JAMIA Open最新文献

Objectives: The study aimed to develop a deep learning-based model, using global and local explainability methods, to process clinical data collected in community pharmacies and identify the key variables influence health-related quality of life in patients with chronic diseases.

Materials and methods: Data from 347 chronic patients, including 257 variables, were analyzed. Five predictive models were compared using 10-way stratified cross-validation: Gradient Boosting, Random Forest, LightGBM, a fully connected neural network (FCNN), and a set of 5 FCNNs. For interpretability, SHapley Additive exPlanations (SHAP) was used for the global importance of variables and Local Interpretable Model-Agnostic Explanations (LIME) for the local interpretation of individual cases.

Results: The FCNN ensemble achieved the best performance (R ² = 0.511 ± 0.126; 95% CI: 0.385-0.637; Mean Absolute Error = 0.0819 ± 0.0088; Mean Squared Error = 0.0122 ± 0.0039). Tree-based models showed slightly lower performance (eg, Gradient Boosting R ² = 0.484 ± 0.113). Explainability analysis identified pain, mobility limitations, beta-blocker use, anxiety/depression symptoms, and difficulties with activities of daily living as the most influential variables.

Discussion: The findings highlight that deep learning models can capture complex relationships among multiple clinical and psychosocial variables. The combination of SHAP and LIME allows for clinically interpretable results, facilitating personalized decisions in chronic disease care. Furthermore, the accessibility of community pharmacies provides a practical setting for data collection and application of these predictive tools.

Conclusions: The study demonstrates the potential of machine learning to support personalized decision-making in the management of chronic diseases from accessible settings such as community pharmacies, identifying the most important factors affecting patients' quality of life.

Objective: To evaluate how electronic health records (EHR) optimization and data analytics supported a large rural health system action plan to mitigate the effects of a nation-wide blood culture bottle shortage.

Materials and methods: Following the announcement of a nationwide blood culture bottle shortage on July 10, 2024, we implemented EHR order modification, alternative alerts (LMAs) for clinical decision support (CDS), and developed a data analytics dashboard to track daily orders and inventory. We analyzed changes in daily blood culture specimen utilization before and after the EHR interventions using run charts. We assessed blood culture positivity and contamination rates, and provider interactions with LMAs as process measures.

Results: The EHR-based interventions led to a sustained 81% reduction in daily blood culture utilization during the shortage. Blood culture contamination rates remained consistent at 5.1% pre- and post-interventions, and positivity rates were stable (13.5% pre vs 12.8% post). Analysis of LMAs showed that 16% of blood culture orders were canceled after the alert, with only 3% reordered within one hour. The utilization and inventory monitoring reports became top 10% most-used within the health system, supporting operational decisions.

Discussion: Combining EHR optimization, CDS via LMAs, and data analytics effectively mitigated a critical resource shortage, demonstrated by a sustained 81% reduction in blood culture bottle utilization without compromising patient care. EHR order modifications contributed to the initial reduction; LMAs sustained the decrease and were well-received compared to traditional alerts. Rapid development of data analytics reports supported data-driven operational decisions.

Objectives: To explore how Spanish-speaking caregivers navigate translation barriers in patient portals and to assess their perspectives on improving language accessibility.

Materials and methods: This qualitative study was conducted at a pediatric academic health system. Semi-structured interviews were conducted with Spanish-speaking caregivers of children with chronic conditions, and inductive thematic analysis was used to generate themes.

Results: Twenty caregivers participated. Three key themes emerged: (1) Caregivers rely on online machine translation tools, which can be inaccurate and time-consuming; (2) Caregivers frequently depend on children and family members for translation, raising concerns about comprehension and appropriateness; (3) Caregivers expressed strong interest in timely and accurate translation features within patient portals to enhance accessibility.

Discussion: Spanish-speaking caregivers develop workarounds to access medical information, but these strategies pose risks to patient safety and exacerbate digital health inequities. While AI-powered machine translation offers a potential solution, concerns about accuracy, regulatory compliance, and equitable implementation must be addressed.

Conclusion: Spanish-speaking caregivers face significant challenges in accessing health information through patient portals. Health systems should prioritize integrated translation solutions, leveraging AI-driven tools while ensuring accuracy and equitable implementation to improve language accessibility.

Objective: This study aimed to evaluate the characteristics of drug allergy alerts (DAAs) and identify false-positive alerts by analyzing prior drug administration records.

Materials and methods: We retrospectively analyzed DAAs fired to providers with prescribing authority in 2023 at a large academic medical center in the Southeast to identify data elements that could help reduce clinically irrelevant alerts.

Results: Overall, 101 492 DAAs were triggered in 2023, with a 98.9% override rate. Alerts were fired for 9111 unique patients (an average of 11.1 alerts per patient). Only 9.7% DAAs had a definite match between the prescribed drug and documented allergen, with the remaining 90.3% fired for different drugs under the same class or allergen group. Overall, 70% DAAs were triggered for patients with prior administration of the drug triggering the alert, of these, 74% (52% of all DDAs) occurred in patients who had received the prescribed drug after initial allergy documentation, and 79% (56% of all DAAs) received the same drug again after 2023. Patients who had received prescribed drug previously, definite match were more likely to be overridden than no match (OR = 1.18, 95% CI: 1.03-1.33, P = .013) with a slightly higher override rate (98.9%% [n = 71 357] vs 98.8% [n = 30 135]).

Discussion: Most DAAs occurred in patients previously exposed to the alerting medication, often after allergy documentation, and over half of the patients continued to receive the same drug after 2023.

Conclusion: Future research should focus on examining strategies to incorporate tolerance assertions into DAA logic to reduce false-positives without compromising safety.

Objectives: To assess the performance of a reasoning large language model (LLM) in identifying medication errors in medical incident reports.

Materials and methods: OpenAI's O4-mini LLM was adapted using prompt engineering on 75 000 anonymized incident reports from the Västmanland region of Sweden (2019-2024). To guide the prompt design, we used a subset of 2434 reports, which were manually reclassified by pharmacists as medication-related or not. For validation, 200 reports (January 2024-March 2024) were independently classified by 2 pharmacists to establish a reference classification. Moreover, the LLM performed binary classification, with concordance rates measured against the expert consensus.

Results: The LLM achieved a concordance rate of 96.0% (192/200; 95% CI, 92.3-98.3) with expert classification. Eight cases (4.0%) showed disagreements, primarily due to linguistic ambiguity or context-dependent interpretation. Five cases involved pharmacists classifying reports as non-medication-related, while the LLM classified them as medication-related, with the reverse in 3 cases. Subcategorization accuracy was 76.5%.

Discussion: The LLM showed expert-level performance, outperforming existing automated methods. Thus, its integration into incident reporting systems might improve the efficiency, accuracy, and consistency of patient safety monitoring.

Conclusion: This validated AI-driven method can be integrated directly into clinical informatics workflows, enabling healthcare organizations to rapidly and consistently identify medication errors, ultimately enhancing patient safety outcomes.

Objective: Clinical studies struggle to locate the right patients, in part because many remain undiagnosed or lack relevant labels. We develop and prospectively test a physiology-based patient representation ("Bioprofile") to evaluate whether it can reduce the number of individuals who must be screened and enable more precise targeting of candidates for steatotic liver disease studies.

Materials and methods: We trained Bioprofile patient representations from routinely collected health data and fine-tuned against multiple endpoints. Bioprofiles were trained using 1 million subjects from the UK Biobank and other cohorts. The trained model was applied to 45 484 research subjects at Vanderbilt University Medical Center. Based on Bioprofile nominations, 31 subjects were recruited for a prospective validation study. The primary outcome measure was proton density fat fraction (PDFF), an imaging-based metric of steatosis severity.

Results: Bioprofile models achieved a Spearman coefficient of 0.65 against PDFF, outperforming existing foundation models (ρ = 0.361) and clinical risk scores (ρ = 0.522-0.542) in retrospective validation. Simulations found that Bioprofiles required half as many subjects needed to screen compared to existing methods depending on task. Bioprofiles were further validated against 5 global cohort studies. Bioprofile predictions aligned strongly with prospective study data (ρ = 0.740).

Discussion and conclusion: Artificial intelligence (AI)-based profiling of patient physiology can reveal individuals who have subclinical signatures of disease. If implemented widely, this approach can identify unknown human subjects, reduce screening failures, and improve trial quality. Bioprofiles may have additional utility in decision support applications in precision medicine and AI-augmented healthcare.

Objectives: To create a theoretical framework of mental health risk factors to inform the development of prediction models for young people's mental health problems.

Materials and methods: We created an initial prototype theoretical framework using a rapid literature search and stakeholder discussion. A snowball sampling approach identified experts for the Delphi study. Round 1 sought consensus on the overall approach, framework domains, and life course stages. Round 2 aimed to establish the points in the life course where exposure to specific risk factors would be most influential. Round 3 ranked risk factors within domains by their predictive importance for young people's mental health problems.

Results: The final framework reached consensus after 3 rounds and included 287 risk factors across 8 domains and 5 life course stages. Twenty-five experts completed round 3. Domains ranked as most important were "Social and Environmental" and "Psychological and Mental Health." Ranked lists of risk factors within domains and heat maps showing the salience of risk factors across life course stages were generated.

Discussion: The study integrated multidisciplinary expert perspectives and prioritized health equity throughout the framework's development. The ranked risk factor lists and life stage heat maps support the targeted inclusion of risk factors across developmental stages in prediction models.

Conclusion: This theoretical framework provides a roadmap of important risk factors for inclusion in early identification models to enhance the predictive accuracy of childhood mental health problems. It offers a useful theoretical reference point to support model building for those without domain expertise.

Objectives: Electronic health record (EHR) work may differently affect women and men physicians. Identifying gender discrepancies in EHR work across different specialties may inform strategies to reduce EHR burdens.

Materials and methods: We retrospectively evaluated EHR use by ambulatory physicians in 4 specialties (2 procedural [cardiology and gastroenterology] and 2 nonprocedural [internal medicine and rheumatology]) during 1 year at a large academic medical institution. Gender differences in EHR and clinical workload across specialties were evaluated by analysis of variance. Mixed-effects linear regression models analyzed gender differences in EHR workload controlling for specialty. Significant differences were additionally examined by stratifying procedural and nonprocedural specialties.

Results: Clinical and EHR workload varied across specialties (P <.05), though scheduled clinical workload did not differ by gender. Controlling for specialty, women physicians spent more time per appointment on In Basket messages (P =.001), sent more Secure Chat messages per appointment (P =.003), and spent more time in the EHR outside 7:00 AM-7:00 PM (P <.001) than men. Gender differences in messaging were concentrated among the procedural physicians. Women procedural physicians spent more time on In Basket messages (P <.001) and sent more Secure Chat messages (P =.007) than men, whereas these differences did not occur among nonprocedural physicians.

Discussion: Women physicians had greater EHR burdens despite similar scheduled clinical workloads as men. The greater messaging workload predominantly affected women procedural physicians.

Conclusion: Gender disparities in EHR burden in ambulatory specialties vary between procedural and nonprocedural fields. Future research is needed to mitigate gender inequity in EHR workloads.

Objectives: The primary objective of this research is to assess the content coverage of nursing data within a publicly available common data model (CDM), focusing on how nursing data, documented in flowsheets, are represented within the model.

Materials and methods: This mapping study was informed by previous evaluation studies and serves as a framework for evaluating information resources, including to guide development and implementation. The overall research process consists of 4 steps: (1) identify a CDM; (2) define evaluation criteria; (3) map nursing flowsheet data; and (4) apply evaluation criteria.

Results: Overall, 65.5% (n = 1170) of the flowsheet concepts were mapped to Systematized Nomenclature of Medicine-Clinical Terms (SNOMED CT) and Logical Observation Identifiers Names and Codes (LOINC) target codes and 56.0% (n = 1831) of the flowsheet values were mapped to SNOMED CT and LOINC target codes. The flowsheet concepts had a higher average mapping time per concept/reviewer (1.19 min) as compared to the average mapping time per value/reviewer (0.64 min).

Discussion: This mapping study demonstrated the progress and ongoing challenges of mapping nursing data to a national common data model. However, the ability to use nursing data at scale in a national CDM remains limited until more comprehensive mapping is completed.

Conclusion: This mapping study identifies a significant gap in integrating nursing data into a national common data model, highlighting an opportunity to enhance patient care through improved real-time insights and evidence-based nursing practices. Addressing this gap can help shape policies that prioritize the inclusion of nursing data. Additionally, aligning nursing data at scale can advance research, increase efficiency, and optimize nurse-sensitive patient outcomes.

[This corrects the article DOI: 10.1093/jamiaopen/ooz061.].