Citations

Importance Timely identification of aortic stenosis (AS) is essential for appropriate clinical management, yet screening remains limited by dependence on comprehensive echocardiography and trained imaging personnel. Objective To develop and validate a deep learning algorithm for detection of moderate or greater AS and prospectively evaluate its performance using artificial intelligence (AI)–guided focused cardiac ultrasound (FoCUS) acquired by novice operators. Design, Setting, and Participants This diagnostic study included retrospective algorithm development and validation and prospective evaluation of AI-guided FoCUS across Mayo Clinic sites in the Midwest, Arizona, and Florida. The model was developed using 6753 patients and evaluated in internal validation (n = 852), internal test (n = 844), and validation (n = 1912) cohorts. Performance was assessed on FoCUS acquired by experienced sonographers (n = 602) and prospectively by novice operators (n = 1302). The retrospective model development and validation cohorts comprised studies performed from January 2005 through September 2022. Prospective study was conducted in 2 enrollment periods from June to August 2024 and from June to September 2025. Participants from both periods were combined to comprise the final prospective cohort. Exposure AI-guided FoCUS acquisition and automated deep learning–based assessment for detection of moderate or greater AS. Main Outcomes and Measures Detection of moderate or greater AS. Performance was assessed using area under the receiver operating characteristic curve (AUROC), sensitivity, specificity, and positive predictive value. Results The model demonstrated excellent discrimination in the internal test cohort (AUROC, 0.99; 95% CI, 0.98-1.00) and geographically distinct validation cohorts in Arizona (AUROC, 0.99; 95% CI, 0.97-1.00) and Florida (AUROC, 0.99; 95% CI, 0.96-1.00). Among FoCUS examinations acquired by experienced sonographers, sensitivity was 95% (95% CI, 82-99) and specificity was 97% (95% CI, 95-98). In the prospective novice-operator cohort, 1258 of 1302 examinations (96.6%) were suitable for automated analysis. Sensitivity was 93% (95% CI, 82-99) and specificity was 96% (95% CI, 95-97). Expert review of AI-positive and uninterpretable examinations increased the positive predictive value from 49.4% to 91.1%, with sensitivity of 85.4%. Conclusions and Relevance A deep learning algorithm accurately detected moderate or greater AS across validation cohorts in this study. In prospective evaluation, novice operators were able to acquire AI-guided FoCUS examinations that enabled accurate detection of moderate or greater AS. These findings support a scalable strategy that may expand access to AS detection in settings with limited echocardiography resources.

Abstract Echocardiography is the most widely used cardiac imaging modality, yet artificial intelligence-enabled interpretation remains limited by the inability of existing models to integrate visual assessment, quantitative measurement, and clinical reasoning within a unified framework. Here we present EchoAtlas, the first autoregressive vision-language model developed for echocardiographic interpretation. Trained on over 12.9 million question-answer pairs derived from approximately 2 million echocardiogram videos, EchoAtlas achieves 0.966 accuracy on multiple-choice questions in our internal test set and establishes a new state-of-the-art on the public MIMIC-EchoQA benchmark (0.699 vs. 0.508 previously). EchoAtlas also provides accurate quantitative measurements, segment-level regional wall motion assessment, longitudinal comparison, and diagnostic reasoning across diverse question formats — capabilities not previously demonstrated in this domain. These results highlight the potential of autoregressive vision-language models as a foundation for interactive echocardiographic interpretation, representing an early step toward scalable, auditable artificial intelligence systems in cardiology practice.

Abstract Objective To investigate the ability of artificial intelligence‐enabled electrocardiogram (AI‐ECG) atrial fibrillation (AF) prediction model output and delta age, defined as AI‐estimated age minus chronological age, to predict adverse vascular outcomes in patients with migraine. Background Migraine, especially migraine with aura, is associated with an increased risk of cardiovascular events. Using normal sinus rhythm ECG, our institution has developed AI‐ECG models that predict the probability of AF and estimate patients' age, and from that, delta age (estimated age minus actual age) has been shown to reflect cardiovascular health and physiologic aging. Methods We conducted a retrospective cohort study that included adult patients with migraine who had at least one digital, standard 12‐lead ECG from 2000 to 2020 at any Mayo Clinic site. The first normal sinus rhythm ECG was defined as the index ECG, and those with adverse vascular events or AF prior to, or within 3 days after, index ECG were excluded. AI‐ECG AF prediction model output and delta age were obtained from the index ECG. The primary outcome was a composite endpoint of adverse vascular events that included acute myocardial infarction, acute ischemic stroke, venous thromboembolism, and death. New‐onset AF was also recorded as a secondary outcome. Results A total of 29,928 patients were included. The mean age was 44.3 ± 14.5 years, and the median follow‐up time was 54 months. During the study period, 4662 (15.6%) patients had adverse vascular events, and 1384 (4.6%) patients developed new‐onset AF. A higher AF prediction model output was associated with a higher risk of adverse vascular outcomes (adjusted hazard ratio [aHR]:1.15 [95% confidence interval: 1.12 to 1.18] for every 10% increase in prediction model output), as was a larger delta age (aHR:1.16 [1.12 to 1.21] for every 10‐year increase in delta age). Moreover, AI‐ECG AF prediction model outputs of ≥1% (aHR: 1.45 [1.36 to 1.54]) and delta age ≥−1 years (aHR: 1.09 [1.02 to 1.16]) were found to be the optimal cutoffs to dichotomize patients into high‐risk and low‐risk groups for adverse vascular events. Higher AI‐ECG AF prediction model output at baseline also predicted a higher likelihood of having new‐onset AF during the study period (aHR: 1.31 [1.26 to 1.37]), and the development of new‐onset AF was associated with a higher risk of adverse vascular events (aHR: 2.43 [2.17 to 2.73]). Conclusions For patients with migraine, a higher AI‐ECG AF prediction model output and delta age are predictive of adverse vascular outcomes. A higher AF prediction model output is also predictive of future AF development. Furthermore, we identified the cutoffs for both AI‐ECG models that can distinguish patients between high‐ and low‐risk groups. Our results showed that AI‐ECG models could inform clinical practice and identify patients with migraine at risk for future adverse vascular events.

Key Points Atrial fibrillation after kidney transplantation is common and is associated with poor prognosis. Artificial intelligence applied to 12-lead electrocardiograms pretransplant accurately predicts the risk of new-onset atrial fibrillation in our cohort. High-risk artificial intelligence electrocardiography score was independently associated with increased risk of mortality and allograft failure. Background Incident atrial fibrillation (AF) is common after kidney transplantation (KTx) and is associated with worse clinical outcomes. Artificial intelligence electrocardiography (AI-ECG) algorithms have demonstrated efficacy in predicting risk of new-onset AF in the general population; however, their prognostic value in KTx recipients is relatively unknown. Methods Retrospective analysis was conducted on KTx recipients without AF, with at least one pretransplant ECG between 2011 and 2021 across three tertiary centers in the United States (Mayo Clinic sites in MN, AZ, and FL). A previously validated AI-ECG algorithm estimated the probability of incident AF for each patient. Based on AI-ECG probabilities, patients were categorized into high-risk and low-risk groups, with the optimal AI-ECG score cutoff determined. The incidence of new-onset AF, allograft failure, and mortality was compared between groups. Results Overall, 6246 patients (age 53.5±13.8 years; 58.9% male) were included. Pretransplant AI-ECG probability of AF ≥5% was the optimal cutoff for high risk of incident AF (sensitivity 72%, specificity 62%). High-risk scores were associated with true new-onset AF at 30 days (adjusted hazard ratio [aHR], 2.89; 95% confidence intervals [CI], 2.05 to 4.09; P < 0.001), 3 years (aHR, 2.54; 95% CI, 1.99 to 3.26, P < 0.001), and 5 years post-transplant (aHR, 2.48; 95% CI, 1.99 to 3.09, P < 0.001). High-risk AI-ECG scores were also associated with increased mortality (aHR, 1.56; 95% CI, 1.30 to 1.88, P < 0.001) and overall allograft failure (aHR, 1.50; 95% CI, 1.30 to 1.75, P < 0.001) through the 5-year follow-up. Conclusions This pretransplant AI-ECG parameter identified patients at increased risk of new-onset AF post-KTx and provided prognostic utility. Overall, this easy to obtain tool allows for risk stratification of patients who may benefit from closer monitoring, targeted risk factor modification, and early intervention.

Abstract The advances in large language models (LLMs) have demonstrated sophisticated potential for automating complex tasks within the radiology workflow. From radiology report generation and report summarization to data collection for research trials, these models have proven to be powerful tools. However, optimal implementation of these models requires careful adaptation to the specialized medical domain. In addition, these models tend to generate information that is not truthful or factual, which can adversely affect patient care and clinical decisions. Strategies such as fine-tuning and prompt optimization have been shown to be impactful in eliminating these errors. Although these models undergo rapid updates and improvements, understanding the principles of prompt engineering and fine-tuning provides a foundation for evaluating and maintaining the performance of any LLM deployment. The current article aims to review the recent advancements in radiology using fine-tuning and prompt optimization to leverage LLMs’ capabilities. It delves into various techniques within each strategy, their advantages and limitations, and presents a framework to facilitate the practical integration of LLMs into radiology settings.

Abstract Background Automated data extraction from echocardiography reports could facilitate large-scale registry creation and clinical surveillance of valvular heart diseases (VHD). We evaluated the performance of open-source large language models (LLMs) guided by prompt instructions and chain of thought (CoT) for this task. Methods From consecutive transthoracic echocardiographies performed in our center, we utilized 200 random reports from 2019 for prompt optimization and 1000 from 2023 for evaluation. Five instruction-tuned LLMs (Qwen2.0-72B, Llama3.0-70B, Mixtral8-46.7B, Llama3.0-8B, and Phi3.0-3.8B) were guided by prompt instructions with and without CoT to classify prosthetic valve presence and VHD severity. Performance was evaluated using classification metrics against expert-labeled ground truth. Mean squared error (MSE) was also calculated for predicted severity’s deviation from actual severity. Results With CoT prompting, Llama3.0-70B and Qwen2.0 achieved the highest performance (accuracy: 99.1% and 98.9% for VHD severity; 100% and 99.9% for prosthetic valve; MSE: 0.02 and 0.05, respectively). Smaller models showed lower accuracy for VHD severity (54.1%-85.9%) but maintained high accuracy for prosthetic valve detection (&gt;96%). Chain of thought reasoning yielded higher accuracy for larger models while increasing processing time from 2-25 to 67-154 seconds per report. Based on CoT reasonings, the wrong predictions were mainly due to model outputs being influenced by irrelevant information in the text or failure to follow the prompt instructions. Conclusions Our study demonstrates the near-perfect performance of open-source LLMs for automated echocardiography report interpretation with the purpose of registry formation and disease surveillance. While larger models achieved exceptional accuracy through prompt optimization, practical implementation requires balancing performance with computational efficiency.