focus on the history of global surgery. LMIC country-based curriculum focused on introducing the concepts of global surgery, quantifying the burden and epidemiology of surgical disease and had a greater emphasis on case studies and use cases, with decreased focus on resources and collaboration. Discussion/Significance of Impact: The research introduces a “precision education” approach that could help close the surgical education access gap globally. Further pilot and qualitative studies are necessary to validate the feasibility of AI-generated needs-based curricula.

Objectives/Goals: We aim to enhance risk prediction in kidney transplantation outcomes by improving models of peptide antigen presentation of mismatched HLA molecules. HLA-derived peptides presented by HLA Class II to T-cells can activate an immune response, ultimately leading to graft failure. We aim to improve peptide prediction by modeling antigen processing. Methods/Study Population: T-cell epitope models for HLA mismatching struggle to predict which peptides are presented because antigen processing by proteases is not well modeled. We model antigen processing of HLA Class II proteins using 3D HLA structures (crystallography data) to create an HLA-specific antigen processing likelihood (APL) model. APL uses conformational stability measurements such as b-factor, COREX, solvent accessible surface area, and sequence entropy to predict cleavage sites from proteolysis. We will integrate APL into a T-cell epitope prediction tool for HLA-derived peptides based on donor and recipient HLA genotypes. Finally, we will associate the risk of graft failure with counts of these peptides derived from APL-integrated prediction models using a historical kidney transplant cohort from 2000 to 2023. Results/Anticipated Results: We expect that applying APL could reduce false-positive peptide binders influencing risk prediction scores. We anticipate improved peptide prediction accuracy compared to existing tools such as NetMHCIIPan, which assumes all possible peptides are equally likely to emerge from antigen processing. NetMHCIIPan is currently used by PIRCHE-II HLA mismatch risk algorithm. We expect that merging antigen processing (APL) and peptide-binding (NetMHCIIPan) models into a unified model would enhance risk stratification for graft failure. Current risk stratification still leads to poor outcomes post-transplant, especially for minority population groups. Our model can identify an alternative pool of well-matched donors and has the potential to improve equity for non-White minority candidates. Discussion/Significance of Impact: Improving the understanding of how HLA matching contributes to kidney transplant outcomes can better stratify risks for kidney transplant recipients, enable personalized treatment, and ultimately improve outcomes for those undergoing kidney transplantation to treat renal diseases.

Objectives/Goals: The research aims to prototype a mobile app for physicians to remotely monitor patients receiving CD19-directed CAR T-cell therapy post-discharge. This app will facilitate standardized data collection across carious CAR-T treatment centers and help harmonize follow-up protocols. Methods/Study Population: A literature review and semi-structured interviews with patients, clinical coordinators, and experts helped identify essential parameters for a mobile app prototype aimed at monitoring adverse effects such as cytokine release syndrome and neurotoxicity. The app was designed through process mapping to combine data from self-reports and wearable devices, such as the Garmin smartwatch. New screens were designed in Figma, drawing from an existing patient monitoring app for allogeneic stem cell transplant follow-up. Finally, a preliminary feasibility study will be conducted to gather feedback on the app prototype from CAR T-cell therapy patients, healthcare providers, and stakeholders, ensuring its effectiveness and usability. Results/Anticipated Results: Semi-structured interviews with people with professional and lived experience with CAR T-cell therapy were conducted to determine what metrics might be monitored and how they could be measured remotely to effectively monitor for side effects. The mobile phone application was then prototyped using process mapping in Visio®, designed in Figma, and preliminary development was completed. The final prototype includes parameters that will be recorded or measured using a combination of self-reporting and devices to monitor body temperature, basic vitals, activity, sleep, and cognitive function, among others. The prototype of the remote monitoring app is the first step in implementing remote monitoring of CAR T patients, standardized data collection, and reduction in the overall cost of CAR T-cell therapy. Discussion/Significance of Impact: This app will enable physicians to monitor patients for routine follow-ups and adverse effects, such as CRS and ICANS. Future research will validate the digitized ICANS assessment and used to establish best practices for standardizing CAR-T follow-up protocols across Canada.

Objectives/Goals: Our research group is focused on care of hospitalized persons with opioid use disorder (OUD) in the era of high-potency synthetic opioids (HPSO). In this work, we describe trends in patient-directed discharge (PDD) and inpatient treatment with medications for opioid use disorder (MOUD). We hypothesized that PDD is associated with MOUD dose and timing. Methods/Study Population: Patient data generated in the routine care of patients was automatically abstracted using a SQL query on Epic Clarity tables in the electronic health record (EHR). We included adult patients admitted to Johns Hopkins Hospital between July 1, 2019 and June 30, 2022, with an ICD-10-CM code for a list of opioid-related disorders (F11.X) consistent with OUD. Demographics, prior medication list, clinical care including hospital service, consultation services, COWS scores, length of time in emergency department, time of triage, time until receipt of methadone or buprenorphine, dosage and timing of MOUD, opioid medications other than methadone or buprenorphine, adjuvant medications; prior methadone or buprenorphine treatment and disposition. Query results were validated by manual abstraction of EHR. Results/Anticipated Results: The SQL identification of the cohort of patients with OUD was found to be accurate. Time of triage, discrete orders completed during hospitalization were well represented in the query. The query was able to identify individual opioid medications but unable to summarize total dose in Morphine Milligram Equivalents. The query did not extract accurate information from patient-controlled analgesia pumps due to the continuous nature of the medication rather than discrete doses reflected in the medication administration record. Finally, the query characterized prior treatment with methadone or buprenorphine as a binary variable – dosage and timing of that prior treatment could not be accurately represented. Finally, stimulant use is not reliably collected in the EHR and was unavailable. Discussion/Significance of Impact: Given the rise of HPSO, patients may not tolerate delay of MOUD. Improving the granularity of data collected will offer more insight into the inpatient treatment for OUD. Real-world data have promise but requires extensive technical expertise. Future work is needed to improve capture of derived variables such as total dosage of opioids in MME.

Objectives/Goals: Sexual minority populations report a disproportionately high prevalence of alcohol use, often attributed to coping with bi/homonegativity and systemic inequities across various social domains. This study aims to explore alcohol use patterns and associated neighborhood and individual factors among sexual minority populations (SMPs) using data from the NIH All of US dataset. Methods/Study Population: Alcohol use was assessed using the AUDIT-C (Alcohol Use Disorders Identification Test—Consumption) scale across a sample of 9,454 gay, 15,284 bisexual, 5,267 lesbian, and 349,748 straight participants. The AUDIT-C measured hazardous alcohol use, and logistic regression models were employed to examine its association with neighborhood-level factors (e.g., socioeconomic status, alcohol outlet density) and individual-level factors (e.g., age, race/ethnicity, income, and education) among SMPs. Interaction terms assessed how these relationships varied by sexual orientation. Sensitivity analyses were conducted to assess the robustness of the findings, including stratified analyses by gender identity and the exclusion of extreme outliers in alcohol use reporting. Results/Anticipated Results: Our analyses revealed that gay participants had the highest AUDIT-C scores (mean  =  3.60, SD  =  2.27), followed by bisexual (mean  =  3.35, SD  =  2.21), other SMPs (mean  =  3.18, SD  =  2.19), lesbian (mean  =  3.04, SD  =  2.08), and straight individuals (mean  =  3.05, SD  =  2.06). Alcohol use was positively associated with neighborhood disorder (β  =  0.12, 95% CI  =  0.07, 0.17), housing insecurity (β  =  0.14, 95% CI  =  0.03, 0.25), and male gender (β  =  0.98, 95% CI  =  0.96, 1.00). In contrast, neighborhood density (β  =  -0.11, 95% CI  =  -0.15, -0.07), food insecurity (β  =  -0.14, 95% CI  =  -0.20, -0.08), being Black, and identifying as bisexual were negatively associated with alcohol use. Sensitivity analyses determined no significant differences among specfic supgroups. Discussion/Significance of Impact: This study highlights important differences in alcohol use across SMPs and emphasizes the influence of neighborhood-level stressors (e.g., disorder and housing insecurity). These findings underscore the need for addressing social and environmental determinants of alcohol use in SMPs to mitigate the negative impacts of alcohol consumption.

Objectives/Goals: We assessed the feasibility of using a large language model (LLM) to create lay language descriptions of study protocols for recruitment, which has the potential to improve accessibility and transparency of clinical studies and enable participants to make informed decisions. Methods/Study Population: All studies from a clinical research recruitment platform were included, which features human-written lay descriptions and titles for study recruitment. Corresponding protocol summaries in the IRB system were extracted and translated into lay language using a LLM (gpt-35-turbo-0613). A subset was used to develop prompt variations through an iterative process. Prompt strategies evaluated include chain-of-thought and few-shot prompting techniques. LLM-generated and human-written descriptions were compared for readability using Flesch–Kincaid and Simple Measure of Gobbledygook (SMOG) reading grade levels and information completeness using Word Movers’ Distance (WMD). Results/Anticipated Results: A total of 55 study descriptions were included – 10 were used to develop prompts and 45 were used for evaluation. The final LLM instructions included multistep prompts. The LLM was first instructed to produce a two- to three-sentence long description without using scientific jargon and included two pairs of examples. The LLM was then asked to shorten the description and finally to provide an engaging title. LLM-generated and human-written summaries were similar in length (median (IQR) 328 (278.5–360.5) vs. 342 (203–532.5) characters, respectively). LLM-generated summaries had lower Flesch-Kincaid grade level (5.15 vs. 8.28, p Discussion/Significance of Impact: An LLM can be used to generate lay language summaries that are readable at a lower grade level while maintaining semantic similarity. This approach can be used to improve the drafting of summaries for recruitment, thereby improving accessibility to potential participants. Future work includes human evaluation and implementation into practice.

Objectives/Goals: Both diabetes and obesity are major health issues in the USA. Though much focus is given on the impact of lifestyle modifications to control diabetes and obesity, more information needs to be established about the association of social determinants of health and them. This study explores these associations, focusing on the DMV area. Methods/Study Population: The data for this study were collected for 158 counties which cover two states Maryland and Virginia, and Washington DC from PolicyMap. PolicyMap is a geography information system (GIS) that aggregates different types of data from different sources for research purposes. County-level data on public health related to different SDoH like median age, and percentages proportions of different ethnicity (Hispanic or Latino), percentages of different race, and nativity (foreign born), gender (ratio of male to female), access to primary healthcare, social vulnerability index (SVI), and median household income was used in this study. Statistical methods like multiple regression, one way ANOVA, and Pearson’s correlation coefficients were used to determine which factors are associated with these two conditions. Results/Anticipated Results: For both diabetes prevalence and obesity, multiple linear regression model with backward elimination was used to select variables which associated with them. The backward elimination process selected the set of factors for which the adjusted R square was the highest. In both cases, median household income, median age of population, social vulnerability level, percentage of white population, and percentage of foreign-born population were found to be significant at 5% level of significant. Pearson’s correlation coefficients showed significant positive relationship between factors like obesity and diabetes, median age and access to health care, and negative relationship between obesity and foreign born. Income, healthcare access, and white population were found to be significantly different SVIs from ANOVA. Discussion/Significance of Impact: This research study found that some SDoH affect diabetes and obesity in the same direction. The association is positive for median age and negative for income, SVI, percentage of white population, and foreign born. The associations were found between actionable and non-actionable factors like percentage of white population with access to health care.

Objectives/Goals: This Weill Cornell Clinical and Translational Science Collaborative (CTSC) project evaluates whether large language models (LLMs) can generate accurate summaries of translational science benefits using the Translational Science Benefits Model (TSBM) framework, aiming to identify optimal LLMs and prompting strategies via expert review. Methods/Study Population: We are using prompt engineering to train multiple LLMs to generate one-page impact profiles based on the TSBM framework. LLMs will be selected via benchmarks, focusing on models excelling in information extraction. Leading LLMs (e.g., Llama 3.2, ChatGPT 4.0, Gemini 1.5 Pro, and Claude) and other high-performing models will be considered. Initial work has utilized Gemini 1.5 Pro. Models use data from CTSC-supported projects in WebCAMP, our local instantiation of a translational research activity tracking system used by >20 CTSA hubs, and manuscripts from the Overton database cited in policy documents. Human experts will evaluate the quality and accuracy of LLM-generated profiles. Results/Anticipated Results: Preliminary results using Gemini 1.5 Pro indicate that LLMs can generate coherent and informative impact profiles encompassing diverse areas within the TSBM. Face validity appears satisfactory, suggesting the outputs align with expectations. We anticipate that further exploration with other LLMs and expert validation will reveal strengths and weaknesses of the LLM approach, including the potential for naccuracies (“hallucinations”), informing further refinement of models and prompting strategies. Analysis of manuscripts cited in policy will provide valuable insights into communicating policy-relevant benefits effectively, and benchmark comparisons will identify optimal LLMs for this use case. Discussion/Significance of Impact: This project demonstrates LLMs’ potential for streamlining and enhancing impact reporting in translational science, enabling broader dissemination of research outcomes and promoting better understanding among stakeholders. Future work will integrate LLM-based reporting into research infrastructure.

Objectives/Goals: Our goal was to determine whether a consensus exists around 1) what the main barriers to innovation in Health artificial intelligence (AI) are 2) where there are gaps in education and training in Health AI and 3) where in their workflows organizations should implement AI to see the most immediate impact on productivity. Methods/Study Population: We employed a three-round Delphi method survey to stakeholders with health and/or engineering expertise. The first round was open-ended to generate responses to the three research questions. The second round asked participants to rank the responses and provide feedback as to their reasoning. The third round provided aggregated results and feedback and asked participants to re-rank the responses. Participants were attendees at a conference that brought people with health and/or engineering backgrounds together to discuss innovation in Health AI. 55 people in total participated across the three rounds. Results/Anticipated Results: Consensus emerged on all three questions: lack of trust was seen as the single greatest barrier to innovation, experience with implementation as the greatest gap in training, and automating health documentation as the point of most immediate impact. Consensus also emerged as to which of the 10–15 responses to each question were top priorities, which were somewhat significant, and which were not that important. Some of the rankings (such as implementation) seemed to reflect hot topics of discussion at the conference, but others (such as documentation) only emerged as significant in the surveys. Discussion/Significance of Impact: We successfully employed the Delphi method to discover what stakeholders think about three important questions in Health AI. Interestingly, although we polled experts from both health and engineering backgrounds, their answers converged on all three questions.

Objectives/Goals: To investigate interventional clinical trial participation overall and by race, gender, and age. Methods/Study Population: We used Epic Cosmos, an aggregated, de-identified EHR platform including over 270 million patients, to examine overall clinical trial participation and the race, gender, and age composition of participants versus non-participants. Patients ≥5 years old with known race and gender and at least one healthcare encounter between 2021 and 2024 were included. Interventional trial enrollment was identified by a “research flag” indicating current or past participation in an interventional study within an Epic system contributing data to Cosmos. Race was categorized as American Indian, Asian, Black, Native Hawaiian, or White. Age-adjusted relative representation (RR) ratios were used to compare participation, with RR >1 indicating over-representation and RR Results/Anticipated Results: Of 130,455,189 patients meeting eligibility criteria, 0.52% (673,425) of patients were active or inactive in an interventional clinical trial. Results are shown in the figure below. The poorest representation was from Asian and NH/PI persons. Representation was most similar to the patient population for whites and AI/AN persons. Black males participated less and women, more than predicted by patient composition. Older patients participated more frequently than younger (age, mean (SD), y, 53 (22) vs. 46 (23); p Discussion/Significance of Impact: This is the first study we know of describing interventional trial participation in the USA across millions and millions of patients. Further research is needed to clarify whether these differences are due to the nature of the studies themselves (e.g., OB/GYN trials including only women, etc.) versus disparities in recruitment or otherwise.

Objectives/Goals: Ventilator-associated pneumonia (VAP) is an infection caused by bacteria, viruses, or fungi during mechanical ventilation. We analyzed a cohort of COVID-19 patients admitted to the intensive care unit with respiratory failure with different VAP outcomes. We hypothesize that the multiomics data can help predict VAP development within this cohort. Methods/Study Population: We recruited participants from a cohort on a NYU IRB protocol (i22–00616), who had COVID19 respiratory failure, admitted to ICU, and required invasive mechanical ventilation (n = 245). We collected and analyzed research specimens (bronchoalveolar lavage [BAL, n = 213], tracheal aspirates [n = 246], background [n = 18]) and clinical cultures (sputum and BAL) for 245 participants. A panel of experts adjudicated VAP within the cohort, resulting in 92 VAP diagnoses. We annotated metatranscriptome (Illumina NovaSeq) using a Kraken/Bracken database, and KEGG for functional annotation of transcriptome data (Illumina HiSeq). We used edgeR (v.4.0.16) to analyze differential expression of metatranscriptome and transcriptome data. Results/Anticipated Results: We diagnosed VAP in n = 92 (38%) participants. We found significant differences in days of overall hospital stay (p Discussion/Significance of Impact: VAP is a serious complication of mechanical ventilation, and oral commensals alter the lung microbiome and host immunity. We identified a transcriptome-metatranscriptome signature that identifies those at VAP risk. VAP was associate with both pro- and anti-inflammatory gene expression resulting in increased risk for lower airway infection.