Development and preliminary testing of Health Equity Across the AI Lifecycle (HEAAL): A framework for healthcare delivery organizations to mitigate the risk of AI solutions worsening health inequities.

The use of data-driven technologies such as Artificial Intelligence (AI) and Machine Learning (ML) is growing in healthcare. However, the proliferation of healthcare AI tools has outpaced regulatory frameworks, accountability measures, and governance standards to ensure safe, effective, and equitable use. To address these gaps and tackle a common challenge faced by healthcare delivery organizations, a case-based workshop was organized, and a framework was developed to evaluate the potential impact of implementing an AI solution on health equity. The Health Equity Across the AI Lifecycle (HEAAL) is co-designed with extensive engagement of clinical, operational, technical, and regulatory leaders across healthcare delivery organizations and ecosystem partners in the US. It assesses 5 equity assessment domains-accountability, fairness, fitness for purpose, reliability and validity, and transparency-across the span of eight key decision points in the AI adoption lifecycle. It is a process-oriented framework containing 37 step-by-step procedures for evaluating an existing AI solution and 34 procedures for evaluating a new AI solution in total. Within each procedure, it identifies relevant key stakeholders and data sources used to conduct the procedure. HEAAL guides how healthcare delivery organizations may mitigate the potential risk of AI solutions worsening health inequities. It also informs how much resources and support are required to assess the potential impact of AI solutions on health inequities.

Establishing priorities for implementation of large language models in pathology and laboratory medicine.

Artificial intelligence and machine learning have numerous applications in pathology and laboratory medicine. The release of ChatGPT prompted speculation regarding the potentially transformative role of large-language models (LLMs) in academic pathology, laboratory medicine, and pathology education. Because of the potential to improve LLMs over the upcoming years, pathology and laboratory medicine clinicians are encouraged to embrace this technology, identify pathways by which LLMs may support our missions in education, clinical practice, and research, participate in the refinement of AI modalities, and design user-friendly interfaces that integrate these tools into our most important workflows. Challenges regarding the use of LLMs, which have already received considerable attention in a general sense, are also reviewed herein within the context of the pathology field and are important to consider as LLM applications are identified and operationalized.

Large Language Models and Medical Education: Preparing for a Rapid Transformation in How Trainees Will Learn to Be Doctors.

Artificial intelligence has the potential to revolutionize health care but has yet to be widely implemented. In part, this may be because, to date, we have focused on easily predicted rather than easily actionable problems. Large language models (LLMs) represent a paradigm shift in our approach to artificial intelligence because they are easily accessible and already being tested by frontline clinicians, who are rapidly identifying possible use cases. LLMs in health care have the potential to reduce clerical work, bridge gaps in patient education, and more. As we enter this era of healthcare delivery, LLMs will present both opportunities and challenges in medical education. Future models should be developed to support trainees to develop skills in clinical reasoning, encourage evidence-based medicine, and offer case-based training opportunities. LLMs may also change what we continue teaching trainees with regard to clinical documentation. Finally, trainees can help us train and develop the LLMs of the future as we consider the best ways to incorporate LLMs into medical education. Ready or not, LLMs will soon be integrated into various aspects of clinical practice, and we must work closely with students and educators to make sure these models are also built with trainees in mind to responsibly chaperone medical education into the next era.

Gender Disparities in Critical Care Procedure Training of Internal Medicine Residents.

Background: Procedural training is a required competency in internal medicine (IM) residency, yet limited data exist on residents' experience of procedural training. Objectives: We sought to understand how gender impacts access to procedural training among IM residents. Methods: A mixed-methods, explanatory sequential study was performed. Procedure volume for IM residents between 2016 and 2020 was assessed at two large academic residencies (Program A and Program B: 399 residents and 4,020 procedures). Procedural rates and actual versus expected procedure volume by gender were compared, with separate analyses by clinical environment (intensive care unit [ICU] or structured procedural service). Semistructured gender-congruent focus groups were conducted. Topics included identity formation as a proceduralist and the resident procedural learning experience, including perceived gender bias in procedure allocation. Results: Compared with men, women residents performed disproportionately fewer ICU procedures per month at Program A (1.4 vs. 2.7; P < 0.05) but not at Program B (0.36 vs. 0.54; P = 0.23). At Program A, women performed only 47% of ICU procedures, significantly fewer than the 54% they were expected to perform on the basis of their time on ICU rotations (P < 0.001). For equal gender distribution of procedural volume at Program A, 11% of the procedures performed by men would have needed to have been performed by women instead. Gender was not associated with differences in the Program A structured procedural service (53% observed vs. 52% expected; P = 0.935), Program B structured procedural service (40% observed vs. 43% expected; P = 0.174), or in Program B ICUs (33% observed vs. 34% expected; P = 0.656). Focus group analysis identified that women from both residencies perceived that assertiveness was required for procedural training in unstructured learning environments. Residents felt that gender influenced access to procedural opportunities, ability to self-advocate for procedural experience, identity formation as a proceduralist, and confidence in acquiring procedural skills. Conclusion: Gender disparities in access to procedural training during ICU rotations were seen at one institution but not another. There were ubiquitous perceptions that assertiveness was important to access procedural opportunities. We hypothesize that structured allocation of procedures would mitigate disparities by allowing all residents to access procedural training regardless of self-advocacy. Residency programs should adopt structured procedural training programs to counteract inequities.

Profiling mouse brown and white adipocytes to identify metabolically relevant small ORFs and functional microproteins.

Microproteins (MPs) are a potentially rich source of uncharacterized metabolic regulators. Here, we use ribosome profiling (Ribo-seq) to curate 3,877 unannotated MP-encoding small ORFs (smORFs) in primary brown, white, and beige mouse adipocytes. Of these, we validated 85 MPs by proteomics, including 33 circulating MPs in mouse plasma. Analyses of MP-encoding mRNAs under different physiological conditions (high-fat diet) revealed that numerous MPs are regulated in adipose tissue in vivo and are co-expressed with established metabolic genes. Furthermore, Ribo-seq provided evidence for the translation of Gm8773, which encodes a secreted MP that is homologous to human and chicken FAM237B. Gm8773 is highly expressed in the arcuate nucleus of the hypothalamus, and intracerebroventricular administration of recombinant mFAM237B showed orexigenic activity in obese mice. Together, these data highlight the value of this adipocyte MP database in identifying MPs with roles in fundamental metabolic and physiological processes such as feeding.

A Clinical Reasoning-Encoded Case Library Developed through Natural Language Processing.

IMPORTANCE: Case reports that externalize expert diagnostic reasoning are utilized for clinical reasoning instruction but are difficult to search based on symptoms, final diagnosis, or differential diagnosis construction. Computational approaches that uncover how experienced diagnosticians analyze the medical information in a case as they formulate a differential diagnosis can guide educational uses of case reports. OBJECTIVE: To develop a "reasoning-encoded" case database for advanced clinical reasoning instruction by applying natural language processing (NLP), a sub-field of artificial intelligence, to a large case report library. DESIGN: We collected 2525 cases from the New England Journal of Medicine (NEJM) Clinical Pathological Conference (CPC) from 1965 to 2020 and used NLP to analyze the medical terminology in each case to derive unbiased (not prespecified) categories of analysis used by the clinical discussant. We then analyzed and mapped the degree of category overlap between cases. RESULTS: Our NLP algorithms identified clinically relevant categories that reflected the relationships between medical terms (which included symptoms, signs, test results, pathophysiology, and diagnoses). NLP extracted 43,291 symptoms across 2525 cases and physician-annotated 6532 diagnoses (both primary and related diagnoses). Our unsupervised learning computational approach identified 12 categories of medical terms that characterized the differential diagnosis discussions within individual cases. We used these categories to derive a measure of differential diagnosis similarity between cases and developed a website ( universeofcpc.com ) to allow visualization and exploration of 55 years of NEJM CPC case series. CONCLUSIONS: Applying NLP to curated instances of diagnostic reasoning can provide insight into how expert clinicians correlate and coordinate disease categories and processes when creating a differential diagnosis. Our reasoning-encoded CPC case database can be used by clinician-educators to design a case-based curriculum and by physicians to direct their lifelong learning efforts.

Protecting reproductive health information in the post-Roe era: interoperability strategies for healthcare institutions.

On June 24, 2022, the US Supreme Court ended constitutional protections for abortion, resulting in wide variability in access from severe restrictions in many states and fewer restrictions in others. Healthcare institutions capture information about patients' pregnancy and abortion care and, due to interoperability, may share it in ways that expose their providers and patients to social stigma and potential legal jeopardy in states with severe restrictions. In this article, we describe sources of risk to patients and providers that arise from interoperability and specify actions that institutions can take to reduce that risk. Institutions have significant power to define their practices for how and where care is documented, how patients are identified, where data are sent or hosted, and how patients are counseled, and thus should protect patients' privacy and ability to receive medical care that is safe and legal where it is performed.

Characterizing styles of clinical note production and relationship to clinical work hours among first-year residents.

OBJECTIVE: To characterize variation in clinical documentation production patterns, how this variation relates to individual resident behavior preferences, and how these choices relate to work hours. MATERIALS AND METHODS: We used unsupervised machine learning with clinical note metadata for 1265 progress notes written for 279 patient encounters by 50 first-year residents on the Hospital Medicine service in 2018 to uncover distinct note-level and user-level production patterns. We examined average and 95% confidence intervals of median user daily work hours measured from audit log data for each user-level production pattern. RESULTS: Our analysis revealed 10 distinct note-level and 5 distinct user-level production patterns (user styles). Note production patterns varied in when writing occurred and in how dispersed writing was through the day. User styles varied in which note production pattern(s) dominated. We observed suggestive trends in work hours for different user styles: residents who preferred producing notes in dispersed sessions had higher median daily hours worked while residents who preferred producing notes in the morning or in a single uninterrupted session had lower median daily hours worked. DISCUSSION: These relationships suggest that note writing behaviors should be further investigated to understand what practices could be targeted to reduce documentation burden and derivative outcomes such as resident work hour violations. CONCLUSION: Clinical note documentation is a time-consuming activity for physicians; we identify substantial variation in how first-year residents choose to do this work and suggestive trends between user preferences and work hours.

Estimation of Surgical Resident Duty Hours and Workload in Real Time Using Electronic Health Record Data.

OBJECTIVE: To explore the use of electronic health record (EHR) data to estimate surgery resident duty hours and monitor real time workload. DESIGN: Retrospective analysis of resident duty hours logged using a voluntary global positioning system (GPS)-based smartphone application compared to duty hour estimates by an EHR-based algorithm. The algorithm estimated duty hours using EHR activity data and operating room logs. A dashboard was developed through Plan-Do-Study-Act cycles for real-time monitoring of workload. SETTING: Single tertiary/quaternary medical center general surgery residency program with approximately 90 categorical, preliminary, and integrated residents at eight clinical sites. PARTICIPANTS: Categorical, preliminary, and integrated surgery residents of all clinical years who volunteered to pilot a GPS application to track duty hours. RESULTS: Of 2,623 work periods by 59 residents were logged with both methods. EHR-estimated work periods started later than GPS logs (median 0.3 hours, interquartile range [IQR] -0.1 - 0.3); EHR-estimated work periods ended earlier than GPS logs (median 0.1 hours, IQR -0.7 - 0.3); and EHR-estimated duty hour totals were less than totals logged by GPS (median -0.3 hours, IQR -0.8 - +0.1). Overall correlation between weekly duty hours logged by EHR and GPS was 0.79. Correlations between the 2 systems stratified from PGY-1 through PGY-5 were 0.76, 0.64, 0.82, 0.87, and 0.83, respectively. The algorithm identified six 80-hour workweek violations (averaged over 4 weeks), while GPS logs identified 8. EHR-based duty hours and operational data were integrated into a dashboard to enable real time monitoring of resident workloads. CONCLUSIONS: EHR-based estimation of surgical resident duty hours has good correlation with GPS-based assessment of duty hours and identifies most workweek duty hour violations. This approach allows for dynamic workload monitoring and may be combined with operational data to anticipate and prevent duty hour violations, thereby optimizing learning.

Health Equity in Artificial Intelligence and Primary Care Research: Protocol for a Scoping Review.

BACKGROUND: Though artificial intelligence (AI) has the potential to augment the patient-physician relationship in primary care, bias in intelligent health care systems has the potential to differentially impact vulnerable patient populations. OBJECTIVE: The purpose of this scoping review is to summarize the extent to which AI systems in primary care examine the inherent bias toward or against vulnerable populations and appraise how these systems have mitigated the impact of such biases during their development. METHODS: We will conduct a search update from an existing scoping review to identify studies on AI and primary care in the following databases: Medline-OVID, Embase, CINAHL, Cochrane Library, Web of Science, Scopus, IEEE Xplore, ACM Digital Library, MathSciNet, AAAI, and arXiv. Two screeners will independently review all abstracts, titles, and full-text articles. The team will extract data using a structured data extraction form and synthesize the results in accordance with PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews) guidelines. RESULTS: This review will provide an assessment of the current state of health care equity within AI for primary care. Specifically, we will identify the degree to which vulnerable patients have been included, assess how bias is interpreted and documented, and understand the extent to which harmful biases are addressed. As of October 2020, the scoping review is in the title- and abstract-screening stage. The results are expected to be submitted for publication in fall 2021. CONCLUSIONS: AI applications in primary care are becoming an increasingly common tool in health care delivery and in preventative care efforts for underserved populations. This scoping review would potentially show the extent to which studies on AI in primary care employ a health equity lens and take steps to mitigate bias. INTERNATIONAL REGISTERED REPORT IDENTIFIER (IRRID): PRR1-10.2196/27799.

Automating Measurement of Trainee Work Hours.

BACKGROUND: Medical training programs across the country are bound to a set of work hour regulations, generally monitored via self-report. OBJECTIVE: We developed a computational method to automate measurement of intern and resident work hours, which we validated against self-report. DESIGN, SETTING, AND PARTICIPANTS: We included all electronic health record (EHR) access log data between July 1, 2018, and June 30, 2019, for trainees enrolled in the internal medicine training program. We inferred the duration of continuous in-hospital work hours by linking EHR sessions that occurred within 5 hours as "on-campus" work and further accounted for "out-of-hospital" work which might be taking place at home. MAIN OUTCOMES AND MEASURES: We compared daily work hours estimated through the computational method with self-report and calculated the mean absolute error between the two groups. We used the computational method to estimate average weekly work hours across the rotation and the percentage of rotations where average work hours exceed the 80-hour workweek. RESULTS: The mean absolute error between self-reported and EHR-derived daily work hours for first- (PGY-1), second- (PGY-2), and third- (PGY-3) year trainees were 1.27, 1.51, and 1.51 hours, respectively. Using this computational method, we estimated average (SD) weekly work hours of 57.0 (21.7), 69.9 (12.2), and 64.1 (16.3) for PGY-1, PGY-2, and PGY-3 residents. CONCLUSION: EHR log data can be used to accurately approximate self-report of work hours, accounting for both in-hospital and out-of-hospital work. Automation will reduce trainees' clerical work, improve consistency and comparability of data, and provide more complete and timely data that training programs need.

Whole transcriptome analysis and validation of metabolic pathways in subcutaneous adipose tissues during FGF21-induced weight loss in non-human primates.

Fibroblast growth factor 21 (FGF21) induces weight loss in mouse, monkey, and human studies. In mice, FGF21 is thought to cause weight loss by stimulating thermogenesis, but whether FGF21 increases energy expenditure (EE) in primates is unclear. Here, we explore the transcriptional response and gene networks active in adipose tissue of rhesus macaques following FGF21-induced weight loss. Genes related to thermogenesis responded inconsistently to FGF21 treatment and weight loss. However, expression of gene modules involved in triglyceride (TG) synthesis and adipogenesis decreased, and this was associated with greater weight loss. Conversely, expression of innate immune cell markers was increased post-treatment and was associated with greater weight loss. A lipogenesis gene module associated with weight loss was evaluated by testing the function of member genes in mice. Overexpression of NRG4 reduced weight gain in diet-induced obese mice, while overexpression of ANGPTL8 resulted in elevated TG levels in lean mice. These observations provide evidence for a shifting balance of lipid storage and metabolism due to FGF21-induced weight loss in the non-human primate model, and do not fully recapitulate increased EE seen in rodent and in vitro studies. These discrepancies may reflect inter-species differences or complex interplay of FGF21 activity and counter-regulatory mechanisms.

Decreasing Hypoglycemia following Insulin Administration for Inpatient Hyperkalemia.

BACKGROUND: Acute hyperkalemia (serum potassium ≥ 5.1 mEq/L) is often treated with a bolus of IV insulin. This treatment may result in iatrogenic hypoglycemia (glucose < 70 mg/dl). OBJECTIVES: The aims of this study were to accurately determine the frequency of iatrogenic hypoglycemia following insulin treatment for hyperkalemia, and to develop an electronic health record (EHR) orderset to decrease the risk for iatrogenic hypoglycemia. DESIGN: This study was an observational, prospective study. SETTING: The setting for this study was a university hospital. PATIENTS: All nonobstetric adult inpatients in all acute and intensive care units were eligible. INTERVENTION: Implementation of a hyperkalemia orderset (Orderset 1.1) with glucose checks before and then one, two, four, and six hours after regular intravenous insulin administration. Based on the results from Orderset 1.1, Orderset 1.2 was developed and introduced to include weight-based dosing of insulin options, alerts identifying patients at higher risk of hypoglycemia, and tools to guide decision-making based on the preinsulin blood glucose level. MEASUREMENTS: Patient demographics, weight, diabetes history, potassium level, renal function, and glucose levels were recorded before, and then glucose levels were measured again at one, two, four, and six hours after insulin was administered. RESULTS: The iatrogenic hypoglycemia rate identified with mandatory glucose checks in Orderset 1.1 was 21%; 92% of these occurred within three hours posttreatment. Risk factors for hypoglycemia included decreased renal function (serum creatinine >2.5 mg/dl), a high dose of insulin (>0.14 units/kg), and re-treatment with blood glucose < 140 mg/dl. After the introduction of Orderset 1.2, the rate of iatrogenic hypoglycemia decreased to 10%. CONCLUSIONS: The use of an EHR orderset for treating hyperkalemia may reduce the risk of iatrogenic hypoglycemia in patients receiving insulin while still adequately lowering their potassium.

Feasibility and cost analysis of day 4 granulocyte colony-stimulating factor mobilized peripheral blood progenitor cell collection from HLA-matched sibling donors.

BACKGROUND: Guidelines recommend treatment with 4-5 days of granulocyte colony-stimulating factor (G-CSF) for optimal donor peripheral blood progenitor cell (PBPC) mobilization followed by day 5 collection. Given that some autologous transplant recipients achieve adequate collection by day 4 and the possibility that some allogeneic donors may maximally mobilize PBPC before day 5, a feasibility study was performed evaluating day 4 allogeneic PBPC collection. METHODS: HLA-matched sibling donors underwent collection on day 4 of G-CSF for peripheral blood (PB) CD34+ counts ≥0.04 × 106/mL, otherwise they underwent collection on day 5. Those with inadequate collected CD34+ cells/kg recipient weight underwent repeat collection over 2 days. Transplant and PBPC characteristics and cost analysis were compared with a historical cohort collected on day 5 per our prior institutional algorithm. RESULTS: Of the 101 patient/donor pairs, 50 (49.5%) had adequate PBPC collection on day 4, with a median PB CD34+ cell count of 0.06 × 106/mL. Day 4 donors were more likely to develop bone pain and require analgesics. Median collected CD34+ count was significantly greater, whereas total nucleated, mononuclear and CD3+ cell counts were significantly lower, at time of transplant infusion for day 4 versus other collection cohorts. There were no significant differences in engraftment or graft-versus-host disease. Cost analysis revealed 6.7% direct cost savings for day 4 versus historical day 5 collection. DISCUSSION: Day 4 PB CD34+ threshold of ≥0.04 × 106/mL identified donors with high likelihood of adequate PBPC collection. Day 4 may be the optimal day of collection for healthy donors, without adverse effect on recipient transplant outcomes and with expected cost savings.

Automated and flexible identification of complex disease: building a model for systemic lupus erythematosus using noisy labeling.

Accurate and efficient identification of complex chronic conditions in the electronic health record (EHR) is an important but challenging task that has historically relied on tedious clinician review and oversimplification of the disease. Here we adapt methods that allow for automated "noisy labeling" of positive and negative controls to create a "silver standard" for machine learning to automate identification of systemic lupus erythematosus (SLE). Our final model, which includes both structured data as well as text processing of clinical notes, outperformed all existing algorithms for SLE (AUC 0.97). In addition, we demonstrate how the probabilistic outputs of this model can be adapted to various clinical needs, selecting high thresholds when specificity is the priority and lower thresholds when a more inclusive patient population is desired. Deploying a similar methodology to other complex diseases has the potential to dramatically simplify the landscape of population identification in the EHR. MeSH terms: Electronic Health Records, Machine Learning, Lupus Erythematosus, Phenotype, Algorithms.

Novel Withdrawal Symptoms of a Neonate Prenatally Exposed to a Fentanyl Analog.

Neonatal abstinence syndrome (NAS) is a withdrawal syndrome observed in neonates exposed to drugs in utero, typically opioids, which is associated with symptoms affecting the central and autonomic nervous systems and the gastrointestinal system. West Virginia, particularly the southeastern region of the state, has remarkably higher rates of NAS than similar communities. Our facility is increasingly faced with complex cases of NAS caused by in utero exposure to multiple substances. We present a case report of a neonate born to a 25-year-old mother enrolled in a medication-assisted treatment program for substance use disorder who was noncompliant in prenatal care, using multiple substances throughout the pregnancy, including gabapentin and fentanyl. After birth, the neonate began to exhibit unusual withdrawal symptoms including arching, tongue thrusting, and irregular eye movements, which are typically associated with in utero gabapentin exposure. The parents denied consent to treat with gabapentin, the suggested management for these symptoms; thus, a treatment protocol for methadone and clonidine were followed. This case exemplifies the medical and social complexities involved in treating polysubstance exposure-associated NAS.