HIV Risk Assessment using Longitudinal Electronic Health Records

BackgroundDisease surveillance of diabetes among youth has relied mainly upon manual chart review. However, increasingly available structured electronic health record (EHR) data have been shown to yield accurate determinations of diabetes status and type. Validated algorithms to determine date of diabetes diagnosis are lacking. The objective of this work is to validate two EHR-based algorithms to determine date of diagnosis of diabetes.MethodsA rule-based ICD-10 algorithm identified youth with diabetes from structured EHR data over the period of 2009 through 2017 within three children’s hospitals that participate in the SEARCH for Diabetes in Youth Study: Cincinnati Children’s Hospital, Cincinnati, OH, Seattle Children’s Hospital, Seattle, WA, and Children’s Hospital Colorado, Denver, CO. Previous research and a multidisciplinary team informed the creation of two algorithms based upon structured EHR data to determine date of diagnosis among diabetes cases. An ICD-code algorithm was defined by the year of occurrence of a second ICD-9 or ICD-10 diabetes code. A multiple-criteria algorithm consisted of the year of first occurrence of any of the following: diabetes-related ICD code, elevated glucose, elevated HbA1c, or diabetes medication. We assessed algorithm performance by percent agreement with a gold standard date of diagnosis determined by chart review.ResultsAmong 3777 cases, both algorithms demonstrated high agreement with true diagnosis year and differed in classification (p = 0.006): 86.5% agreement for the ICD code algorithm and 85.9% agreement for the multiple-criteria algorithm. Agreement was high for both type 1 and type 2 cases for the ICD code algorithm. Performance improved over time.ConclusionsYear of occurrence of the second ICD diabetes-related code in the EHR yields an accurate diagnosis date within these pediatric hospital systems. This may lead to increased efficiency and sustainability of surveillance methods for incidence of diabetes among youth.

Foundation models pretrained on structured electronic health record (EHR) data promise improved predictive performance, sample efficiency and resilience to distribution shifts. However, model design, scale and use remain unclear. Objectives were to characterize foundation models pretrained on structured EHR data; examine temporal trends in model application and scale, architecture and design; and assess the extent to which publications omitted methodological details. We searched MEDLINE and Embase (2018-October 2025) for foundation models pretrained on structured EHR data using self-supervised learning and applied to clinical prediction tasks. Study selection and data abstraction were performed in duplicate. Characteristics were summarized and stratified by median publication year. Fifty-three studies were included; publications increased over time. Most datasets (79%) originated from the United States. None pretrained exclusively on pediatric cohorts. Model architecture shifted towards transformers (P = .013) with longer context windows (P = .028), while application shifted from exclusively embedding-based toward generative or mixed use (P < .001). Choices regarding feature inclusion, temporal representation, self-supervised objective and downstream adaptation remained heterogeneous. Only 26% of studies evaluated transfer to external datasets, and none described clinical deployment. Key indicators of scale and compute were frequently unreported. EHR foundation models are proliferating and increasingly transformer-based and generative. Yet methodological choices and reporting remain fragmented, indicating design trade-offs and best practices for EHR foundation models have not yet been established. None describe clinical deployment. Future work should clarify which design choices improve performance, robustness and transferability, increase reporting transparency and identify if they can be implemented to improve patient-important outcomes.

580 Background: Measurements of cancer care quality are critical for internal administration and participation in value-based care programs, towards the goal of improving patient outcomes. Data from electronic health records (EHRs) have the potential to allow for automatic computation of quality measures at a larger scale than manual chart review. This requires the creation of algorithms that process data and apply custom logic to define a clinically meaningful measurement. In this work, we describe the development and performance of an algorithm for deriving time from first primary oncologist visit to first cancer treatment using EHR data. Methods: This EHR-derived measure of time to first treatment (TTT) was implemented at an academic medical center for all patients who had a visit since 01/01/2011 and a cancer-related diagnosis code. We defined a patient’s primary oncologist as the first medical oncologist who they visited at least 5 times. These criteria were chosen to find the first provider at the site who the patient likely established care with and received treatment from. Then the patient’s TTT is computed as the number of days between their first visit date with this oncologist and the date of their first antineoplastic drug administration. Data on other treatment modalities, such as surgery and radiation therapy, were not available at the time of algorithm development. We allow for patients to receive their first treatment up to 60 days before their first visit to account for inpatient treatments before a primary outpatient oncologist is established. This algorithm’s performance was assessed through manual chart review among a sample of 38 patients. Results: Our algorithm correctly assigned 37 out of 38 patients (97.4%) to their primary oncologist, and computed the correct TTT for 28 out of 38 patients (73.4%). Incorrect TTT computations were mostly due to missingness of systemic therapy administrations in the structured EHR data; these were primarily for oral therapies. Conclusions: This work demonstrates the feasibility of measuring a patient’s TTT from structured EHR data alone. In particular, a patient’s initial visit with their primary oncologist can be derived with high accuracy. From this index date, the accuracy of the metric is dependent on the availability of structured treatment data. To further improve the EHR-derived metric, future work should incorporate other treatment modalities through structured data sources such as billing codes, and improve completeness of systemic therapy data in a scalable way, such as from natural language processing techniques.[Table: see text]

2043 Background: Electronic health records (EHR) are used for retrospective cancer outcomes analysis. Sites and timing of recurrence are not captured in structured EHR data. Novel computerized methods are necessary to use unstructured longitudinal EHR data for large scale studies. Methods: We previously developed a neural network-based NLP algorithm to identify no recurrence vs. metastatic recurrence cases by analyzing physician notes, pathology and radiology reports in Stanford’s breast cancer database, Oncoshare (Cohort A). To validate this algorithm for local vs. distant recurrence, we identified a distinct Oncoshare cohort (Cohort B). Cases were manually curated for longitudinal development of local or distant recurrence and metastatic sites. A two-sided t-test was used to compare mean probabilities between local and distant recurrence cases. Next, we combined cases in Cohorts A and B to train and validate a novel NLP classifier that identifies metastatic site. The combined cohort was randomly divided into training and validation sets. Sensitivity and specificity were calculated for the NLP algorithm’s ability to detect metastatic sites compared to manual curation. Results: In Cohort B: 350 metastatic cases were identified. Mean probability for local and distant recurrence was 0.43 and 0.79, respectively and differed significantly for patients with local vs. distant recurrence (p&lt;0.01). In Cohorts A and B: 632 metastatic cases were used for determination of sites. Sensitivity and specificity were highest for detection of peritoneal metastasis followed by liver, lung, skin, bone and central nervous system (table). Conclusions: This NLP algorithm is a scalable tool that uses unstructured EHR data to capture breast cancer recurrence, distinguishing local from distant recurrence and identifying metastatic site. This method may facilitate analysis of large datasets and correlation of outcomes with metastatic site. [Table: see text]

ABSTRACTObjective: The US population of older adults is growing, with an increase in diseases like cancer. As cancer rates increase, there is a concomitant increase in adverse correlates, such as cognitive impairment and depressive symptomatology. In order to develop appropriate interventions, it is vital to assess relationships among cancer, depressive symptoms and cognitive functioning.Methods: The sample consisted of 403 older adults with cancer diagnoses from the Health and Retirement Study. Using latent class growth analysis, longitudinal data were explored. The goals were to investigate trajectories of cognitive functioning, and to identify whether depressive symptoms and demographic factors predicted membership in the cognitive classes.Results: Three classes of cognitive functioning best fit the data: High, Middle and Low Recall, fairly stable trajectories from pre-diagnosis to a period four years after diagnosis. More depressive symptoms after diagnosis (but not prior) significantly predicted membership in the Low Recall class. Depressive symptoms did not distinguish between the High and Middle Recall classes.Conclusion: Depressive symptomatology is thought to affect cognition in late life. We found that depressive symptoms after a cancer diagnosis, but not before, successfully differentiated between those who had Low Recall from those with Middle and High Recall. Implications are discussed.

Background Alzheimer’s disease and related dementias (ADRD) are among the top causes of mortality and disability in older adults. Early detection is critical for timely clinical intervention, caregiver planning, and patient enrollment in clinical trials. Electronic health records (EHRs) offer a readily accessible data source for automated dementia risk screening. However, existing predictive models often struggle to accurately identify nuanced or complex patient presentations, limiting their clinical utility. Recently, large language models (LLMs)—advanced artificial intelligence systems proficient at reasoning and interpreting medical data—have demonstrated promising capabilities for improving diagnostic accuracy. Methods We propose a novel dementia risk prediction pipeline that integrates traditional supervised machine learning methods (SLs) with advanced LLMs. Our method uses SLs to accurately classify clear-cut dementia risk cases from structured EHR data, while leveraging LLMs’ sophisticated reasoning capabilities to handle ambiguous or challenging patient records. We evaluated this collaborative approach using a large, real-world EHR dataset from Oregon Health &amp; Science University (OHSU), consisting of over 2.5 million patient records spanning more than 20 million clinical encounters. Results Our integrated approach significantly outperformed traditional predictive models alone, demonstrating improved accuracy and reliability in dementia risk prediction. The combined use of SLs and LLMs effectively captured subtle clinical indicators that traditional methods overlooked. Conclusion Integrating advanced language models into existing machine learning approaches markedly improves dementia screening from routine clinical data. This method shows substantial potential for enhancing early ADRD detection, offering physicians powerful tools for proactive, personalized patient management.

IntroductionReal-world evidence is receiving considerable attention as a way to evaluate the efficacy and safety of medical products for substance use disorders (SUDs). However, the feasibility of using real-world data (RWD) to emulate clinical trials evaluating treatments for SUDs is uncertain. The aim of this study is to identify the number of clinical trials evaluating treatments for SUDs with reported results that could be feasibly emulated using observational data from contemporary insurance claims and/or electronic health record (EHR) data.MethodsIn this cross-sectional study, all phase 2–4 trials evaluating treatments for SUDs registered on ClinicalTrials.gov with reported results were identified. Each trial was evaluated to determine if the indications, interventions, at least 80% of eligibility criteria, comparators, and primary end points could be ascertained using contemporarily available administrative claims and/or structured EHR data.ResultsThere were 272 SUD trials on ClinicalTrials.gov with reported results. Of these, when examining feasibility using contemporarily available administrative claims and/or structured EHR data, 262 (96.3%) had indications that were ascertainable; 194 (71.3%) had interventions that were ascertainable; 21 (7.7%) had at least 80% of eligibility criteria that were ascertainable; 17 (6.3%) had active comparators that were ascertainable; and 61 (22.4%) had primary end points that were ascertainable. In total, there were no trials for which all 5 characteristics were ascertainable using contemporarily available administrative claims and/or structured EHR data. When considering placebo comparators as ascertainable, there were 6 (2.2%) trials that had all 5 key characteristics classified as ascertainable from contemporarily available administrative claims and/or structured EHR data.ConclusionsNo trials evaluating treatments for SUDs could be feasibly emulated using contemporarily available RWD, demonstrating a need for an increase in the resolution of data capture within a public health system to facilitate trial emulation.

Identifying the 'gray zone': Developing scalable methods to detect opioid misuse in veterans on long-term opioid therapy for pain.

The goal of this study was to evaluate an artificial intelligence approach, namely deep learning, on clinical text in electronic health records (EHRs) to identify patients with cirrhosis. Accurate identification of cirrhosis in EHR is important for epidemiological, health services, and outcomes research. Currently, such efforts depend on International Classification of Diseases (ICD) codes, with limited success. We trained several machine learning models using discharge summaries from patients with known cirrhosis from a patient registry and random controls without cirrhosis or its complications based on ICD codes. Models were validated on patients for whom discharge summaries were manually reviewed and used as the gold standard test set. We tested Naive Bayes and Random Forest as baseline models and a deep learning model using word embedding and a convolutional neural network (CNN). The training set included 446 cirrhosis patients and 689 controls, while the gold standard test set included 139 cirrhosis patients and 152 controls. Among the machine learning models, the CNN achieved the highest area under the receiver operating characteristic curve (0.993), with a precision of 0.965 and recall of 0.978, compared with 0.879 and 0.981 for the Naive Bayes and Random Forest, respectively (precision 0.787 and 0.958, and recalls 0.878 and 0.827). The precision by ICD codes for cirrhosis was 0.883 and recall was 0.978. A CNN model trained on discharge summaries identified cirrhosis patients with high precision and recall. This approach for phenotyping cirrhosis in the EHR may provide a more accurate assessment of disease burden in a variety of studies.

Delirium is an acute syndrome characterized by fluctuating attention, cognitive impairment, and severe disorganization of behavior, which has been shown to affect up to 31% of patients in the intensive care unit (ICU). Early detection can enable timely interventions and improved health outcomes. While artificial intelligence (AI) models have shown great potential for ICU delirium prediction using structured electronic health records (EHR), most studies have either not leveraged state-of-the-art AI models, been limited to single-center cohorts, or relied on small datasets for development and validation. In this study, we introduce DeLLiriuM, a novel LLM-based delirium prediction model that utilizes EHR data from the first 24 hours of ICU admission to estimate a patient’s risk of developing delirium for the remainder of their ICU stay. We developed and validated DeLLiriuM using ICU admissions from 104,303 patients across 195 hospitals in three large databases: the eICU Collaborative Research Database, the Medical Information Mart for Intensive Care (MIMIC)-IV, and the University of Florida’s Integrated Data Repository. Our DeLLiriuM model achieved superior performance compared to all baseline models on the external validation set, measured by the area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPRC) metric. DeLLiriuM attained AUROC 82.4 (95% confidence interval 81.8–83.0) and AUPRC 11.8 (95% confidence interval 11.3–12.4) across 77,543 patients spanning 194 hospitals. Our approach of transforming structured EHR data into an unstructured text format, the primary data modality for LLMs, enables our DeLLiriuM model to capture clinical contextual information, resulting in improved predictive performance. To the best of our knowledge, DeLLiriuM is the first LLM-based delirium prediction tool for the ICU that utilizes structured EHR data with LLMs rather than clinical notes with LLMs or traditional structured feature representations used in AI models.

Large Language Models (LLMs) hold significant promise for improving clinical decision support and reducing physician burnout by synthesizing complex, longitudinal cancer Electronic Health Records (EHRs). However, their implementation in this critical field faces three primary challenges: the inability to effectively process the extensive length and fragmented nature of patient records for accurate temporal analysis; a heightened risk of clinical hallucination, as conventional grounding techniques such as Retrieval-Augmented Generation (RAG) do not adequately incorporate process-oriented clinical guidelines; and unreliable evaluation metrics that hinder the validation of AI systems in oncology. To address these issues, we propose CliCARE, a framework for Grounding Large Language Models in Clinical Guidelines for Decision Support over Longitudinal Cancer Electronic Health Records. The framework operates by transforming unstructured, longitudinal EHRs into patient-specific Temporal Knowledge Graphs (TKGs) to capture long-range dependencies, and then grounding the decision support process by aligning these real-world patient trajectories with a normative guideline knowledge graph. This approach provides oncologists with evidence-grounded decision support by generating a high-fidelity clinical summary and an actionable recommendation. We validated our framework using large-scale, longitudinal data from a private Chinese cancer dataset and the public English MIMIC-IV dataset. In these settings, CliCARE significantly outperforms baselines, including leading long-context LLMs and Knowledge Graph-enhanced RAG methods. The clinical validity of our results is supported by a robust evaluation protocol, which demonstrates a high correlation with assessments made by oncologists.

To examine the value of unstructured electronic health record (EHR) data (free-text notes) in identifying a set of geriatric syndromes. Retrospective analysis of unstructured EHR notes using a natural language processing (NLP) algorithm. Large multispecialty group. Older adults (N=18,341; average age 75.9, 58.9% female). We compared the number of geriatric syndrome cases identified using structured claims and structured and unstructured EHR data. We also calculated these rates using a population-level claims database as a reference and identified comparable epidemiological rates in peer-reviewed literature as a benchmark. Using insurance claims data resulted in a geriatric syndrome prevalence ranging from 0.03% for lack of social support to 8.3% for walking difficulty. Using structured EHR data resulted in similar prevalence rates, ranging from 0.03% for malnutrition to 7.85% for walking difficulty. Incorporating unstructured EHR notes, enabled by applying the NLP algorithm, identified considerably higher rates of geriatric syndromes: absence of fecal control (2.1%, 2.3 times as much as structured claims and EHR data combined), decubitus ulcer (1.4%, 1.7 times as much), dementia (6.7%, 1.5 times as much), falls (23.6%, 3.2 times as much), malnutrition (2.5%, 18.0 times as much), lack of social support (29.8%, 455.9 times as much), urinary retention (4.2%, 3.9 times as much), vision impairment (6.2%, 7.4 times as much), weight loss (19.2%, 2.9 as much), and walking difficulty (36.34%, 3.4 as much). The geriatric syndrome rates extracted from structured data were substantially lower than published epidemiological rates, although adding the NLP results considerably closed this gap. Claims and structured EHR data give an incomplete picture of burden related to geriatric syndromes. Geriatric syndromes are likely to be missed if unstructured data are not analyzed. Pragmatic NLP algorithms can assist with identifying individuals at high risk of experiencing geriatric syndromes and improving coordination of care for older adults.

BackgroundSuicide is an important public health concern in the United States and around the world. There has been significant work examining machine learning approaches to identify and predict intentional self-harm and suicide using existing data sets. With recent advances in computing, deep learning applications in health care are gaining momentum.ObjectiveThis study aimed to leverage the information in clinical notes using deep neural networks (DNNs) to (1) improve the identification of patients treated for intentional self-harm and (2) predict future self-harm events.MethodsWe extracted clinical text notes from electronic health records (EHRs) of 835 patients with International Classification of Diseases (ICD) codes for intentional self-harm and 1670 matched controls who never had any intentional self-harm ICD codes. The data were divided into training and holdout test sets. We tested a number of algorithms on clinical notes associated with the intentional self-harm codes using the training set, including several traditional bag-of-words–based models and 2 DNN models: a convolutional neural network (CNN) and a long short-term memory model. We also evaluated the predictive performance of the DNNs on a subset of patients who had clinical notes 1 to 6 months before the first intentional self-harm event. Finally, we evaluated the impact of a pretrained model using Word2vec (W2V) on performance.ResultsThe area under the receiver operating characteristic curve (AUC) for the CNN on the phenotyping task, that is, the detection of intentional self-harm in clinical notes concurrent with the events was 0.999, with an F1 score of 0.985. In the predictive task, the CNN achieved the highest performance with an AUC of 0.882 and an F1 score of 0.769. Although pretraining with W2V shortened the DNN training time, it did not improve performance.ConclusionsThe strong performance on the first task, namely, phenotyping based on clinical notes, suggests that such models could be used effectively for surveillance of intentional self-harm in clinical text in an EHR. The modest performance on the predictive task notwithstanding, the results using DNN models on clinical text alone are competitive with other reports in the literature using risk factors from structured EHR data.

Electronic health records (EHRs) with clinical decision support tools are now ubiquitous in healthcare organizations. Clinical foundation models (CFMs) pretrained on large-scale, heterogeneous structured EHR data have emerged as a powerful approach to improve predictive performance and generalizability. Meanwhile, large language models (LLMs) pretrained on broad data sources are being applied to an expanding range of healthcare tasks. However, it remains unclear whether generalist LLMs can match specialized CFMs for disease risk prediction using structured clinical data. We compared CFMs (Med-BERT, CLMBR) against fine-tuned generalist LLMs (Mistral, LLaMA-2/3/3.1), a clinical LLM (Me-LLaMA), and LLM-generated embeddings paired with simple classifiers (using DeepSeek, Qwen3, and GPT-OSS) on two disease risk prediction tasks: heart failure risk among diabetic patients (DHF) and pancreatic cancer diagnosis (PaCa). Evaluations spanned multi-site EHR data, claims data, and an open-source single-institution benchmark (EHRSHOT). Performance was assessed using the area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPRC). On larger EHR and claims cohorts (>30,000 patients), fine-tuned CFMs outperformed fine-tuned LLMs by a small but statistically significant margin (<1% AUROC). The clinical LLM performed comparably to generalist LLMs despite being smaller. On the open-source PaCa cohort (3,810 patients, 199 cases), LLMs achieved slightly higher AUROCs that were not statistically significant (LLaMA-3.1-70B 86.1% vs. Med-BERT 85.3%, p=0.27), but CFMs achieved significantly higher AUPRC (Med-BERT 55.9% vs. LLaMA-3.1-70B 41.1%, p=0.001). Notably, LLM-generated trajectory embeddings paired with logistic regression or a simple MLP, without any LLM fine-tuning, achieved the best overall performance, with AUROC exceeding 90% (Qwen3) and AUPRC reaching 66% (GPT-OSS 20B). LLM-generated embeddings with lightweight classifiers outperformed both fine-tuned CFMs and fine-tuned LLMs on AUROC and AUPRC. While these results demonstrate the potential of generalist models to match or surpass specialized CFMs, their substantially greater computational cost and variable AUPRC performance in the fine-tuning setting warrant caution. We provide a reproducible evaluation framework and codebase to support continued benchmarking.

Background. There is growing evidence that social and behavioral determinants of health (SBDH) play a substantial effect in a wide range of health outcomes. Electronic health records (EHRs) have been widely employed to conduct observational studies in the age of artificial intelligence (AI). However, there has been limited review into how to make the most of SBDH information from EHRs using AI approaches.Methods. A systematic search was conducted in six databases to find relevant peer-reviewed publications that had recently been published. Relevance was determined by screening and evaluating the articles. Based on selected relevant studies, a methodological analysis of AI algorithms leveraging SBDH information in EHR data was provided.Results. Our synthesis was driven by an analysis of SBDH categories, the relationship between SBDH and healthcare-related statuses, natural language processing (NLP) approaches for extracting SBDH from clinical notes, and predictive models using SBDH for health outcomes.Discussion. The associations between SBDH and health outcomes are complicated and diverse; several pathways may be involved. Using NLP technology to support the extraction of SBDH and other clinical ideas simplifies the identification and extraction of essential concepts from clinical data, efficiently unlocks unstructured data, and aids in the resolution of unstructured data-related issues.Conclusion. Despite known associations between SBDH and diseases, SBDH factors are rarely investigated as interventions to improve patient outcomes. Gaining knowledge about SBDH and how SBDH data can be collected from EHRs using NLP approaches and predictive models improves the chances of influencing health policy change for patient wellness, ultimately promoting health and health equity.