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

Electronic Health Record (EHR) audit log data are increasingly utilized for clinical tasks, from workflow modeling to predictive analyses of discharge events, adverse kidney outcomes, and hospital readmissions. These data encapsulate user-EHR interactions, reflecting both healthcare professionals' behavior and patients' health statuses. To harness this temporal information effectively, this study explores the application of Large Language Models (LLMs) in leveraging audit log data for clinical prediction tasks, specifically focusing on discharge predictions. Utilizing a year's worth of EHR data from Vanderbilt University Medical Center, we fine-tuned LLMs with randomly selected 10,000 training examples. Our findings reveal that LLaMA-2 70B, with an AUROC of 0.80 [0.77-0.82], outperforms both GPT-4 128K in a zero-shot, with an AUROC of 0.68 [0.65-0.71], and DeBERTa, with an AUROC of 0.78 [0.75-0.82]. Among various serialization methods, the first-occurrence approach-wherein only the initial appearance of each event in a sequence is retained-shows superior performance. Furthermore, for the fine-tuned LLaMA-2 70B, logit outputs yield a higher AUROC of 0.80 [0.77-0.82] compared to text outputs, with an AUROC of 0.69 [0.67-0.72]. This study underscores the potential of fine-tuned LLMs, particularly when combined with strategic sequence serialization, in advancing clinical prediction tasks.

Many sepsis prediction models use the Sepsis-3 definition or its variants as a training label. However, among the few sepsis models ever deployed in practice, there is scant evidence that they offer clinically meaningful decision support at the bedside. As a potential mechanism to explain this limitation, we hypothesized that clinician-recommended treatment times for sepsis would diverge from onset time defined by Sepsis-3. We conducted an electronic survey that was completed by 153 clinicians at three large and geographically diverse medical centers using vignettes derived from eight real cases of sepsis. After reviewing these vignettes, participants suggested antibiotic treatment to start an average of 7.0 hours (95% confidence interval 5.3 to 8.8) before the Sepsis-3 definition onset. Thus, predicting Sepsis-3 onset as a treatment prompt could lead to inappropriate and delayed treatment recommendations. Building predictive decision support systems that identify outcomes aligned with bedside decisions would increase their clinical utility.

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.

Causal inference, or the task of estimating the causal effect of an exposure or interventional variable on an outcome from an observational dataset, requires precise and rigorous methods, based on assumptions about the system under study. Such assumptions can be articulated as a causal diagram, however use of this technique in medicine is uncommon due to challenges in causal diagram construction in high-dimensional settings. Recent introduction of cluster causal diagrams or C-DAGs promise to ease the task of diagram construction by allowing for the representation of some unknown or partially defined relationships. We evaluate the practical application of C-DAGs in simulated medical contexts. We estimate causal effects under varying sets of assumptions, determined by both causal diagrams and C-DAGs and compare our results. Our findings show empirically similar results, with little discrepancy between causal effect sizes or variance across experimental runs, although estimation and efficiency challenges remain to be explored.

Integrating Electronic Health Records (EHR) and the application of machine learning present opportunities for enhancing the accuracy and accessibility of data-driven diabetes prediction. In particular, developing data-driven machine learning models can provide early identification of patients with high risk for diabetes, potentially leading to more effective therapeutic strategies and reduced healthcare costs. However, regulation restrictions create barriers to developing centralized predictive models. This paper addresses the challenges by introducing a federated learning approach, which amalgamates predictive models without centralized data storage and processing, thus avoiding privacy issues. This marks the first application of federated learning to predict diabetes using real clinical datasets in Canada extracted from the Canadian Primary Care Sentinel Surveillance Network (CPCSSN) without cross-province patient data sharing. We address class-imbalance issues through downsampling techniques and compare federated learning performance against province-based and centralized models. Experimental results show that the federated MLP model presents a similar or higher performance compared to the model trained with the centralized approach. However, the federated logistic regression model showed inferior performance compared to its centralized peer.

This study proposes to use the K-medoids clustering method to identify subtypes of Intensive Care Unit (ICU)-acquired acute kidney injury (AKI) patients based on serum electrolyte data. Three distinct AKI subtypes with different serum electrolyte characteristics were identified by clustering analysis. Further, descriptive analysis was employed to characterize in-hospital mortality and renal replacement therapy, diuretic and vasopressor usage in the three subtypes, and Chi-square tests were conducted to check the differences of prognosis and treatments among the identified subtypes. This study enables the subclassification of AKI patients in the ICU, facilitating ICU physicians to make timely clinical decisions about AKI, and ultimately may contribute to patient outcome improvement.

Clinical reporting of cardiac magnetic resonance (CMR) imaging exams is commonly performed with a dictation approach which requires great care to capture both consistent and comprehensive data. We sought to transform the reporting process by utilizing a structured report framework for reporting standardization, by incorporating automated transfer of data semi-automated segmentation tools for efficiency, and rule-based reporting requirements to improve quality and standardization. Interfaces between the applications used to schedule and protocol exams and to analyze the acquired images were created to bring the source information directly into the structured reporting environment. The physicians reporting CMR were surveyed to determine satisfaction and improved efficiency with the new process through self-reported reporting time. Quality improvement was assessed by examining the consistency of reported parameters with the inclusion of rule-based requirements. The designed structured reporting process with automated measurements and rule-based requirements resulted in significant improvement in report efficiency and quality.

Building on the success of the Segment Anything Model (SAM) in image segmentation, "PathSAM: SAM for Pathological Images in Oral Cancer Detection" addresses the unique challenges associated with diagnosing oral cancer. Although SAM is versatile, its application to pathological images is hindered by its inherent complexity and variability. PathSAM advances beyond traditional deep-learning methods by delivering superior accuracy and detail in segmenting critical datasets like ORCA and OCDC, as demonstrated through both quantitative and qualitative evaluations. The integration of Large Language Models (LLMs) further enhances PathSAM by providing clear, interpretable segmentation results, facilitating accurate tumor identification, and improving communication between patients and healthcare providers. This innovation positions PathSAM as a valuable tool in medical diagnostics.

Patient portal messages represent a unique source of clinical data due to how they represent the voice of the patient, provide a glimpse into care delivery between episodic synchronous appointments, and capture variations in patient behavior and health literacy. There is little understanding of how to best apply modern natural language processing (NLP) approaches, such as large, pre-trained language models (LLMs), to patient messages. In this study, we aim to explore different approaches in incorporating patient messages into an existing Emergency Departments (ED) visit risk prediction model currently deployed at Stanford Health Care. With the addition of patient message frequencies to the baseline we were able to achieve an improved AUC of .77 and a jump in the F1 score. In future work, we aim to build upon these findings and further test combination models to incorporate features around patient message content, in addition to message frequencies.

Mapping electronic health records (EHR) data to common data models (CDMs) enables the standardization of clinical records, enhancing interoperability and enabling large-scale, multi-centered clinical investigations. Using 2 large publicly available datasets, we developed transformer-based natural language processing models to map medication-related concepts from the EHR at a large and diverse healthcare system to standard concepts in OMOP CDM. We validated the model outputs against standard concepts manually mapped by clinicians. Our best model reached out-of-box accuracies of 96.5% in mapping the 200 most common drugs and 83.0% in mapping 200 random drugs in the EHR. For these tasks, this model outperformed a state-of-the-art large language model (SFR-Embedding-Mistral, 89.5% and 66.5% in accuracy for the two tasks), a widely used software for schema mapping (Usagi, 90.0% and 70.0% in accuracy), and direct string match (7.5% and 7.5% accuracy). Transformer-based deep learning models outperform existing approaches in the standardized mapping of EHR elements and can facilitate an end-to-end automated EHR transformation pipeline.