Citations

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.

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.

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 .

Graph neural networks (GNNs) are increasingly used for medical imaging analysis and biological data modeling, where the integration of radiomics, topology, geometry, and generative artificial intelligence (AI) may improve representation learning from medical images and related biomedical data. Across the reviewed literature, GNNs show particular value for modeling spatial relationships, multimodal interactions, graph-structured biological networks, and non-Euclidean imaging features that are difficult to capture using conventional convolutional architectures alone. Topology- and geometry-aware approaches further expand this capability by encoding multi-scale structure, higher-order relationships, curvature, geodesic organization, and equivariant spatial priors. Hybrid graph–transformer models and generative graph methods represent emerging directions for modeling long-range dependencies, augmenting scarce datasets, supporting synthetic pretraining, and improving representation learning in low-label or heterogeneous biomedical settings. However, clinical translation remains limited by variability in graph construction, limited external validation, computational cost, scalability constraints, interpretability challenges, and uncertainty regarding the biological realism of synthetic data. Overall, this review highlights that GNN-based medical imaging analysis is most likely to advance when graph construction is biologically justified, model performance is evaluated across diverse clinical cohorts, and technical gains are paired with transparent validation, interpretability, and implementation strategies.

The integration of artificial intelligence (AI) in medical imaging has revolutionized diagnostic capabilities, yet the black-box nature of deep learning models poses significant challenges for clinical adoption. Current explainable AI (XAI) approaches, including SHAP, LIME, and Grad-CAM, predominantly focus on post hoc explanations that may inadvertently undermine clinical decision-making by providing misleading confidence in AI outputs. This paper presents a systematic review and meta-analysis of 67 studies (covering 23 radiology, 19 pathology, and 25 ophthalmology applications) evaluating XAI fidelity, stability, and performance trade-offs across medical imaging modalities. Our meta-analysis of 847 initially identified studies reveals that LIME achieves superior fidelity (0.81, 95% CI: 0.78–0.84) compared to SHAP (0.38, 95% CI: 0.35–0.41) and Grad-CAM (0.54, 95% CI: 0.51–0.57) across all modalities. Post hoc explanations demonstrated poor stability under noise perturbation, with SHAP showing 53% degradation in ophthalmology applications (ρ = 0.42 at 10% noise) compared to 11% in radiology (ρ = 0.89). We demonstrate a consistent 5–7% AUC performance penalty for interpretable models but identify modality-specific stability patterns suggesting that tailored XAI approaches are necessary. Based on these empirical findings, we propose a comprehensive three-pillar accountability framework that prioritizes transparency in model development, interpretability in architecture design, and a cautious deployment of post hoc explanations with explicit uncertainty quantification. This approach offers a pathway toward genuinely accountable AI systems that enhance rather than compromise clinical decision-making quality and patient safety.

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.