AMIA ... Annual Symposium proceedings. AMIA Symposium最新文献

While there is a growing recognition of the association between depression and asthma, few studies have leveraged deep learning-based (DL-based) models in a retrospective cohort study with a large sample size. We analyzed the association between depression and asthma among 239,161 participants of the All of Us Research Program through DL-based, logistic regression, and Cox Proportional Hazards (CoxPH) models. We used SHAP values to help interpret DL-based models and c-index to evaluate model performance. Results suggest a significant odds ratio for depression in asthma. The c-indices for the CoxPH, DeepSurv, and DeepHit models were 0.619, 0.625, and 0.596, respectively. SHAP indicated a different set of important variables when compared with the CoxPH model. In conclusion, we provide strong evidence of a positive relationship between depression and asthma. Also, DL-based models did not outperform the CoxPH model on the c-index. Sex at birth and income may play important roles in occurrence of depression in asthma patients.

The advent of clinical language models integrated into electronic health records (EHR) for clinical decision support has marked a significant advancement, leveraging the depth of clinical notes for improved decision-making. Despite their success, the potential vulnerabilities of these models remain largely unexplored. This paper delves into the realm of backdoor attacks on clinical language models, introducing an innovative attention-based backdoor attack method, BadCLM (Bad Clinical Language Models). This technique clandestinely embeds a backdoor within the models, causing them to produce incorrect predictions when a pre-defined trigger is present in inputs, while functioning accurately otherwise. We demonstrate the efficacy of BadCLM through an in-hospital mortality prediction task with MIMIC III dataset, showcasing its potential to compromise model integrity. Our findings illuminate a significant security risk in clinical decision support systems and pave the way for future endeavors in fortifying clinical language models against such vulnerabilities.

RealRisks is a decision aid that integrates patient-generated and electronic health record (EHR) data using Fast Healthcare Interoperability Resources (FHIR). It offers modules to enhance understanding of breast cancer risk and a way for individuals to review and modify their EHR data before it is used in their personal risk assessment. RealRisks intends to encourage high-risk patients to take risk-reducing measures. To better understand how patients understand risk and barriers to action, we conducted in-depth interviews as part of a usability study to assess the clarity and interpretability of RealRisks. Overall, participants demonstrated an improved understanding of breast cancer risk after using RealRisks. However, challenges were noted for certain concepts, in particular, lifetime risk, how benign breast disease affects your risk, and the differences between hereditary, sporadic, and familial cancer. The EHR download feature was well-received, but some raised concerns about insurance and privacy/security.

The All of Us (AoU) Research Program and UK Biobank (UKBB) boast a wealth of EHR data, which can be harnessed to refine cohort selection via rule-based phenotyping algorithms. The Observational Health Data Sciences and Informatics (OHDSI) Phenotype Library (PL) hosts many complex phenotyping rules. Here, we compare prevalence for 423 OHDSI PL cohorts in AoU and UKBB. For three select diseases (T2D, COPD, Acute MI), we analyze differences in demographics, social determinants of health (SDOH), geographic prevalence, and genome-wide association study (GWAS) results. We found that AoU has a significantly higher prevalence for 80% of phenotypes compared to UKBB. We also found that for the select diseases, SDOH variables between the two biobanks differ significantly. Findings for each of these three diseases confirm known regions of high risk. Additionally, GWAS in UKBB discovered more genes associated with each of the three diseases than GWAS in AoU.

With data considered as the 'oxygen' of public health, the Data Modernization Initiative (DMI) to enhance the public health data and information infrastructure is critical. The DMI Stories from the Field features data modernization from public health agencies to highlight success/progress. These stories (n=241) were analyzed, with outbreak response, information systems capacity, epidemiology/laboratory capacity being some of the common topics. A total of 199 codes across DMI stories were organized into 7 themes and the top 3 codes were communication, collaboration and public health agencies. Key takeaways and next steps were identified and validated with expert input across people, product, process and partnership categories and people factor was critical along with funding/sustainability. Ongoing DMI stories and future studies for evaluating impact are recommended. DMI stories are a great option to communicate the projects and impact of DMI to a larger public audience and garner support for this vital endeavor.

Large Language Model (LLM)-based conversational agents offer promising solutions for mental health support, but lack cultural responsiveness for diverse populations. This study evaluated the effectiveness of cultural prompting in improving cultural responsiveness and perceived empathy of LLM-generated therapeutic responses for Chinese American family caregivers. Using a randomized controlled experiment, we compared GPT-4o and Deepseek-V3 responses with and without cultural prompting. Thirty-six participants evaluated input-response pairs on cultural responsiveness (competence and relevance) and perceived empathy. Results showed that cultural prompting significantly enhanced GPT-4o's performance across all dimensions, with GPT-4o with cultural prompting being the most preferred, while improvements in DeepSeek-V3 responses were not significant. Mediation analysis revealed that cultural prompting improved empathy through improving cultural responsiveness. This study demonstrated that prompt-based techniques can effectively enhance the cultural responsiveness of LLM-generated therapeutic responses, highlighting the importance of cultural responsiveness in delivering empathetic AI-based therapeutic interventions to culturally and linguistically diverse populations.

This study explores the challenges faced by justice-impacted Black women during their reintegration into society, with a focus on mental health care access and the potential for technology-assisted interventions to address barriers. Participants from focus groups emphasized significant obstacles, including inadequate mental health resources during incarceration, insufficient post-release support, and barriers such as discrimination, lack of insurance, and transportation issues. When designing technology-assisted interventions, such as the Welcome Home app, additional considerations for justice-impacted Black women include trauma-informed design, tiered support systems, integration with electronic health records, privacy protection, and culturally tailored content. The study underscores the importance of culturally relevant, user-centered digital solutions to improve health outcomes and facilitate the successful reintegration of Black women impacted by the criminal legal system. Apps that provide a sense of community promote engagement, which may improve health outcomes.

The University of Texas Southwestern Medical Center (UTSW) and Texas Health Resources (THR) implemented an Observational Medical Outcomes Partnership (OMOP) common data model (CDM) that utilizes the Epic electronic health record's (EHR) extract, transform, and load (ETL) system to enable collaborative research with other health institutions within the OHDSI network. We mapped EHR data from core Epic reporting tables to 25 OMOP CDM tables and transferred the data to a shared OMOP database housed within the Caboodle infrastructure using Epic's pre-existing ETL system, minimizing the need for customization. ETL processes occur weekly at THR and daily at UTSW. OMOP CDM mapping resulted in data quality assessment values of 97% and 98% for THR and UTSW respectively. Our study established a reproduceable, collaborative pipeline using the OMOP CDM with Epic's native ETL framework, expanding the OHDSI research network resulting in better quality and more generalizable data sets available for future research.

Clinical variability in prostate radiation therapy (RT) planning is well documented, but little is known about how radiation oncologists experience and adapt to the factors that drive it. This study explores variability as a human-centered design challenge, with the goalofinformingclinicaldecision support (CDS) design through real-timeinsight into planning decisions. We conducted observation sessions with the think aloud method followed by semi-structured interviews with five radiation oncologists while they contoured prostate cases. Using the Systems Engineering Initiative for Patient Safety (SEIPS) framework, we thematically analyzed the contributors to variability across tasks, technology, and organizational conditions. Results suggest that variability arises not only from anatomical or guidelineambiguity, butalso fromindividual interpretations of inputs, variation in contouring decisions, andadaptive strategies such as reliance on prior experience and estimation under uncertainty. Findings support the design of context-sensitive CDS tools that reflect real-world clinical reasoning while preserving clinical flexibility.

Influenza vaccine effectiveness (VE) estimation plays a critical role in public health decision-making by quantifying the real-world impact of vaccination campaigns and guiding policy adjustments. Current approaches to VE estimation are constrained by limited population representation, selection bias, and delayed reporting. To address some of these gaps, we propose leveraging large language models (LLMs) with few-shot chain-of-thought (CoT) prompting to mine social media data for real-time influenza VE estimation. We annotated over 4,000 tweets from the 2020-2021 flu season using structured guidelines, achieving high inter-annotator agreement. Our best prompting strategy achieves F₁ scores above 87% for identifying influenza vaccination status and test outcomes, outperforming traditional supervised fine-tuning methods by large margins. These findings indicate that LLM-based prompting approaches effectively identify relevant social media information for influenza VE estimation, offering a valuable real-time surveillance tool that complements traditional epidemiological methods.