Citations

Mayo Clinic

Clear

12 citation(s) found for "Lee, Frank G".

AI-Based Screening Approach for Primary Aldosteronism Using Longitudinal Electronic Health Records
Choudhary, Ashok , Lee, Frank G

No container title available | October 8, 2026 | Volume: No volume available | Issue: No issue available | Pages: No page available

10.64898/2026.10.06.26364909 Get It

Abstract Context Current guidelines for primary aldosteronism (PA) recommend screening all patients with hypertension. Given low screening rates, an automated artificial intelligence (AI)-based risk-stratified approach may facilitate guideline implementation. Objective To develop an AI-based prescreening tool using electronic health record (EHR) data to identify at-risk patients. Design Model development retrospective cohort study. Ensemble architecture using eXtreme Gradient Boosting, random forest, and extremely randomized trees was trained on EHR data (1986-2025) and evaluated on a test set of patients with hypertension. Setting Single institution system spanning community clinics and tertiary/quaternary centers reported via Mayo Clinic Platform. Patients Adults with hypertension, International Classification of Disease (ICD) diagnosis of PA, or had plasma renin/aldosterone testing. Main Outcome Measures Predicting PA diagnosis (disease) or negative workup (control) using age, sex, ICD diagnoses, vitals, labs, and medications by area under receiver operating characteristic (AUROC), estimated calibration error (ECE), and specificity. Results Of 22,264 patients (1,833 disease and 20,431 controls), median time from initial encounter to screening was 5.0 years (IQR 0.5-12.6). The model demonstrated AUROC 0.709 (95% CI 0.688-0.730) with ECE 0.037. At a 0.5 cut-off threshold, the model had high specificity (71.6%) and low false positivity (28.4%). In a test cohort of 1,279,455 adults with hypertension, the model identified 5,295 (0.4%) for first-pass screening. Conclusions A high-specificity model can provide a pre-screening PA pathway using routine EHR data to identify and prioritize patients for PA screening. This AI-based approach can risk-stratify patients, guide sequencing of testing/referrals, and pragmatically facilitate adoption of current guidelines into practice.
Assessment of Fairness and Bias of an Image-based Surgical Site Infection Detection AI Model
Lee, Frank G. , Choudhary, Ashok , Li-Han, Leo Yan , Dale, Erica L., Bankes, Audrey C , Varghese, Chris , Anderson, Stephanie S., Habermann, Elizabeth B. , Etzioni, David A. , McLaughlin, Sarah A. , Salehinejad, Hojjat , Thiels, Cornelius A. , Varghese, Chris , Habermann, Elizabeth , Dale, Erica , Bankes, Audrey , Choudhary, Ashok , Lee, Frank G

No container title available | September 23, 2026 | Volume: No volume available | Issue: No issue available | Pages: No page available

10.64898/2026.09.22.26363636 Get It

ABSTRACT Background Evaluating fairness in artificial intelligence (AI) is essential before clinical implementation to prevent unintended harm across diverse patient populations. This study assesses performance differences of an image-based surgical site infection (SSI) detection model across age, gender, geography, race, and skin color. Methods Model development/validation study using retrospective, multicenter data across surgical subspecialties from adults who submitted an image of their incision via the electronic health record within 30 days of their procedure between 2019-2024. Area under receiver operating curve (AUROC) was compared by age (<55 vs. ≥55 years old), gender (male vs. female), geography (Rochester, Minnesota vs. Jacksonville, Florida vs. Scottsdale, Arizona vs. regional Minnesota/Wisconsin areas), race (White/Caucasian vs. non-White/Caucasian), and skin color (Fitpatrick I-III vs. IV-VI and Monk A-E vs. F-J). Results Training data contained 13,702 images from 4,274 patients between 2019-2022 (50.1% ≥55 years old; 63.2% female; 91.6% White/Caucasian; 58.5% Rochester, 17.3% Florida, 10.5% Arizona; 96.3% light skin tone), of which 1,348 images (9.8%) had an SSI. No differences in SSI detection existed for age, gender, race, or skin color (AUROC 0.83-0.87, p>0.09), however geographical differences were significant: lowest AUROC 0.83 (95% CI 0.77-0.88) vs. highest AUROC 0.87 (95% CI 0.84-0.89, p=0.01). Temporal validation using 4,242 images from 1,353 patients between 2022-2023 showed intact fairness across subgroups (ΔAUROC ≤0.06) except geography (ΔAUROC ≤0.16). Conclusions The model behaved fairly across age, gender, race, and skin color, however geographical differences persisted. These findings have practical implications underscoring the importance of evaluating bias in clinical AI before deployment.
NSQIP Outcomes for Open and Minimally Invasive Gastrectomy for Malignancy
Lee, Frank G. , Cross, Ashton E. , Habermann, Elizabeth B. , Thiels, Cornelius A. , Grotz, Travis E. , Hanson, Kristine , Cross, Ashton E , Habermann, Elizabeth , Thiels, Cornelius A , Grotz, Travis E , Hanson, Kristine T. , Lee, Frank G

Journal of Surgical Oncology | August 16, 2026 | Volume: No volume available | Issue: No issue available | Pages: No page available

10.1002/jso.70359 | PMID 42604579 Get It

ABSTRACT Background ACS‐NSQIP began capturing operative approach in 2022. With increased adoption of minimally invasive (MIS) techniques for malignancy, a contemporary review of benchmark outcomes for open, laparoscopic and robotic gastrectomy is warranted. Methods We performed a retrospective cohort study of adults in ACS‐NSQIP (2022–2024) undergoing partial or total gastrectomy (CPT 43631‐4 and 43620‐2) for cancer. Outcomes included operative time, length of stay (LOS), unplanned conversion‐to‐open, and 30‐day morbidity, mortality, and readmission. Multivariable logistic regression evaluated patient and operative factors associated with morbidity. Results Among 1,459 patients (median age 66, 60.4% male), 66.8% underwent partial gastrectomy and 33.2% total gastrectomy. Approaches were 61.2% open, 19.5% laparoscopic, and 19.3% robotic. Thirty‐day mortality (1.8%), major complications (18.2%), and readmissions (10.3%) did not differ by approach. Unplanned conversion‐to‐open occurred in 14.0% of MIS. Robotic had longer median operative time of +1.2 and +1.7 h than laparoscopic and open and median −1 and −2 days shorter LOS respectively ( p < 0.001). Hypoalbuminemia (OR 3.10), CKD ≥ 3 (OR 1.75), and anemia (OR 1.48) were independently associated with major complications ( p < 0.05). Conclusion Operative approach influences operative time and LOS but not short‐term morbidity and mortality. The newly available ACS‐NSQIP operative approach variable enables approach‐specific benchmarks for cancer‐related gastrectomy.
Computer-Vision Approach to Triaging Patient-Submitted Photos of Intestinal Ostomies
Varghese, Chris, Choudhary, Ashok, Larson, E W, Lee, Frank, Baier, Michele, Glabe, CG, Just, Jora N., Friedrich, Stacy, Merchea, Amit, Brady, Justin, Selig, Jody L., Thiels, Cornelius A., Salehinejad, Hojjat , Perry, William R.G., Varghese, Chris , Varghese, Chris , Baier, Michele C, Glabe, Cody L, Friedrich, Stacy A, Perry, William R , Lee, Frank, Glabe, Cody L., Merchea, Amit , Brady, Justin T. , Thiels, Cornelius A. , Perry, William

Diseases of the Colon & Rectum | July 17, 2026 | Volume: No volume available | Issue: No issue available | Pages: No page available

10.1097/dcr.0000000000004366 | PMID 42466540 Get It

BACKGROUND: Surgeons and ostomy nurses receive a high volume of stoma photos from patients. OBJECTIVE: This study aimed to develop and validate an automated and scalable artificial intelligence pipeline for identification and triage of patient-submitted intestinal ostomy photos. DESIGN: Retrospective derivation and validation study with stakeholder engagement to guide model design. Clinical teams reviewed photos in duplicate and categorized them into: (a) obstructed view, (b) healthy stoma, (c) suitable for conservative management, or (d) requiring in-person review. Pre-trained neural networks including MobileNetV4, ResNet50, ViT and CLIP-ViT were fine-tuned with 5-fold cross-validation. SETTINGS: Nine Mayo Clinic hospitals (2019–2022). PATIENTS: Adult patients (≥18 years) undergoing surgery who submitted images within 30 days after surgery. MAIN OUTCOME MEASURES: Model performance was evaluated with area under the receiver operator curve, precision, recall, and F1-scores. RESULTS: In this study, 538 photos of abdominal ostomies were sent in by 191 patients (median age 52 years, IQR 40–63; 60.4% female). Expert consensus triaged 236 (43.9%) photos as having an obstructed view, 30 (5.6%) as healthy, 177 (32.9%) for conservative management, and 95 (17.7%) for in-person review. All models achieved area under the receiver operator curve >0.97 for identifying stomas. The CLIP-ViT model performed best at triage (macro- area under the receiver operator curve 0.94 ± 0.06; F1 0.77 ± 0.10). End-to-end detection and triage using CLIP-ViT achieved area under the receiver operator curve of 0.99 ± 0.01, precision of 0.87 ± 0.08, recall of 0.93 ± 0.04, and F1 of 0.89 ± 0.07. Attention maps showed that models focused on stomas to determine classification. LIMITATIONS: Low sample size and lack of prospective validation. CONCLUSIONS: A vision language model pipeline accurately detected and triaged patient-submitted ostomy photos. Prospective evaluation is now needed to support integration into multidisciplinary digital workflows. See Video Abstract .
Cross-Domain Vision-Language Transfer Learning for Surgical Drain Detection in Real-World Postoperative Images
Choudhary, Ashok, Thiels, Cornelius A., Salehinejad, Hojjat , Lee, Frank G. , Larson, Ellen L. , Thiels, Cornelius A.

2026 IEEE 23rd International Symposium on Biomedical Imaging (ISBI) | May 20, 2026 | Volume: No volume available | Issue: No issue available | Pages: 1-4

10.1109/isbi61048.2026.11515732 Get It

Predictors and impact of multiple listing for liver transplantation in the United States across the acuity circles eras
Cywes, Claire , Bankes, Audrey C , Hamandi, Ali, Daniels, Robert J, Fipps, David C. , Lee, Frank G, Ellertson, Marka F, Limburg, Benjamin , Beaman, Lindsay R. , Diwan, Tayyab S, Ellertson, Marka , Hamandi, Ali M, Ellertson, Marka F. , Fipps, David C , Beaman, Lindsay , Bankes, Audrey , Lee, Frank G , Diwan, Tayyab S.

Liver Transplantation | April 28, 2026 | Volume: No volume available | Issue: No issue available | Pages: No page available

10.1097/lvt.0000000000000892 | PMID 42048126 Get It

Multi-listing (ML) increases the odds of transplantation compared with single-center listing. However, not all patients have equivalent access to ML. The recent allocation policy change prioritized distance from the donor hospital in an attempt to decrease geographic disparities. This study assessed access to liver transplant (LT) given ML practices before and after the acuity circles policy change. United Network for Organ Sharing data were queried for LTs between January 1, 2016, and November 1, 2024. Patients were grouped by listing at one versus multiple centers and by acuity circle era (ACE). Cohort characteristics and comparisons were described using chi-square and Wilcoxon rank tests. Multivariable logistic regressions and propensity score match (PSM) were performed to assess associations between ML and LT. The final cohort consisted of 105,030 patients; 4084 (3.9%) were multi-listed. Racial/ethnic minority groups, less educated, and publicly insured candidates were less often multi-listed. ML was positively associated with LT in all analyses (overall: OR=1.180, p <0.001; PSM: OR=1.164, p =0.001; pre-ACE: OR=1.127, p =0.015; post-ACE: OR=1.218, p <0.001). Education, race/ethnicity, and insurance type were also associated with LT across all analyses. ML patients were more often transplanted at their primary center with a significant increase in primary center LT post-ACE (53.4% vs. 63.4%, p <0.001). ML and LT had a positive correlation. There was a negative association between minority groups and ML as well as LT. Although acuity circles decreased geographic disparities, they did not change ML accessibility across candidates.
Can Large Language Models Be a Viable Tool for Consensus Working Groups? Experience of the Ventral Rectopexy Expert Consensus Group
Larson, Ellen L. , Vermunt, Jane , Thiels, Cornelius A., Christensen, P., Collinson, Rowan , Cornish, Julie , D’Hoore, André , Gurland, Brooke , Mellgren, Anders, Bordeianou, Liliana , Perry, William , Group, On behalf of the Ventral Rectopexy Consensus AI Group and the Ventral Rectopexy Working, Thiels, Cornelius A. , Christensen, Peter , Larson, Ellen L , Vermunt, Jane V , Thiels, Cornelius A , Christensen, Peter , Collinson, Rowan J , Cornish, Julie A , Gurland, Brooke H , Ratto, Carlo , Ris, Frederic , Bordeianou, Liliana G , Perry, William R , Lee, Frank G. , Ris, Frédéric , Stevenson, Andrew R. L.

Diseases of the Colon & Rectum | January 6, 2026 | Volume: 69 | Issue: 5 | Pages: 836-846

10.1097/dcr.0000000000004058 | PMID 41492895 Get It

BACKGROUND: The Ventral Rectopexy International Expert Panel recently published a consensus update on ventral rectopexy. The ability of large language models to synthesize the literature on ventral rectopexy without an explicit knowledge base was studied before the publication of the consensus. OBJECTIVE: To compare the responses and citations of different large language models on ventral rectopexy using the expert panel consensus as the reference standard. DESIGN: ChatGPT-4o, Gemini 1.5 Pro, and OpenEvidence were compared on content appropriateness (1 = inappropriate to 5 = appropriate), readability (Flesch reading ease), response length, citation fabrication, and citation quality per Oxford Levels of Evidence. The most content-appropriate chatbot response selected by the expert panel was de-identified and presented alongside the consensus text to 15 colorectal surgeons who attempted to identify the chatbot-generated text. SETTINGS: Questions were submitted from September 18 to 19, 2024. Analysis was performed before the online publication of the consensus on January 30, 2025. MAIN OUTCOME MEASURES: The main outcome measures included content appropriateness, fabricated citation rate, citation quality, and the accuracy of identifying human versus chatbot text by colorectal surgeons. RESULTS: OpenEvidence ranked highest for content appropriateness (mean 3.5/5), above Gemini (3.0/5) and ChatGPT (2.8/5; p < 0.001). ChatGPT was most verbose with the highest readability scores ( p = 0.021). ChatGPT fabricated 53% of citations; Gemini fabricated 12%; OpenEvidence fabricated 0% ( p < 0.001). All OpenEvidence citations were peer reviewed, with 40 of 117 (34%) citing level I to III studies compared with 46 of 94 (49%) of the references cited in the consensus ( p = 0.043). Chatbot-generated responses were identified with 28 of 51 (55%) accuracy. LIMITATIONS: Reproducibility may be affected due to the nature of large language models and the availability of the consensus after publication. CONCLUSIONS: OpenEvidence outperformed Gemini 1.5 Pro and ChatGPT-4o in content appropriateness and peer-reviewed citation quantity and quality. Chatbot-generated text was indistinguishable from expert-authored consensus to subject matter experts. Large language models, as an early-stage research tool, may be viable for future consensus working groups, provided transparent disclosure and rigorous oversight. See Video Abstract . ¿PUEDEN LOS MODELOS LINGÜÍSTICOS DE GRAN TAMAÑO SER UNA HERRAMIENTA VIABLE PARA LOS GRUPOS DE TRABAJO DE CONSENSO? LA EXPERIENCIA DEL GRUPO DE CONSENSO DE EXPERTOS EN RECTOPEXIA VENTRAL ANTECEDENTES: El Panel Internacional de Expertos en Rectopexia Ventral publicó recientemente una actualización consensuada sobre la rectopexia ventral. Antes de la publicación del consenso, se estudió la capacidad de los grandes modelos lingüísticos para sintetizar la bibliografía sobre la rectopexia ventral sin una base de conocimientos explícita. Objetivo: Comparar las respuestas y citas de diferentes grandes modelos lingüísticos sobre la rectopexia ventral utilizando el consenso del panel de expertos como estándar de referencia. DISEÑO: Se compararon ChatGPT-4o, Gemini 1.5 Pro y OpenEvidence en cuanto a la idoneidad del contenido (1 = inadecuado a 5 = adecuado), la legibilidad (índice de legibilidad de Flesch), la longitud de la respuesta, la fabricación de citas y la calidad de las citas según los niveles de evidencia de Oxford. La respuesta del chatbot más adecuada en cuanto al contenido seleccionada por el panel de expertos se anonimizó y se presentó junto con el texto del consenso a 15 cirujanos colorrectales que intentaron identificar el texto generado por el chatbot. ENTORNO: Las preguntas se enviaron los días 18 y 19 de septiembre de 2024. El análisis se realizó antes de la publicación del consenso en línea el 30 de enero de 2025. PRINCIPALES MEDIDAS DE RESULTADOS: Adecuación del contenido, tasa de citas inventadas, calidad de las citas, precisión en la identificación del texto humano frente al del chatbot por parte de los cirujanos colorrectales. RESULTADOS: OpenEvidence obtuvo la puntuación más alta en cuanto a la adecuación del contenido (media de 3,5/5), por encima de Gemini (3,0/5) y ChatGPT (2,8/5) (p<0,001). ChatGPT fue el más prolijo, con las puntuaciones más altas en legibilidad (p = 0,021). ChatGPT inventó el 53% de las citas; Gemini inventó el 12 %; OpenEvidence inventó el 0 % (p < 0,001). Todas las citas de OpenEvidence fueron revisadas por pares, con 40/117 (34 %) que citaban estudios de nivel I-III frente a 46/94 (49 %) de las referencias citadas en el consenso (p = 0,043). Las respuestas generadas por el chatbot se identificaron con una precisión del 28/51 (55 %). LIMITACIONES: Reproducibilidad debido a la naturaleza de los modelos de lenguaje grandes y a la disponibilidad del consenso tras la publicación. CONCLUSIONES: OpenEvidence superó a Gemini 1.5 Pro y ChatGPT-4o en cuanto a la idoneidad del contenido y la cantidad y calidad de las citas revisadas por pares. El texto generado por el chatbot era indistinguible del consenso redactado por expertos para los expertos en la materia. Los modelos de lenguaje grandes como herramienta de investigación en fase inicial pueden ser viables para futuros grupos de trabajo de consenso, siempre que se garantice una divulgación transparente y una supervisión rigurosa. ( AI-generated translation )
Dual-Space Augmentation and Sampling for Training Language Models on Limited and Imbalanced Surgical Reports
Choudhary, Ashok, Lee, Frank, Thiels, Cornelius, Salehinejad, Hojjat , Thiels, Cornelius A. , Lee, Frank G. , Choudhary, Ashok , Thiels, Cornelius A , Choudhary, Ashok , Lee, Frank G

Proceedings of the IEEE/ACM 12th International Conference on Big Data Computing, Applications and Technologies | December 24, 2025 | Volume: No volume available | Issue: No issue available | Pages: 1-6

10.1145/3773276.3774293 Get It

ViTCAI: A Vision-Language Model for Automated Triage and Captioning of Postoperative Incision Images Submitted by Remote Patients
Choudhary, Ashok , Lee, Frank G. , Muaddi, Hala , Thiels, Cornelius A. , Salehinejad, Hojjat , Thiels, Cornelius A , Choudhary, Ashok , Lee, Frank G

IEEE Journal of Biomedical and Health Informatics | December 8, 2025 | Volume: 29 | Issue: 12 | Pages: 8615-8622

10.1109/jbhi.2025.3624318 | PMID 41359691 Get It

Imaging-based Surgical Site Infection Detection Using Artificial Intelligence
Muaddi, Hala , Anderson, Stephanie , Etzioni, David A. , McLaughlin, Sarah A. , Kendrick, Michael L. , Salehinejad, Hojjat , Anderson, Stephanie S., Etzioni, David, Kendrick, Michael, Choudhary, Ashok, Lee, Frank, Habermann, Elizabeth B. , Thiels, Cornelius A.

Annals of Surgery | July 3, 2025 | Volume: 282 | Issue: 3 | Pages: 419-428

10.1097/sla.0000000000006826 | PMID 40607706 Get It

Objective: To develop an artificial intelligence–based pipeline to assess and triage patient-submitted postoperative wound images. Background: The rise of outpatient surgeries, remote monitoring, and patient-submitted wound images via online portals has contributed to a growing administrative workload on clinicians. Early identification of surgical site infection (SSI) is essential for reducing postoperative morbidity. Methods: Patients ≥18 years old who underwent surgery at 9 affiliated Mayo Clinic hospitals (2019-2022) and were captured by the National Surgical Quality Improvement Program (NSQIP) were included. Eligibility required submission of one image via the patient portal within 30 days after surgery. Images were independently screened in duplicate to determine the presence of an incision. SSI outcomes were obtained from NSQIP. The developed model consisted of 2 stages: incision detection and SSI detection in images with incisions. Four pretrained architectures were evaluated using 10-fold cross-validation, with upsampling and data augmentation to mitigate class imbalance. An end-to-end pipeline, image quality assessment and sensitivity analysis stratified by race were also performed. Results: Among 6060 patients, the median age was 54 years (interquartile range: 40–65), 61.4% (n=3805) were female, and 92.5% (n=5731) identified as White. SSIs were confirmed in 6.2% (n=386) images. Vision Transformer outperformed all others, achieving an incision detection accuracy of 0.94 (area under the curve=0.98) and an SSI detection accuracy of 0.73 (area under the curve=0.81). In addition, it demonstrated strong performance in assessing image quality. Sensitivity analysis revealed comparable performance across racial subgroups. Conclusion: This artificial intelligence pipeline demonstrates promising performance in automating wound image assessment and SSI detection, reducing clinical workload and improving postoperative care.