Faculty Bibliography

PURPOSE/OBJECTIVE:To develop Trainee Attributable & Automatable Care Evaluations in Real‑Time (TRACERs) for inpatient diabetes management, collect validity evidence for their use in formative assessment, and explore performance variation across residents and institutions. METHOD/METHODS:In 2023, a multi‑institutional team created two TRACERs based on type 2 diabetes guidelines-discourage bolus‑only insulin (TRACER #1) and encourage basal (± bolus) insulin (TRACER #2)-for internal medicine residents at three large residency programs. Residents were attributed to inpatient admissions based on placing the most medication orders in the first 12 hours. Structured queries extracted 35 discrete variables from the electronic health record (EHR). Two experts per institution reviewed random admissions (July-August 2022) to establish criterion validity. A retrospective cohort (July 2020-June 2023) added validity evidence. RESULTS:Automated extraction achieved ≥96% sensitivity and ≥95% specificity when compared to manual review. Among 615 residents attributed to 6,192 admissions of patients with type 2 diabetes at high risk for hyperglycemia, TRACER #1 occurred in 42.6% (1,689/3,965) of admissions at Program A, 28.9% (408/1,410) at Program B, and 26.7% (218/817) at Program C. TRACER #2 occurred in 44.9% (367/817) of Program C admissions versus 24.2% (959/3,965) at Program A and 28.2% (397/1,410) at Program B (all P < .001). Four resident-level insulin‑ordering profiles were identified-consistent basal-bolus insulin use (most guideline-concordant), basal-predominant, bolus-predominant, and bolus-only (most guideline-discordant)-with between‑resident variation exceeding between‑program differences. Longitudinally, cohort-level trends masked opposing individual trajectories-some residents improved with exposure while others worsened-and performance tertiles were distinguishable early in training. CONCLUSIONS:TRACERs revealed substantial institution‑ and resident‑level variation in insulin‑ordering practices, including guideline deviations, demonstrating potential for real‑time formative feedback. Multi‑institutional implementation highlighted scalability barriers, including EHR heterogeneity and workflow‑dependent attribution, underscoring the need for continued refinement and broader validation.

Integrated electronic health record databases provide an unprecedented opportunity to enhance knowledge of disease prediction, prevention, and management in real-world settings. The Precision Medicine (PMED) Registry is a cohort of approximately 2 million patients seen at NYU Langone Health inpatient and outpatient centers, capturing data generated during clinical care from January 1, 2010, to the present, with regular data updates. Data have been used for several research investigations, including international meta-analyses, validation of disease identification algorithms, local evaluation of risk tools, testing analytical pipelines for imaging data, and the investigation of novel correlates of established risk prediction models. Additionally, the assessment of local practice has provided insights into clinical practice patterns and aided quality improvement efforts to assess and promote the uptake of guideline-directed therapies at the system and provider level. This study illustrates how real-world integrated electronic health record data with multi-modal clinical information can be leveraged to support research in prediction, diagnosis, prevention, and treatment optimization across health systems.

BACKGROUND:Understanding the current risk of inpatient mortality following hip fracture in the United States is of significant value to patient families and the health system. Currently, the literature lacks a national representation of the inpatient mortality following hip fracture. PURPOSE/OBJECTIVE:The purpose of this study was to investigate the incidence of inpatient mortality following hip fracture using Epic Cosmos-an aggregated, de-identified, multi-institutional data that includes over 280 million patients in the United States. METHODS:A "Cosmos hip fracture cohort" that included all adults (18 years or older) who sustained a femoral neck, intertrochanteric, or subtrochanteric hip fracture (ICD 72.0, S72.1, S72.2) between January 2019 and December 2024 was created. The dataset was queried for demographic data including age, sex, geographic location, incidence of inpatient mortality, and bone health medication use at the time of admission. RESULTS:The Cosmos database included 284,455,033 patients, of which 1,232,250 hip fracture hospital admissions between January 2019 and December 2024 were identified. Of these patients, 47,773 (3.9%) expired during their hip fracture hospital admission. The most common age bracket was 85 years or older (39.8%), followed by 75-85 (30.0%), and 65-75 (17.8%). Most patients were white (91%) females (55.5%). Most inpatient mortalities occurred in the South (38.4%), followed by the Midwest (31.8%), followed by the Northeast (23.6%), and last by the West (6.2%). CONCLUSION/CONCLUSIONS:The current inpatient mortality following hip fracture is 3.9%. Most inpatient mortalities occurred in white females above the age of 85 in the South of the country. LEVEL OF EVIDENCE/METHODS:Level III.

Liver transplant (LT) recipients experience a wide range of comorbidities, leading to frequent healthcare encounters. Until now, national registries, which have limited exposures and outcomes, and laborious small cohort studies have been the main data sources for LT research. Cosmos database offers electronic health record (EHR)-based insights into LT recipients at the national level with granular data. We evaluated if Cosmos data is representative of the entire US LT recipient population. Using Cosmos (N=20,235) and the national Scientific Registry of Transplant Recipients (SRTR) (N=51,281), we identified adult, first-time LT recipients between 7/2016-12/2022. We compared demographics, clinical data, and mortality across datasets, calculating Kaplan-Meier survival estimates and multi-variable Cox regressions. Recipient characteristics were highly comparable (e.g., female: Cosmos=36.5% vs. SRTR=36.4%, Black: 6.8% vs. 7.2%; BMI: 28.5 kg/m2 [24.8-32.9] vs. 28.2 [24.6-32.4]). Lab values were similar across cohorts, including MELD (24 [17-30] vs. 23 [16-30]). Transplant indications, donor characteristics, and 5-year survival (Cosmos 83.1% [82.3-83.8) vs. SRTR 80.9% [80.4-81.3]) were similar. The associations of clinical factors with survival were similar across both groups. Cosmos database demonstrated acceptable generalizability to the general US LT recipient population, which may advance LT research through a better understanding about LT recipients' experiences and outcomes.

OBJECTIVE:Continuous glucose monitoring (CGM) is essential in diabetes care and research; however, extracting key data (e.g., time above, in, or below range) from CGM reports is manual, time-consuming, and inefficient. Natural language processing (NLP) can extract data from unstructured sources (e.g., images), but its application in CGM remains unexplored. We aimed to evaluate the accuracy of extracting CGM data using NLP. RESEARCH DESIGN AND METHODS/METHODS:We analyzed CGM reports stored as PDF files from the electronic health record at New York University Langone Health. The steps of our algorithm pipeline consist of 1) performing optical character recognition (OCR) to obtain glucose matrix data from CGM reports, 2) determining the type of CGM documents based on keywords in OCR results, 3) extracting variables of glucose based on CGM document type, and 4) storing the extracted glucose data in a structured database. Two experts with experience in CGM research and clinical practice conducted an independent manual review of 1% of the documents (n = 226). We calculated accuracy (correct extraction of CGM data) by comparing the algorithm's results with the manual review. RESULTS:Of the documents analyzed, 36.8% were Freestyle Libre and 63.2% were Dexcom. For information extraction, the agreement in evaluating Libre results between two experts was 99.93%. When comparing algorithm accuracy with manual review, the accuracy for Libre was 99.87% and, for Dexcom, 100.00%. CONCLUSIONS:Using an NLP approach to extract valuable glucose data from CGM PDF files is feasible and accurate, which can benefit clinical practice and diabetes research.

BACKGROUND:Abdominal computed tomography (CT) is commonly performed in adults. Abdominal aortic calcification (AAC) can be visualized and quantified using artificial intelligence (AI) on CTs performed for other clinical purposes (opportunistic CT). We sought to investigate the value of AI-enabled AAC quantification as a predictor of coronary artery disease and its association with cardiovascular events. METHODS:A fully automated AI algorithm to quantify AAC from the diaphragm to aortic bifurcation using the Agatston score was retrospectively applied to a cohort of patient that underwent both non-contrast abdominal CT for routine clinical care and cardiac CT for coronary artery calcification (CAC) assessment. Subjects were followed for a median of 36 months for major adverse cardiovascular events (MACE, composite of death, myocardial infarction [MI], ischemic stroke, coronary revascularization) and major coronary events (MCE, MI or coronary revascularization). RESULTS:Our cohort included 3599 patients (median age 60 years, 62% male, 74% white) with an evaluable abdominal and cardiac CT. There was a positive correlation between presence and severity of AAC and CAC (r=0.56, P<0.001). AAC showed excellent discriminatory power for detecting or ruling out any CAC (AUC for PREVENT risk score 0.701 [0.683 to 0.718]; AUC for PREVENT plus AAC 0.782 [0.767 to 0.797]; P<0.001). There were 324 MACE, of which 246 were MCE. Following adjustment for the 10-year cardiovascular disease PREVENT score, the presence of AAC was associated with a significant risk of MACE (adjHR 2.26, 95% CI 1.67-3.07, P<0.001) and MCE (adjHR 2.58, 95% CI 1.80-3.71, P<0.001). A doubling of the AAC score resulted in an 11% increase in the risk of MACE and a 13% increase in the risk of MCE. CONCLUSIONS:Using opportunistic abdominal CTs, assessment of AAC using a fully automated AI algorithm, predicted CAC and was independently associated with cardiovascular events. These data support the use of opportunistic imaging for cardiovascular risk assessment. Future studies should investigate whether opportunistic imaging can help guide appropriate cardiovascular prevention strategies.

BACKGROUND:Clinical use of continuous glucose monitoring (CGM) is increasing storage of CGM-related documents in electronic health records (EHR); however, the standardization of CGM storage is lacking. We aimed to evaluate the sensitivity and specificity of CGM Ambulatory Glucose Profile (AGP) classification criteria. METHODS:We randomly chose 2244 (18.1%) documents from NYU Langone Health. Our document classification algorithm: (1) separated multiple-page documents into a single-page image; (2) rotated all pages into an upright orientation; (3) determined types of devices using optical character recognition; and (4) tested for the presence of particular keywords in the text. Two experts in using CGM for research and clinical practice conducted an independent manual review of 62 (2.8%) reports. We calculated sensitivity (correct classification of CGM AGP report) and specificity (correct classification of non-CGM report) by comparing the classification algorithm against manual review. RESULTS:Among 2244 documents, 1040 (46.5%) were classified as CGM AGP reports (43.3% FreeStyle Libre and 56.7% Dexcom), 1170 (52.1%) non-CGM reports (eg, progress notes, CGM request forms, or physician letters), and 34 (1.5%) uncertain documents. The agreement for the evaluation of the documents between the two experts was 100% for sensitivity and 98.4% for specificity. When comparing the classification result between the algorithm and manual review, the sensitivity and specificity were 95.0% and 91.7%. CONCLUSION/CONCLUSIONS:Nearly half of CGM-related documents were AGP reports, which are useful for clinical practice and diabetes research; however, the remaining half are other clinical documents. Future work needs to standardize the storage of CGM-related documents in the EHR.

Developing real-world evidence from electronic health records (EHR) is vital to advance kidney transplantation (KT). We assessed the feasibility of studying KT using the Epic Cosmos aggregated EHR dataset, which includes 274 million unique individuals cared for in 238 U.S. health systems, by comparing it with the Scientific Registry of Transplant Recipients (SRTR). We identified 69,418 KT recipients transplanted between January 2014 and December 2022 in Cosmos (39.4% of all US KT transplants during this period). Demographics and clinical characteristics of recipients captured in Cosmos were consistent with the overall SRTR cohort. Survival estimates were generally comparable, although there were some differences in long-term survival. At 7 years post-transplant, patient survival was 80.4% in Cosmos and 77.8% in SRTR. Multivariable Cox regression showed consistent associations between clinical factors and mortality in both cohorts, with minor discrepancies in the associations between death and both age and race. In summary, Cosmos provides a reliable platform for KT research, allowing EHR-level clinical granularity not available with either the transplant registry or healthcare claims. Consequently, Cosmos will enable novel analyses to improve our understanding of KT management on a national scale.