ObjectivesAccurate characterization of patients with congenital heart disease is fundamental to research, outcomes reporting, quality improvement, and clinical decision-making. Here we present an approach to computing the anatomy of patients with congenital heart disease based on the whole of their diagnostic and surgical codes.Materials and MethodsAll diagnostic and procedure codes for patients cared for between 1981 and 2020 at Boston Children’s Hospital were extracted from a database containing diagnostic codes from echocardiograms, and procedural codes from surgical and catheterization procedures. The pipeline sequentially (1) mapped each of the 7500 native codes to algorithm codes; (2) computed the parent anatomy for each study using a pre-defined hierarchy; (3) computed the parent anatomy for the patient, based on highest ranking parent anatomy; and (4) computed the subcategories and mandatory co-variate findings for each patient. Thereafter, diagnostic accuracy of 500 unseen patients was adjudicated against clinical documentation by clinical experts.ResultsA total of 514 541 echocardiograms on 161 735 patients were available for this study. Phenotypes of congenital cardiac diseases were assigned in 84 285 patients (52%), and the remainder were computed to have normal anatomy. Clinicians agreed with algorithm assignments in 96.4% (482 of 500 patients), with disagreements most often representing definitional differences. An interactive dashboard enabled by the output of this algorithm is presented.ConclusionsThe computation of detailed congenital heart defect phenotypes from raw diagnostic and procedure codes is possible with a high degree of accuracy and efficiency. This framework may enable tools to support interactive outcomes reporting and clinical decision support.

ObjectivesAcute chest pain is a common presentation in the emergency department, characterized by sudden onset with high morbidity and mortality. Traditional diagnostic methods, such as computed tomography (CT) and CT angiography (CTA), are often time-consuming and fail to meet the urgent need for rapid triage in emergency settings.Materials and MethodsWe developed a multimodal model that integrates Bio-ClinicalBERT and ensemble learning (AdaBoost, Gradient boosting, and XGBoost) based on 41 382 patient data from April 1, 2013 to April 1, 2025 at Chongqing Daping Hospital. By integrating clinical texts and laboratory indicators, the model aims to classify the 3 major causes of fatal chest pain (acute coronary syndrome, pulmonary embolism, and aortic dissection), as well as other causes of chest pain, aiding rapid triage. We adopt strict data preprocessing and rank importance feature selection.ResultsThe multimodal fusion model based on Gradient boosting exhibits the best performance: accuracy of 88.40%, area under the curve of 0.951, F1-score of 74.56%, precision of 77.50%, and recall of 72.52%. SHapley Additive exPlanations (SHAP) analysis confirmed the clinical relevance of key features such as d-dimer and high-sensitivity troponin. When reducing the number of numerical features to 30 key indicators, the model enhanced robustness without compromising performance.Discussion and ConclusionWe developed an artificial intelligence model for chest pain classification that effectively addresses the problem of overlapping clinical symptoms through multimodal fusion, and the model has high accuracy. However, future work needs to better integrate model development with clinical workflows and practical constraints.

ObjectiveMedical examiners and coroners (ME/C) oversee medicolegal death investigations which determine causes of death and other contextual factors that may have influenced a death. We utilize open data releases from ME/C offices covering 6 different geographic areas to demonstrate the strengths and limitations of ME/C data for forensic epidemiology research.Materials and MethodsWe use our novel geoPIPE tool to establish a pipeline that (a) automates ingesting open data releases, (b) geocodes records where possible to yield a spatial component, (c) enhances data with variables useful for overdose research, such as flagging substances contributing to each death, and (d) publishes the enriched data to our open repository. We use results from this pipeline to highlight similarities and differences of overdose data across different sources.ResultsText processing to extract drugs contributing to each death yielded compatible data across all locations. Conversely, geospatial analyses are sometimes incompatible due to differences in available geographic resolution, which range from fine-grain latitude and longitude coordinates to larger regions identified by zip codes. Our pipeline pushes weekly results to an open repository.DiscussionOpen ME/C data are highly useful for research on substance use disorders; our visualizations demonstrate the ability to contextualize overdose data within and across specific geographic regions. Furthermore, the spatial component of our results enables clustering of overdose events and accessibility studies for resources related to preventing overdose deaths.ConclusionsGiven the utility to public health researchers, we advocate that other ME/C offices explore releasing open data and for policy makers to support and fund transparency efforts.

Abstract Objective To compare clinical summaries generated from simulated patient primary care electronic health records (EHRs) by GPT-4, to summaries generated by clinicians on multiple domains of quality including utility, concision, accuracy, and bias. Materials and Methods Seven primary care physicians generated 70 simulated patient EHR notes, each representing 10 patient contacts with the practice over at least 2 years. Each record was summarized by a different clinician and by GPT-4. artificial intelligence (AI)- and clinician-authored summaries were rated blind by clinicians according to 8 domains of quality and an overall rating. Results The median time taken for a clinician to read through and assimilate the information in the EHRs before summarizing, was 7 minutes. Clinicians rated clinician-authored summaries higher than AI-authored summaries overall (7.39 vs 7.00 out of 10; P = .02), but with greater variability in clinician-authored summary ratings. AI and clinician-authored summaries had similar accuracy and AI-authored summaries were less likely to omit important information and more likely to use patient-friendly language. Discussion Although AI-authored summaries were rated slightly lower overall compared with clinician-authored summaries, they demonstrated similar accuracy and greater consistency. This demonstrates potential applications for generating summaries in primary care, particularly given the substantial time taken for clinicians to undertake this work. Conclusion The results suggest the feasibility, utility and acceptability of using AI-authored summaries to integrate into EHRs to support clinicians in primary care. AI summarization tools have the potential to improve healthcare productivity, including by enabling clinicians to spend more time on direct patient care.

ObjectivesTo develop and test a novel machine learning approach for monitoring impact of computerized clinical decision support (CDS) tools on clinicians’ electronic health record (EHR) activities.Materials and MethodsOur CDS monitoring approach leverages topic modeling, a latent-variable statistical machine learning method, to infer health providers’ EHR activities from EHR audit logs. We applied this approach to monitor the impact of a tobacco cessation support CDS tool newly implemented in 5 cancer clinics (2018-2021). We trained the topic model on EHR audit log data from 3445 encounters (pre-CDS-implementation: 1734, post-CDS-implementation: 1711) for patients with active smoking status. The number of topics was automatically determined based on within-topic coherence and across-topic divergence, and the identified topics were assigned clinically relevant EHR activity labels by 4 domain experts.ResultsThe topic model identified 2 distinct activities focusing on CDS (act on CDS, bypass/postpone CDS), 2 activities related to CDS (review patient records and address alerts, use note templates and acknowledge the completion of CDS), 6 related to accessing (access patient station) and reviewing patient data (external records, synopsis data, snapshot of patient data, problem list/diagnosis/notes, treatment plan), and 4 related to modifying EHR (modify diagnosis/problem lists, document visit with record review, perform administrative activities for visit and billing, and document follow-up care plan). Comparing matched 1-hour after-check-in windows post-implementation (n = 841) versus pre-implementation (n = 841) of CDS, the mean prevalence (expressed as proportions out of 1.0) of providers’ EHR-use activity increased on CDS-focused activities (0.073, 95% CI, 0.066-0.079) and CDS-related activities (0.098, 95% CI, 0.089-0.106) and decreased on modifying EHR (−0.113, 95% CI, −0.124 to −0.102) and reviewing patient data (−0.058, 95% CI, −0.072 to −0.044).DiscussionOur topic model-based CDS monitoring approach can identify shifts in prevalence of EHR-use activities pre-implementation versus post-implementation. This approach can be applied to detect unintended changes in EHR activities on a large population scale following CDS implementation, providing valuable insights to guide focused qualitative investigations for CDS improvement or de-implementation.ConclusionOur approach offers a scalable, data-driven framework for evaluating the real-world impact of EHR-embedded CDS tools. Built on a generic machine learning framework, this approach could be adapted to explore impact of other healthcare quality improvement strategies using EHR-integrated CDS interventions.

BackgroundStudies on symptom concordance between patients and their caregivers often use cross-sectional designs, which may fail to capture the longitudinal, dynamic symptom experience. The Behavioral and Environmental Sensing and Intervention for Cancer (BESI-C) is a remote health monitoring system that utilizes smartwatches and ecological momentary assessments (EMAs) to empower patients and caregivers to monitor and manage cancer pain at home. BESI-C collects real-time symptom data in naturalistic settings, enabling longitudinal tracking and analysis of symptom patterns over time.ObjectiveTo define and examine dyadic concordance using participant-initiated symptom reports collected via remote health monitoring.MethodsDyads of patients with advanced cancer and their family caregivers were recruited to use BESI-C for 2 weeks, reporting pain in real time through EMAs. We used Bangdiwala’s B statistic to determine the concordance of patient-reported pain and caregiver-reported perceived patient pain under different contextual criteria (eg, co-location of participants; user engagement with BESI-C) that we hypothesized would impact concordance. We also explored a hypothesis that concordance would improve between study week 1 versus week 2.ResultsData from 21 patient-caregiver dyads were used for analysis. The reporting of pain events was highly variable between patients and their caregivers. Concordance of pain reporting improved when patients and caregivers were co-located and both wearing their BESI-C smartwatches. We did not observe consistent patterns in patient-caregiver concordance between week 1 and week 2.ConclusionWe propose an analytical approach to define and evaluate concordance between patients’ and caregivers’ real-time symptom reports that can be applied to dyadic, longitudinal symptom data collected using remote health monitoring. Future work should examine the relationship between patient-caregiver symptom concordance with key quality-of-life metrics and sociodemographic factors that impact participant engagement with remote health monitoring technologies.

ObjectiveConversational Health Agents (CHAs) are interactive systems providing healthcare services, such as assistance and diagnosis. Current CHAs, especially those utilizing Large Language Models (LLMs), primarily focus on conversation aspects. However, they offer limited agent capabilities, specifically needing more multistep problem-solving, personalized conversations, and multimodal data analysis. We aim to overcome these limitations.Materials and methodsWe propose openCHA, an open-source LLM-powered framework, designed to enable the development of conversational agents. OpenCHA offers a foundational and structured architecture and codebase, enabling researchers and developers to build and customize their CHA based on the specifics of their intended application. The framework leverages knowledge acquisition, problem-solving capabilities, multilingual, and multimodal conversations, and allows interaction with various AI platforms. We have released the framework as open source for the community on GitHub (https://github.com/Institute4FutureHealth/CHA and https://opencha.com).ResultsWe demonstrated the openCHA’s capability to develop CHAs across multiple health domains using 2 demos and 5 use cases. In diabetic patient management, developed CHA achieved a 92.1% accuracy rate, surpassing GPT4’s 51.8%. In food recommendations, developed CHA outperformed GPT4. The developed CHA excelled as an evaluator for mental health chatbots, recording the lowest Mean Absolute Error at 0.31, compared to competitors like GPT, Misteral, Gemini, and Claude. Additionally, the empathy enabled CHA identified emotional states with 89% accuracy, and in physiological data analysis of heart rate from Photoplethysmography (PPG) signals, the developed CHA achieved an mean absolute error of 2.83, far lower than GPT-4o’s 8.93.DiscussionThe openCHA framework enhances CHAs by enabling features such as explainability, personalization, and reliability through its integration with LLMs and external data sources. The developed CHAs face challenges like latency, token limits, and scalability. Future efforts will focus on improving planning robustness, enhancing accuracy and evaluation methods, and resolving user query ambiguity to further refine the framework’s effectiveness.ConclusionThe diverse demos and use cases of openCHA demonstrate the framework’s capacity to empower the development of a wide range of CHAs for various healthcare tasks.

Artificial Intelligence (AI) is transforming healthcare by improving diagnostics, treatment recommendations, and resource allocation. However, its implementation also raises ethical concerns, particularly regarding biases in AI algorithms trained on inequitable data, which may reinforce health disparities. This article introduces the AI COmmunity-based Ethical Dialogue and DEcision-making (CODE) framework to embed ethical deliberation into AI development, focusing on Electronic Health Records (EHRs). We propose the AI CODE framework as a structured approach to addressing ethical challenges in AI-driven healthcare and ensuring its implementation supports health equity. The framework outlines 5 steps to advance health equity: (1) Contextual diversity and priority: Ensuring inclusive datasets and that AI reflects the community needs; (2) Sharing ethical propositions: Structured discussions on privacy, bias, and fairness; (3) Dialogic decision-making: Collaboratively with stakeholders to develop AI solutions; (4) Integrating ethical solutions: Applying solutions into AI design to enhance fairness; and (5) Evaluating effectiveness: Continuously monitoring AI to address emerging biases. We examine the framework's role in mitigating AI biases through structured community engagement and its relevance within evolving healthcare policies. While the framework promotes ethical AI integration in healthcare, it also faces challenges in implementation. The framework provides practical guidance to ensure AI systems are ethical, community-driven, and aligned with health equity goals.

ObjectivesTo evaluate clinical decision support (CDS) of varying complexities and care settings represented using Health Information Technology (HIT) standards—Clinical Quality Language (CQL) for representing clinical logic and Fast Healthcare Interoperability Resources (FHIR) for health information exchange.Materials and MethodsThis Institutional Review Board-approved, retrospective study was performed at an academic medical center (January 1, 2023-December 31, 2023). Recommendations extracted from patient-centered outcomes guidelines were translated into standardized syntax (SNOMED CT) and representations (CQL, FHIR). Clinical decision support Hooks applications were developed for: CDS1—provides education for emergency department (ED) patients with venous thromboembolism; CDS2—recommends CT pulmonary angiogram in ED patients with suspected pulmonary embolism (PE) and uses FHIR Questionnaire resources for representing interactive content; CDS3—recommends mammography/breast magnetic resonance imaging surveillance in outpatients with breast cancer history. We randomly selected 50 ED patients with suspected PE and 50 outpatients undergoing breast imaging surveillance. We compared outcomes of false-positive alerts and the accuracy of CDS1, the more complex CDS2, and CDS3 for outpatients.ResultsClinical decision support Hooks applications used CQL logic for trigger expressions and logic files and provided recommendations to ED and outpatient providers. CDS1 had a false-positive alert and accuracy of 11.1% and 98%, respectively, not significantly different from CDS2 (0.0% false-positive alerts, P = .33 and 96% accuracy, P = .56) or from CDS3 (0.0% false-positive alerts, P = .15 and 100% accuracy, P = .31).DiscussionHealth Information Technology standards can represent recommendations of varying complexities in various care settings.ConclusionThe potential to represent CDS using standardized syntax and formats can help facilitate the dissemination of CDS-consumable artifacts.

This work evaluated algorithmic bias in biomarkers classification using electronic pathology reports from female breast cancer cases. Bias was assessed across 5 subgroups: cancer registry, race, Hispanic ethnicity, age at diagnosis, and socioeconomic status. We utilized 594875 electronic pathology reports from 178121 tumors diagnosed in Kentucky, Louisiana, New Jersey, New Mexico, Seattle, and Utah to train 2 deep-learning algorithms to classify breast cancer patients using their biomarkers test results. We used balanced error rate (BER), demographic parity (DP), equalized odds (EOD), and equal opportunity (EOP) to assess bias. We found differences in predictive accuracy between registries, with the highest accuracy in the registry that contributed the most data (Seattle Registry, BER ratios for all registries >1.25). BER showed no significant algorithmic bias in extracting biomarkers (estrogen receptor, progesterone receptor, human epidermal growth factor receptor 2) for race, Hispanic ethnicity, age at diagnosis, or socioeconomic subgroups (BER ratio <1.25). DP, EOD, and EOP all showed insignificant results. We observed significant differences in BER by registry, but no significant bias using the DP, EOD, and EOP metrics for socio-demographic or racial categories. This highlights the importance of employing a diverse set of metrics for a comprehensive evaluation of model fairness. A thorough evaluation of algorithmic biases that may affect equality in clinical care is a critical step before deploying algorithms in the real world. We found little evidence of algorithmic bias in our biomarker classification tool. Artificial intelligence tools to expedite information extraction from clinical records could accelerate clinical trial matching and improve care.