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

Implicit bias can impede patient-provider interactions and lead to inequities in care. Raising awareness is key to reducing such bias, but its manifestations in the social dynamics of patient-provider communication are difficult to detect. In this study, we used automated speech recognition (ASR) and natural language processing (NLP) to identify social signals in patient-provider interactions. We built an automated pipeline to predict social signals from audio recordings of 782 primary care visits that achieved 90.1% average accuracy across codes, and exhibited fairness in its predictions for white and non-white patients. Applying this pipeline, we identified statistically significant differences in provider communication behavior toward white versus non-white patients. In particular, providers expressed more patient-centered behaviors towards white patients including more warmth, engagement, and attentiveness. Our study underscores the potential of automated tools in identifying subtle communication signals that may be linked with bias and impact healthcare quality and equity.

Training Large Language Models (LLMs) with billions of parameters on a dataset and publishing the model for public access is the current standard practice. Despite their transformative impact on natural language processing (NLP), public LLMs present notable vulnerabilities given the source of training data is often web-based or crowdsourced, and hence can be manipulated by perpetrators. We delve into the vulnerabilities of clinical LLMs, particularly BioGPT which is trained on publicly available biomedical literature and clinical notes from MIMIC-III, in the realm of data poisoning attacks. Exploring susceptibility to data poisoning-based attacks on de-identified breast cancer clinical notes, our approach is the first one to assess the extent of such attacks and our findings reveal successful manipulation of LLM outputs. Through this work, we emphasize on the urgency of comprehending these vulnerabilities in LLMs, and encourage the mindful and responsible usage of LLMs in the clinical domain.

Evidence-based clinical guidelines serve to support clinical decision making, but implementing such guidelines into practice remains a challenge. Familial hypercholesterolemia (FH) is a high impact clinical condition that exemplifies this disconnect. Using implementation science methods, we designed clinical decision support tools embedded into the electronic health record, including a FH-focused electronic health record Smart Set and clinic note template, to improve the care of adult and pediatric patients at high-risk of FH. End-user feedback gathered through direct observations, semi-structured interviews, and deliberative engagement sessions was used to inform the development of the tools before and after pilot-testing. Clinicians desired comprehensive, guidelines-based tools that promoted collaborative care. During pilot testing, end-users provided insights into technical issues encountered with the tool's first iteration and suggested regular check-in sessions to monitor issues moving forward. This methodology can be used to surmount challenges that prevent the uptake of evidence-based guidelines into practice.

Introduction: Clinical trials play a crucial role in precision cancer care. Patients generally learn of trials from their physician, and physician recognition of potential matches can be enhanced through decision support tools. But automated trial matching remains challenging, particularly for molecular eligibility criteria. Objective: We assessed the feasibility of FHIR Genomics plus CQL to enable trial matching, particularly for molecular criteria. Methods: We developed a prototype that included (1) encoded trial criteria in CQL; (2) synthetic patient clinical and genomic data; (3) trial eligibility computation. Results: We found that even complex molecular eligibility criteria can be represented in CQL given that the semantics of a criterion are formalized in base FHIR specifications. The proof of concept "CQL for Clinical Trials Matching" is available at [https://elimu.io/downloads/]. Discussion and Conclusions: Proof of concept work suggests FHIR and CQL as viable options for enhancing clinical trial matching.

Health Related Social Needs (HRSN) is an important driver of patient health outcomes. Healthcare organizations address patient HRSN with screening and community resource referral fulfillment (S&RF) processes for which they lack patient retention data, due to information silos. The process is complex and not fully represented in available conceptual models nor adequately assessed for effectiveness. The objective was to develop an evidence-based HRSN S&RF complex model and identify patient retention parameters. Model development drew from the literature and expert input to create a complex S&RF model, and identify parameters for model stages and factors. Studies (50) involved manual S&RF processes in small, specialized populations. The model organized 88 factors among five S&RF stages. Half the studies reported parameters, for which stage and factor ranges were wide and indicated reduced patient retention along the process. Needed is data from routine care in which HRSN platforms are used, and information silos overcome.

In the classroom, artificial intelligence techniques help automate student behavior analysis, and teachers are able to understand students' class status more effectively. We developed an intelligent method for classroom behavior analysis by building a CQStu datasets and annotating 6,687 images through active learning. OpenPose was used to detect the key points of the student's body, and the key points of the key parts of the body were utilized to generate representative points of the student, and the idea of coordinates was used to assign the student's position. Using YOLOV5 to recognize students' classroom behaviors and count the number of times, our experimental results show that the average classroom behavior recognition accuracy is 84.23%, and the overall location accuracy is about 79.6%. In addition, we introduced a nonlinear weighting factor to evaluate the effectiveness of teaching and constructed corresponding classroom behavior weights based on different classroom scenarios. A method for student classroom behavior identification and analysis is provided, and a framework for future intelligent classroom teaching evaluation methods is established, providing objective data support for student performance analysis.

Introduction: In the evolving field of health informatics, the American Medical Informatics Association (AMIA) highlights the need for professionals skilled in current research. Journal clubs bridge academic learning with practical application, addressing challenges like limited literature review time and fostering critical analysis. Aim: This study evaluates the impact of an interdisciplinary journal club on 33 Master's students in Health and Bioinformatics program at Grand Valley State University. Thirteen students participated, analyzing contemporary literature and applying findings to real-world problems. Results: Significant improvements were observed in key capstone assessments among journal club participants: Capstone Overall Percentage (mean difference 15.23 points, p < 0.05), Project Proposal (mean difference 13.62 points, p < 0.05), and Research Topic Presentations (mean difference 27.30 points, p < 0.05). Conclusion: These findings support integrating journal clubs into curricula to enhance evidence-based practice, interdisciplinary collaboration, and practical application of knowledge, aligning with AMIA's vision of continuous professional development.

Antibiotics have been crucial in advancing medical treatments, but the growing threat of antibiotic resistance challenges these achievements and emphasizes the need for innovative stewardship strategies. In this study, we developed machine learning models, 'personalized antibiograms', to predict antibiotic resistance across five key antibiotics using Stanford's electronic health record data of 49,872 urine, blood, and respiratory infections. We aimed to ascertain the efficacy of these models in predicting antibiotic susceptibility and identify the clinical factors most indicative of resistance. Employing LightGBM, we incorporated demographics, prior resistance, prescriptions, and comorbidities as features. The models demonstrated notable discriminative ability, with AUROCs between 0.74 and 0.78, and highlighted prior resistance and prescriptions as significant predictive factors. The high specificity demonstrates machine learning models' potential to inform antibiotic de-escalation, aiding stewardship without risking safety. By leveraging machine learning with relevant clinical features, we show that it is feasible to improve empirical antibiotic prescribing.

Deep learning advancements have revolutionized scalable classification in many domains including computer vision, healthcare and Natural Language Processing (NLP). However, when it comes to classification and domain adaptation based on wearables, it suffers from persistent underperformance, largely due to the scarcity of pre-trained deep learning models that are abundantly available for computer vision and NLP. This is primarily because wearable sensor data need sensor-specific preprocessing, architectural modification, and extensive data collection. We present a novel modified-recurrent plot-based image representation that seamlessly integrates both temporal and frequency domain information. We employ an efficient Fourier Transform-based frequency domain angular difference estimation scheme in conjunction with existing temporal recurrent plots. We validated proposed method in two different domains: accelerometer-based activity-recognition and real-time glucose level prediction from wearable sensors. Our findings demonstrated the method we developed not only improves accuracy at recognizing activity but also makes a big leap in glucose level prediction.

Indexing articles by their publication type and study design is essential for efficient search and filtering of the biomedical literature, but is understudied compared to indexing by MeSH topical terms. In this study, we leveraged the human-curated publication types and study designs in PubMed to generate a dataset of more than 1.2M articles (titles and abstracts) and used state-of-the-art Transformer-based models for automatic tagging of publication types and study designs. Specifically, we trained PubMedBERT-based models using a multi-label classification approach, and explored undersampling, feature verbalization, and contrastive learning to improve model performance. Our results show that PubMedBERT provides a strong baseline for publication type and study design indexing; undersampling, feature verbalization, and unsupervised constrastive loss have a positive impact on performance, whereas supervised contrastive learning degrades the performance. We obtained the best overall performance with 80% undersampling and feature verbalization (0.632 macro-F1, 0.969 macro-AUC). The model outperformed previous models (MultiTagger) across all metrics and the performance difference was statistically significant (p < 0.001). Despite its stronger performance, the model still has room for improvement and future work could explore features based on full-text as well as model interpretability. We make our data and code available at https://github.com/ScienceNLP-Lab/MultiTagger-v2/tree/main/AMIA.