Citations

ImportanceSpecialized ophthalmology terminology limits comprehension for nonophthalmology clinicians and professionals, hindering interdisciplinary communication and patient care. The clinical implementation of large language models (LLMs) into practice has to date been relatively unexplored.ObjectiveTo evaluate LLM–generated plain language summaries (PLSs) integrated into standard ophthalmology notes (SONs) in improving diagnostic understanding, satisfaction, and clarity.Design, Setting, and ParticipantsRandomized quality improvement study conducted from February 1, 2024, to May 31, 2024, including data from inpatient and outpatient encounters in a single tertiary academic center. Participants were nonophthalmology clinicians and professionals and ophthalmologists. The single inclusion criterion was any encounter note generated by an ophthalmologist during the study dates. Exclusion criteria were (1) lack of established nonophthalmology clinicians and professionals for outpatient encounters and (2) procedure-only patient encounters.InterventionAddition of LLM-generated plain language summaries to ophthalmology notes.Main Outcomes and MeasuresThe primary outcome was survey responses from nonophthalmology clinicians and professionals assessing understanding, satisfaction, and clarity of ophthalmology notes. Secondary outcomes were survey responses from ophthalmologists evaluating PLS in terms of clinical workflow and accuracy, objective measures of semantic quality, and safety analysis.ResultsA total of 362 (85%) nonophthalmology clinicians and professionals (33.0% response rate) preferred the PLS to SON. Demographic data on age, race and ethnicity, and sex were not collected. Nonophthalmology clinicians and professionals reported enhanced diagnostic understanding (percentage point increase, 9.0; 95% CI, 0.3-18.2; P = .01), increased note detail satisfaction (percentage point increase, 21.5; 95% CI, 11.4-31.5; P < .001), and improved explanation clarity (percentage point increase, 23.0; 95% CI, 12.0-33.1; P < .001) for notes containing a PLS. The addition of a PLS was associated with reduced comprehension gaps between clinicians who were comfortable and uncomfortable with ophthalmology terminology (from 26.1% [95% CI, 13.7%-38.6%; P < .001] to 14.4% [95% CI, 4.3%-24.6%; P > .06]). PLS semantic analysis found high meaning preservation (bidirectional encoder representations from transformers score mean F1 score: 0.85) with greater readability than SONs (Flesch Reading Ease: 51.8 vs 43.6; Flesch-Kincaid Grade Level: 10.7 vs 11.9). Ophthalmologists (n = 489; 84% response rate) reported high PLS accuracy (90% [320 of 355] a great deal) with minimal review time burden (94.9% [464 of 489] ≤1 minute). PLS error rate on ophthalmologist review was 26% (126 of 489). A total of 83.9% (104 of 126) of errors were deemed low risk for harm and none had a risk of severe harm or death.Conclusions and RelevanceIn this study, use of LLM-generated PLSs was associated with enhanced comprehension and satisfaction among nonophthalmology clinicians and professionals, which might aid interdisciplinary communication. Careful implementation and safety monitoring are recommended for clinical integration given the persistence of errors despite physician review.

Background: Cognitive change is an important factor in understanding dementia. Estimating effects of exposures on cognitive change requires choosing an analytical timescale, typically time on study or current age. There is limited consensus regarding timescale choice in epidemiologic cognitive aging research. Methods: Using a coordinated analytic approach in 10 cohorts of older adults, we evaluated whether estimated effects of two exposures on memory change differed depending on timescale (time on study or current age). We modeled effects of apolipoprotein-E (APOE) ε4 genotype (a time-invariant exposure) and diabetes (a potentially time-varying/acquired exposure evaluated at baseline) using mixed-effects models with linear and nonlinear time specifications for both timescales. Results: Among APOE ε4 carriers, model-estimated memory scores at baseline (time on study timescale) or at each cohort’s median baseline age (current age timescale) were lower, and the average rate of decline was faster than among APOE ε4 noncarriers. Model-estimated memory scores at baseline or at median baseline age were generally similar across baseline diabetes status, with variability across cohorts in the diabetes-memory change association. In some cohorts, trends in diabetes-memory change associations differed in direction across timescales. Conclusions: Timescale was largely inconsequential for estimated effects of APOE genotype, but yielded differences in estimated diabetes effects, raising questions about the appropriate scale. The current age scale may be problematic because diabetes was measured heterogeneously in age across individuals, a common issue in aging cohorts. Our work demonstrates approaches to evaluate alternative timescales, including in multicohort analyses, and highlights potential implications for estimated exposure effects on cognitive change.

Abstract Studies have identified genetic and epidemiologic factors associated with mammographic density (MD) phenotypes. However, MD-associated genetic variants only account for a small proportion of the total estimated heritability. Interrogating interactions between genetic and epidemiologic factors could potentially identify additional MD-associated loci, expand our understanding of the genetic basis of MD phenotypes, and clarify how epidemiologic factors modulate relationships between genetic variants and MD. We conducted six separate genome-wide, gene–environment (GxE) interaction analyses, applying 2 degrees of freedom (df) and 1df interaction tests, for each of three MD phenotypes (percent density, dense area (DA), and nondense area (NDA)). The six epidemiologic factors considered were height, ever parous, parity, ever menopausal hormone therapy, ever breastfeeding, and months of breastfeeding. We included European ancestry participants from multiple studies within the Markers of Density consortium and the Breast Cancer Association Consortium (n = 4895-16 218 depending on specific analyses). We identified 11 loci with genome-wide significant (P < 5 × 10−8) interaction tests including two novel common genetic signals interacting with parity (8p21.2) and ever breastfeeding (19p13.2) for NDA. Our results suggest that epidemiologic risk factors might influence relationships between common genetic variants and MD phenotypes at particular genomic loci.

Key Points Pooled discrimination metrics were acceptable (area under the receiver operating characteristic curve >0.70) for all AKI risk classification categories in both internal and external validation. Better performance was observed in most recently published studies and those with a low or unclear risk of bias. Significant heterogeneity in patient populations, definitions, clinical predictors, and methods limit implementation in real-world clinical scenarios. Background Artificial intelligence through machine learning models seems to provide accurate and precise AKI risk classification in some clinical settings, but their performance and implementation in real-world settings has not been established. Methods PubMed, Excerpta Medica (EMBASE) database, Web of Science, and Scopus were searched until August 2023. Articles reporting on externally validated models for prediction of AKI onset, AKI severity, and post-AKI complications in hospitalized adult and pediatric patients were searched using text words related to AKI, artificial intelligence, and machine learning. Two independent reviewers screened article titles, abstracts, and full texts. Areas under the receiver operating characteristic curves (AUCs) were used to compare model discrimination and pooled using a random-effects model. Results Of the 4816 articles initially identified and screened, 95 were included, representing 3.8 million admissions. The Kidney Disease Improving Global Outcomes (KDIGO)-AKI criteria were most frequently used to define AKI (72%). We identified 302 models, with the most common being logistic regression (37%), neural networks (10%), random forest (9%), and eXtreme gradient boosting (9%). The most frequently reported predictors of hospitalized incident AKI were age, sex, diabetes, serum creatinine, and hemoglobin. The pooled AUCs for AKI onset were 0.82 (95% confidence interval, 0.80 to 0.84) and 0.78 (95% confidence interval, 0.76 to 0.80) for internal and external validation, respectively. Pooled AUCs across multiple clinical settings, AKI severities, and post-AKI complications ranged from 0.78 to 0.87 for internal validation and 0.73 to 0.84 for external validation. Although data were limited, results in the pediatric population aligned with those observed in adults. Between-study heterogeneity was high for all outcomes (I 2 >90%), and most studies presented high risk of bias (86%) according to the Prediction Model Risk of Bias Assessment Tool. Conclusions Most externally validated models performed well in predicting AKI onset, AKI severity, and post-AKI complications in hospitalized adult and pediatric populations. However, heterogeneity in clinical settings, study populations, and predictors limits their generalizability and implementation at the bedside.

INTRODUCTION: The recent US Food and Drug Administration approval of resmetirom for treating metabolic dysfunction-associated steatohepatitis in patients necessitates patient selection for significant fibrosis or higher (≥F2). No existing vibration-controlled transient elastography (VCTE) algorithm targets ≥F2. METHODS: The mAchine Learning ADvanceD fibrosis and rIsk metabolic dysfunction-associated steatohepatitis Novel predictor (ALADDIN) study addressed this gap by introducing a machine-learning-based web calculator that estimates the likelihood of significant fibrosis using routine laboratory parameters with and without VCTE. Our study included a training set of 827 patients, a testing set of 504 patients with biopsy-confirmed metabolic dysfunction-associated steatotic liver disease from 6 centers, and an external validation set of 1,299 patients from 9 centers. Five algorithms were compared using area under the curve (AUC) in the test set: ElasticNet, random forest, gradient boosting machines, XGBoost, and neural networks. The top 3 (random forest, gradient boosting machines, and XGBoost) formed an ensemble model. RESULTS: In the external validation set, the ALADDIN-F2-VCTE model, using routine laboratory parameters with VCTE (AUC 0.791, 95% confidence interval [CI]: 0.764–0.819), outperformed VCTE alone (0.745, 95% CI 0.717–0.772, P < 0.0001), FibroScan-aspartate aminotransferase (0.710, 0.679–0.748, P < 0.0001), and Agile-3 model (0.740, 0.710–0.770, P < 0.0001) regarding the AUC, decision curve analysis, and calibration. The ALADDIN-F2-Lab model, using routine laboratory parameters without VCTE, achieved an AUC of 0.706 (95% CI: 0.668–0.749) and outperformed Fibrosis-4, steatosis-associated fibrosis estimator, and LiverRisk scores. DISCUSSION: Along with the steatosis-associated fibrosis estimator model developed to target significant fibrosis or higher, ALADDIN-F2-VCTE (https://aihepatology.shinyapps.io/ALADDIN1) uniquely supports a refined noninvasive approach to patient selection for resmetirom without the need for liver biopsy. In addition, ALADDIN-F2-Lab (https://aihepatology.shinyapps.io/ALADDIN2) offers an effective alternative when VCTE is unavailable.