Keith E. Morse

Background: Health systems answer most questions by having expert analysts hand-write queries against a complex electronic health record data warehouse, a slow, resource-intensive process. Whether an autonomous coding agent can do this accurately is unknown. Methods: In a single-center quality-improvement evaluation, we posed ten questions about a common pediatric infection, acute otitis media. The questions went to analysts, whose adjudicated answers were the reference, and to an autonomous coding agent (OpenAI Codex), which wrote and ran read-only queries on a full copy of the production data warehouse (Epic Caboodle). We ran the agent under four conditions: an autonomous baseline (each question answered three times at two reasoning-effort settings), a variant in which it listed its assumptions, interactive analyst feedback, and reuse of a corrected definition across related questions. Outcomes were accuracy, patient-level agreement (F1), reproducibility, and efficiency (time and tokens). Results: Working autonomously, the agent wrote valid queries and never fabricated data, but rarely produced the exact answer. At medium effort, it came within 5% of the reference on 27 of 30 runs but exact on only 10. Higher effort produced no improvement. Reproducibility was the greater weakness, with the three runs returning an identical answer on only 3 of 10 questions. When a run matched the reference, it had found the same patients (F1 = 1.00); the exception was an over-counted procedure (F1 = 0.72). With analyst feedback on four questions, it answered two exactly, and reusing the corrected definition on related questions restored reproducibility and accuracy. Conclusions: A general-purpose coding agent matched an adjudicated analyst reference on most routine questions but not reproducibly, because key definitions depended on warehouse knowledge that the data dictionary omits. Supplying that knowledge, by prompt or feedback, restored reproducibility. Under expert analyst supervision, the agent is already a capable drafting aid and a promising step toward broader hospital analytic support.

Health systems are gaining early experience with artificial intelligence (AI) as AI tools are implemented in hospitals. AI-savvy leaders are needed to oversee these tools and clinical informatics is an ideal background for such leaders.[1] This letter highlights the relevant expertise of board-certified clinical informaticists (BCIs) and offers recommendations for how organizations can leverage this expertise to ensure their AI adoption aligns with the “FAVES” (fair, appropriate, valid, effective, safe) principles endorsed by US Department of Health and Human Services (HHS).[2] Clinical informatics, officially recognized by the American Board of Medical Specialties (ABMS) in 2011, centers on using data and technology to enhance patient care.[3] Specialized training through a 2-year clinical informatics fellowship is now offered at over 60 institutions nationally.[4] This subspecialty demands proficiency in areas such as health information systems, data governance, and change management. While its practitioners have historically been tapped to lead the implementation of electronic health record (EHR) systems, their skill set makes them uniquely qualified to lead the successful adoption of AI technologies. These skills are outlined explicitly in requirements for training[5] [6] [7] and board certification,[8] [9] as well as evidenced in the real-world experience of its practitioners.[10] [11] Critically, BCIs are among the few medical professionals trained in the fundamentals of AI and data science. AI is a core content area on board examinations and a perennial focus at national conferences. A recent survey showed that, during their 2 years of fellowship, 37% of trainees built AI tools themselves.[11] This experience in AI is augmented by expertise in the health data fed into models. BCIs understand the ontologies and coding schema (e.g., SNOMED, LOINC) that organize the data, the database systems that store it, and the interoperability building blocks (e.g., FHIR, HL7) that move it. BCIs are adept at designing and implementing clinical workflows to integrate new technologies. As clinicians, they understand the frustration of poorly designed interventions and can develop AI systems that enhance rather than disrupt existing clinical processes.[12] Once implemented, they can apply best practices from EHR optimization by monitoring for common issues arising from workflow disruptions, alert fatigue, or automation bias. Finally, they can evaluate the AI system's ultimate impact on the care process, quantifying its overall impact on patient outcomes and provider experience.[13] More broadly, BCIs bring essential leadership skills needed to guide health care organizations through AI integration. Deployment relies heavily on successful change management and BCIs are skilled in managing the human elements of technology adoption, fostering stakeholder buy-in through interdisciplinary communication, comprehensive training, and feedback collection. They are uniquely positioned to lead multidisciplinary teams bridging gaps among clinicians, data scientists, IT departments, and operational leaders. The novel regulatory and ethical challenges posed by AI require coordination among disciplines, with health policy, compliance, and ethical experts now vital contributors to AI governance.[14] Experts advocate for a more defined role for BCIs in AI regulatory oversight, such as at a recent FDA Digital Health Advisory Committee meeting, calling for formal clinical informaticist review of model safety information prior to integration into patient care.[15] To take advantage of BCIs skill sets, health system leaders should consider recruitment, new role creation, and ongoing training. To recruit BCIs, consider both dedicated efforts to hire qualified BCIs and including “clinical informatics” as a desirable ancillary experience in traditional clinical job postings. Interested early career physicians can be encouraged (and funded) to complete clinical informatics fellowship and board certification. And finally, define new, clear job descriptions in which the BCI will serve. New roles must be created that both reflect the organization's needs and provide a reasonable chance for the candidate's success. Depending on strategic priorities, this role could focus on the development of an AI roadmap, leading AI governance, coordinating the creation of core AI infrastructure, identifying high-priority applications of AI, or evaluating the efficacy of deployed AI tools.[1] It should also lead the efforts to educate and train those expected to use the AI algorithms. Success in these tasks relies heavily on collaboration among medical specialties and health system roles, thus the role should sit at a system-wide level to enable multidisciplinary engagement. Such expansive responsibilities cannot be met without funded support, typically starting at 1 to 2 days per week. This role should report to executive leadership with strong ties to the organization's IT Department, typically the CIO, CMIO, or Chief AI Officer. Finally, because of the rapid evolution of AI's capabilities and its applications, ongoing training opportunities are needed. This comes in the form of regular conference attendance, online coursework, vendor demonstrations, and informal engagement with other AI leaders. In an era where much of the health care workforce needs AI upskilling, BCIs are already well-prepared. These individuals stand ready to lead the comprehensive integration of AI into patient care as well as ensure its ongoing monitoring and evaluation as envisioned by regulators. It is incumbent upon health systems to recruit these leaders to ensure that health care innovation proceeds not by chance, but by deliberate design. ^* co-first authors. Received: 01 December 2024 Accepted: 11 March 2025 Article published online: 02 July 2025 © 2025. Thieme. All rights reserved. Georg Thieme Verlag KG Oswald-Hesse-Straße 50, 70469 Stuttgart, Germany

Background: Patients are important participants in their medical care, yet artificial intelligence (AI) models are used to guide care with minimal patient input. This limitation is made partially worse due to a paucity of rigorous methods to measure and incorporate patient values of the tradeoffs inherent in AI applications. This paper presents AI-VOICE (Values-Oriented Implementation and Context Evaluation), a novel method to collect patient values, or utilities, of the downstream consequences stemming from an AI model’s use to guide care. The results are then used to select the model’s risk threshold, offering a mechanism by which an algorithm can concretely reflect patient values. Methods: The entity being evaluated by AI-VOICE is an AI-informed workflow , which is composed of the patient’s health state, an action triggered by the AI model, and the benefits and harms accrued as a consequence of that action. The utilities of these workflows are measured through a survey-based, standard gamble experiment. These utilities define a patient-specific ratio of the cost of an inaccurate prediction versus the benefits of an accurate one. This ratio is mapped to the receiver-operator-characteristic curve to identify the risk threshold that reflects the patient’s values. The survey instrument is made freely available to researchers through a web-based application. Results: A demonstration of AI-VOICE is provided using a hypothetical sepsis prediction algorithm. Conclusion: AI-VOICE offers an accessible, quantitative method to incorporate patient values into AI-informed healthcare workflows.

Study-specific data quality testing is an essential part of minimizing analytic errors, particularly for studies making secondary use of clinical data. We applied a systematic and reproducible approach for study-specific data quality testing to the analysis plan for PRESERVE, a 15-site, EHR-based observational study of chronic kidney disease in children. This approach integrated widely adopted data quality concepts with healthcare-specific evaluation methods. We implemented two rounds of data quality assessment. The first produced high-level evaluation using aggregate results from a distributed query, focused on cohort identification and main analytic requirements. The second focused on extended testing of row-level data centralized for analysis. We systematized reporting and cataloguing of data quality issues, providing institutional teams with prioritized issues for resolution. We tracked improvements and documented anomalous data for consideration during analyses. The checks we developed identified 115 and 157 data quality issues in the two rounds, involving completeness, data model conformance, cross-variable concordance, consistency, and plausibility, extending traditional data quality approaches to address more complex stratification and temporal patterns. Resolution efforts focused on higher priority issues, given finite study resources. In many cases, institutional teams were able to correct data extraction errors or obtain additional data, avoiding exclusion of 2 institutions entirely and resolving 123 other gaps. Other results identified complexities in measures of kidney function, bearing on the study’s outcome definition. Where limitations such as these are intrinsic to clinical data, the study team must account for them in conducting analyses. This study rigorously evaluated fitness of data for intended use. The framework is reusable and built on a strong theoretical underpinning. Significant data quality issues that would have otherwise delayed analyses or made data unusable were addressed. This study highlights the need for teams combining subject-matter and informatics expertise to address data quality when working with real world data.

Foundation models are transforming artificial intelligence (AI) in healthcare by providing modular components adaptable for various downstream tasks, making AI development more scalable and cost-effective. Foundation models for structured electronic health records (EHR), trained on coded medical records from millions of patients, demonstrated benefits including increased performance with fewer training labels, and improved robustness to distribution shifts. However, questions remain on the feasibility of sharing these models across hospitals and their performance in local tasks. This multi-center study examined the adaptability of a publicly accessible structured EHR foundation model (FM_SM), trained on 2.57 M patient records from Stanford Medicine. Experiments used EHR data from The Hospital for Sick Children (SickKids) and Medical Information Mart for Intensive Care (MIMIC-IV). We assessed both adaptability via continued pretraining on local data, and task adaptability compared to baselines of locally training models from scratch, including a local foundation model. Evaluations on 8 clinical prediction tasks showed that adapting the off-the-shelf FM_SMmatched the performance of gradient boosting machines (GBM) locally trained on all data while providing a 13% improvement in settings with few task-specific training labels. Continued pretraining on local data showed FM_SMrequired fewer than 1% of training examples to match the fully trained GBM’s performance, and was 60 to 90% more sample-efficient than training local foundation models from scratch. Our findings demonstrate that adapting EHR foundation models across hospitals provides improved prediction performance at less cost, underscoring the utility of base foundation models as modular components to streamline the development of healthcare AI.

To characterize the complex interplay between multiple clinical conditions in a time-to-event analysis framework using data from multiple hospitals, we developed two novel one-shot distributed algorithms for competing risk models (ODACoR). By applying our algorithms to the EHR data from eight national children’s hospitals, we quantified the impacts of a wide range of risk factors on the risk of post-acute sequelae of SARS-COV-2 (PASC) among children and adolescents. Our ODACoR algorithms are effectively executed due to their devised simplicity and communication efficiency. We evaluated our algorithms via extensive simulation studies as applications to quantification of the impacts of risk factors for PASC among children and adolescents using data from eight children’s hospitals including the Children’s Hospital of Philadelphia, Cincinnati Children’s Hospital Medical Center, Children’s Hospital of Colorado covering over 6.5 million pediatric patients. The accuracy of the estimation was assessed by comparing the results from our ODACoR algorithms with the estimators derived from the meta-analysis and the pooled data. The meta-analysis estimator showed a high relative bias (∼40%) when the clinical condition is relatively rare (∼0.5%), whereas ODACoR algorithms exhibited a substantially lower relative bias (∼0.2%). The estimated effects from our ODACoR algorithms were identical on par with the estimates from the pooled data, suggesting the high reliability of our federated learning algorithms. In contrast, the meta-analysis estimate failed to identify risk factors such as age, gender, chronic conditions history, and obesity, compared to the pooled data. Our proposed ODACoR algorithms are communication-efficient, highly accurate, and suitable to characterize the complex interplay between multiple clinical conditions. Our study demonstrates that our ODACoR algorithms are communication-efficient and can be widely applicable for analyzing multiple clinical conditions in a time-to-event analysis framework.

Development of electronic health records (EHR)-based machine learning models for pediatric inpatients is challenged by limited training data. Self-supervised learning using adult data may be a promising approach to creating robust pediatric prediction models. The primary objective was to determine whether a self-supervised model trained in adult inpatients was noninferior to logistic regression models trained in pediatric inpatients, for pediatric inpatient clinical prediction tasks. This retrospective cohort study used EHR data and included patients with at least one admission to an inpatient unit. One admission per patient was randomly selected. Adult inpatients were 18 years or older while pediatric inpatients were more than 28 days and less than 18 years. Admissions were temporally split into training (January 1, 2008 to December 31, 2019), validation (January 1, 2020 to December 31, 2020), and test (January 1, 2021 to August 1, 2022) sets. Primary comparison was a self-supervised model trained in adult inpatients versus count-based logistic regression models trained in pediatric inpatients. Primary outcome was mean area-under-the-receiver-operating-characteristic-curve (AUROC) for 11 distinct clinical outcomes. Models were evaluated in pediatric inpatients. When evaluated in pediatric inpatients, mean AUROC of self-supervised model trained in adult inpatients (0.902) was noninferior to count-based logistic regression models trained in pediatric inpatients (0.868) (mean difference = 0.034, 95% CI=0.014-0.057; P < .001 for noninferiority and P = .006 for superiority). Self-supervised learning in adult inpatients was noninferior to logistic regression models trained in pediatric inpatients. This finding suggests transferability of self-supervised models trained in adult patients to pediatric patients, without requiring costly model retraining.

Prior work has shown that large language models like GPT-4 and Med-PaLM 2 can answer sample questions from the USMLE Step 2 Clinical Knowledge (CK) exam with greater than 80% accuracy. But can these generative AI create USMLE-like exam questions? This capability could augment humans in writing or preparing for such exams. Here we assess the ability of GPT-4 to generate realistic exam questions by asking licensed physicians to (1) distinguish AI-generated questions from genuine USMLE Step 2 CK questions, and (2) assess the validity of AI-generated questions and answers. We find that GPT-4 can generate question/answer pairs that are largely indistinguishable from human-generated ones, with a majority (64%) deemed “valid” by a panel of licensed physicians.