Citations

Abstract Background and Aims Artificial intelligence (AI) algorithms in 12-lead electrocardiogram (ECG) provides promising age prediction methods. This study investigated whether the discrepancy between ECG-derived AI-predicted age (AI-ECG age) and chronological age, termed electrocardiographic aging (ECG aging), is associated with atrial fibrillation (AF) risk. Methods An AI-ECG age prediction model was developed using a large-scale dataset (1 533 042 ECGs from 689 639 participants) and validated with six independent and multi-national datasets (737 133 ECGs from 330 794 participants). The AI-ECG age gap was calculated across two South Korean cohorts [mean (standard deviation) follow-up: 4.1 (4.3) years for 111 483 participants and 6.1 (3.8) years for 37 517 participants], one UK cohort [3.0 (1.6) years; 40 973 participants], and one US cohort [12.9 (8.6) years; 90 639 participants]. Participants were classified into two groups: normal group (age gap < 7 years) and ECG-aged group (age gap ≥ 7 years). The predictive capability of ECG aging for new- and early-onset AF risk was assessed. Results The mean AI-ECG ages were 51.9 (16.2), 47.4 (12.5), 68.4 (7.8), and 56.7 (14.6) years with age gaps of .0 (6.8), −.1 (6.0), 4.7 (8.7), and −1.4 (8.9) years in the two South Korean, UK, and US cohorts, respectively. In the ECG-aged group, increased risks of new-onset AF were observed with hazard ratios (95% confidence intervals) of 2.50 (2.24–2.78), 1.89 (1.46–2.43), 1.90 (1.55–2.33), and 1.76 (1.67–1.86) in the two South Korean, UK, and US cohorts, respectively. For early-onset AF, odds ratios were 2.89 (2.47–3.37), 1.94 (1.39–2.70), 1.58 (1.06–2.35), and 1.79 (1.62–1.97) in these cohorts compared with the normal group. Conclusions The AI-derived ECG aging was associated with the risk of new- and early-onset AF, suggesting its potential utility to identify individuals for AF prevention across diverse populations.

Abstract Aims The increasing need for diagnostic echocardiography tests presents challenges in preserving the quality and promptness of reports. While Large Language Models (LLMs) have proven effective in summarizing clinical texts, their application in echo remains underexplored. Methods and results Adult echocardiography studies, conducted at the Mayo Clinic from 1 January 2017 to 31 December 2017, were categorized into two groups: development (all Mayo locations except Arizona) and Arizona validation sets. We adapted open-source LLMs (Llama-2, MedAlpaca, Zephyr, and Flan-T5) using In-Context Learning and Quantized Low-Rank Adaptation fine-tuning (FT) for echo report summarization from ‘Findings’ to ‘Impressions.’ Against cardiologist-generated Impressions, the models’ performance was assessed both quantitatively with automatic metrics and qualitatively by cardiologists. The development dataset included 97 506 reports from 71 717 unique patients, predominantly male (55.4%), with an average age of 64.3 ± 15.8 years. EchoGPT, a fine-tuned Llama-2 model, outperformed other models with win rates ranging from 87% to 99% in various automatic metrics, and produced reports comparable to cardiologists in qualitative review (significantly preferred in conciseness (P < 0.001), with no significant preference in completeness, correctness, and clinical utility). Correlations between automatic and human metrics were fair to modest, with the best being RadGraph F1 scores vs. clinical utility (r = 0.42) and automatic metrics showed insensitivity (0–5% drop) to changes in measurement numbers. Conclusion EchoGPT can generate draft reports for human review and approval, helping to streamline the workflow. However, scalable evaluation approaches dedicated to echo reports remains necessary.

Background: We previously developed deep-learning algorithms that identify coronary artery disease (CAD) risk based on (i) coronary artery calcium (CAC), (ii) obstructive CAD by angiography, and (iii) left ventricular akinesis in ≥1 segment by echocardiogram, using a 12-lead electrocardiogram (CAD ECG-AI). We tested the hypothesis that those with increased probability of CAD by CAD ECG-AI algorithms will have an increased risk of atherosclerotic cardiovascular disease (ASCVD) events and would refine the AHA Predicting Risk of Cardiovascular Disease Events (PREVENT TM ) equations’ predictive capabilities. Methods: We assessed a group of consecutive patients who sought primary care in Olmsted County, MN, between 1997 and 2003. Passive follow-up was conducted using the Rochester Epidemiology Project's record linkage system. Patients included met the same criteria of the original PREVENT TM equations. The probability output of each CAD ECG-AI algorithm was used to predict ASCVD (fatal and non-fatal myocardial infarction and ischemic stroke) and ASCVD-Plus [further including percutaneous coronary intervention (PCI), coronary artery bypass grafting (CABG), and all-cause mortality]. Events were validated in duplicate. Cox proportional hazard models adjusted for age, sex, and risk factors while modeling each CAD ECG-AI algorithm as an individual and additive predictor (see footnote). PREVENT TM risk categories stratified analysis to evaluate the effect of the CAD ECG-AI models on predicted ASCVD risk. Results: We included 21,193 patients, with mean ± SD age 51.6 ± 12.0, 54% women and 95% white. After 14.2±12.4 years follow-up, 1,351 (6.4%) developed ASCVD, and 2,808 (13.2%) had ASCVD-Plus. The risk of ASCVD and ASCVD-Plus increased with positivity for the CAD ECG-AI algorithms after adjustments for age, sex, and PREVENT TM equations factors, all p for trend <0.001 (data not shown). In additive analyses, the risk of ASCVD and ASCVD-Plus increased with increased CAD ECG-AI factors, all p for trend <0.001 ( Fig. A-B ). Furthermore, the CAD ECG-AI algorithm enhanced the predictive capabilities of PREVENT TM equations across most risk subgroups ( Fig. C-E ). Conclusions: The deep-learning CAD ECG-AI algorithms displayed an independent and additive association with long-term ASCVD and ASCVD-Plus events in the community and improved the PREVENT TM predicted ASCVD risk. The CAD ECG-AI algorithms could help identify individuals at risk in primary prevention of cardiovascular events.

Background: Artificial intelligence (AI) can be used to predict a person’s age from the electrocardiogram (ECG) (AI-ECG age) and has been proposed as a measurement of biological age. The difference between AI-ECG age and chronological age (defined as delta-age) is an independent predictor of mortality in the general population. This study assessed the relationship between AI ECG estimated age and mortality among adults living with congenital heart disease (CHD). Methods: A previously validated neural network to estimate AI-ECG age was used to analyze standard digital 12-lead ECGs in a cohort of subjects aged >18 years seen in the adult CHD (ACHD) clinic between 1992 to 2023. A single ECG, collected during the year of the first visit to the ACHD clinic was analyzed to compute the delta-age.. A positive delta-age indicates an AI ECG estimated age higher than the subject’s chronological age. The relationship between the delta-age and mortality was evaluated using Cox proportional hazard models adjusting for influential clinical factors including chronologic age, sex, ACHD severity, NYHA class and history of arrhythmia. Results: Of 5,780 subjects tested (50% females), the mean chronological age was 39.1 ± 15.0 years. AI ECG estimated age was 52.3 ± 16.6 years. Complexity of CHD was classified as mild, moderate, and severe in 7.4%, 73.9% and 18.7% of patients respectively. Patients with severe CHD had the highest delta-age of 17.1 (±18.2) years followed by moderate 12.7 (±13.6) y and simple 7.7 (±11.5) y. Patients with single ventricle had the highest deviation in AI estimated age [median delta-age 21.0 (IQR 28.6) y] followed by those with a systemic right ventricle [median delta-age of 16.3 (IQR 29.3) y]. During a median follow-up of 6.4 y (25th – 75th centile 1.58 – 13.7 y), mortality occurred in 839 (14.5%) patients. After adjusting for chronological age, CHD severity, and other clinical variables, delta-age gap was associated with increased mortality risk (HR 1.06 (1.03 – 1.09) per 5-year increment in delta-age, p<0.005), Table 1. Conclusion: The difference between AI ECG and chronological age is an independent predictor of all-cause mortality in ACHD patients.

Background: Our team recently developed an Artificial Intelligence model that enables the identification of increased left ventricular filling pressures using a single-lead electrocardiogram (AI-ECG) with excellent performance (AUC 0.91). However, its ability to predict incident heart failure (HF) in the community is yet to be evaluated. Hypothesis: We tested the hypothesis that the output from the AI-ECG will predict the long-term incidence of heart failure (HF) and HF with preserved ejection fraction (HFpEF), and would refine the HF predictive capabilities of the AHA Predicting Risk of Cardiovascular Disease EVENTs (PREVENT TM ) Equations. Methods: Leveraging the resources of the Rochester Epidemiology Project (REP) we analyzed data from patients who sought primary care in Olmsted County, MN, between 1997-2003 with passive follow-up until April 2024. We included those with a digital ECG at baseline and excluded those who were part of the original algorithm development or history of heart failure. The diagnosis of heart failure was ascertained by using ICD codes (ICD9-428,ICD10-I.50) and a word search for the terms related to heart failure. HFpEF was ascertained using the Ejection Fraction (EF) data closest to the heart failure diagnosis date, using EF more than 40% as the criterion. We tested the association between AI-ECG predicted probabilities and new-onset HF and HFpEF using Cox proportional hazard models. These models were also adjusted for cardiovascular risk factors and were stratified to evaluate the effect of the AI-ECG output on PREVENT TM -predicted HF at 10 years. Results: We included 20,276 subjects, mean age 55±14.03, 52% women, 92% non-Hispanic white. After 18.2±8.5 years of follow-up, 4,008 (19.76%) were diagnosed with HF and 3,812 (18.8%) HFpEF. Risks of HF and HFpEF significantly increased in patients with increased left ventricular filling pressures by AI-ECG, independent of risk factors, all p for trend <0.001, see Fig A-B , respectively. AI-ECG enhanced the prediction capabilities of the PREVENT TM score across all risk groups, see Fig-C. Conclusions: AI-ECG-Derived increased left ventricular filling pressure is associated with new-onset HF and HFpEF in the community and was associated with incidental heart failure at any PREVENT TM risk category. Those results suggest a potential role of AI ECG in predicting heart failure.

Background: Many medications, electrolyte perturbations, and several diseases can prolong the QTc beyond its 99 th percentile value resulting in acquired QT prolongation (AQTP). In contrast, approximately 1 in 2,000 people have congenital long QT syndrome (LQTS) hallmarked by pathological QT prolongation. Recently, we developed an AI-ECG model to distinguish these two groups, however this model does not elucidate which ECG segments distinguish LQTS from AQTP. Methods: Using the cohort from our previous study, all patients with LQTS evaluated in the Mayo Clinic Genetic Heart Rhythm Clinic and controls were included. Next, every patient/control with ≥1 ECG above age- and sex-specific 99 th percentile for QTc was selected and matched at 1:100 ratio. Subsequently, 5 ML algorithms (logistic regression, Naïve Bayes, random forest, Gradient boosting and XGBoost) were trained using 52 features representing segments of the 12-lead ECG (e.g. wave peak times, amplitudes, areas under wave, duration; 52x12 features). Training and testing sets were spilt at 80:20 ratio and 5-fold cross validation was performed to prevent overfitting. Following analysis, the top 100 features were re-analyzed to refine results. Results: Among >1,600 patients with LQTS, 690 had ≥ 1 ECG with a QTc above the established threshold compared to 28,186 controls. Following age- and sex-matching and splitting, 22,410 (training), and 5,776 (testing) ECGs were used. Of the 5 ML algorithms, XGBoost demonstrated best performance (AUC 0.913, accuracy 0.836, sensitivity 0.857, specificity 0.835) to distinguish LQTS from AQTP. Following filtering for the top 100 features, performance of the algorithm remained high with AUC of 0.912 (accuracy 0.84, sensitivity 0.841, specificity 0.84). The features with the greatest impact on classification output included time from T wave onset to T wave peak (leads V1, V4, V4), QRS interval duration (leads V1, V3, V4, aVR), and amplitude at end of ST segment (aVF). Conclusions: Distinguishing monogenetically driven LQTS from multi-factorial AQTP is critical in clinical practice for proper management, mitigation of offending QT-prolonging factors, and treatment. Building on our previously developed AI-ECG-LQTS ‘mutation detector’, this model identifies some of the elements within the ‘black box’ that distinguish an ECG and QTc stemming from a patient with LQTS compared to a similar prolonged QTc value that is arising due to non-genetic, acquired QT-prolonging factors.

Background: Left ventricular diastolic dysfunction (LVDD) predicts mortality in cardiac intensive care unit (CICU) patients. A novel artificial intelligence enhanced electrocardiogram (AIECG) algorithm can predict LVDD and mortality in general populations but has not been examined in the cardiac intensive care unit (CICU). We aim to assess if LVDD by AI-ECG is associated with in-hospital and one-year mortality in CICU patients. Methods: In this retrospective cohort study, we included consecutive unique adults admitted to the Mayo Clinic CICU from 2007 to 2018 with an admission AIECG, which AI assigned LVDD grade (0 to 3). Medial mitral E/e’ ratio >15 on transthoracic echocardiogram (TTE) defined elevated filling pressures. AIECG and TTE assessment of LVDD were used to assign patients to four groups based on TTE as gold standard: true and false negative, and true and false positive. In-hospital mortality was evaluated using multivariable logistic regression. One-year survival was evaluated using Kaplan-Meier curves and multivariable Cox proportional-hazards. Results: We included 11,868 patients (median age 69.5 years, 37.7% females); 48% had heart failure and 44% had acute coronary syndromes. AIECG LVDD grades were: grade 0 (normal), 33%; grade 1, 7%; grade 2, 39%; grade 3, 21%. In-hospital (adjusted OR) and one-year (adjusted HR) mortality increased in each higher AIECG LVDD grade (Figure A/B), before and after adjustment including TTE measurements. Patients with grade 2-3 LVDD by AIECG and medial mitral E/e’ ratio >15 by TTE (true positive) had the highest in-hospital (adjusted OR 2.5 [1.7-3.9]) and one-year (adjusted HR 1.9 [1.5-2.5]) mortality (Figure C/D), and mortality was elevated similarly in patients with either grade 2-3 LVDD by AIECG (false positive) or medial mitral E/e’ ratio >15 by TTE (false negative). Conclusions: The AIECG LVDD grade was strongly associated with in-hospital and one-year mortality in CICU patients, even after adjusting for clinical variables and TTE measurements. Patients with concordant AIECG and TTE for elevated filling pressures (true positive) were at highest risk.

BACKGROUND: While clinical practice guidelines advocate for multidisciplinary heart team (MDHT) discussions in coronary revascularization, variability in implementation across health care settings remains a challenge. This variability could potentially be addressed by language learning models like ChatGPT, offering decision-making support in diverse health care environments. Our study aims to critically evaluate the concordance between recommendations made by MDHT and those generated by language learning models in coronary revascularization decision-making. METHODS: From March 2023 to July 2023, consecutive coronary angiography cases (n=86) that were referred for revascularization (either percutaneous or surgical) were analyzed using both ChatGPT-3.5 and ChatGPT-4. Case presentation formats included demographics, medical background, detailed description of angiographic findings, and SYNTAX score (Synergy Between Percutaneous Coronary Intervention With Taxus and Cardiac Surgery; I and II), which were presented in 3 different formats. The recommendations of the models were compared with those of an MDHT. RESULTS: ChatGPT-4 showed high concordance with decisions made by the MDHT (accuracy 0.82, sensitivity 0.8, specificity 0.83, and kappa 0.59), while ChatGPT-3.5 (0.67, 0.27, 0.84, and 0.12, respectively) showed lower concordance. Entropy and Fleiss kappa of ChatGPT-4 were 0.09 and 0.9, respectively, indicating high reliability and repeatability. The best correlation between ChatGPT-4 and MDHT was achieved when clinical cases were presented in a detailed context. Specific subgroups of patients yielded high accuracy (>0.9) of ChatGPT-4, including those with left main disease, 3 vessel disease, and diabetic patients. CONCLUSIONS: The present study demonstrates that advanced language learning models like ChatGPT-4 may be able to predict clinical recommendations for coronary artery disease revascularization with reasonable accuracy, especially in specific patient groups, underscoring their potential role as a supportive tool in clinical decision-making.

Background Early detection of left and right ventricular systolic dysfunction (LVSD and RVSD respectively) in children can lead to intervention to reduce morbidity and death. Existing artificial intelligence algorithms can identify LVSD and RVSD in adults using a 12‐lead ECG; however, its efficacy in children is uncertain. We aimed to develop novel artificial intelligence–enabled ECG algorithms for LVSD and RVSD detection in pediatric patients. Methods and Results We identified 10 142 unique pediatric patients (age≤18) with a 10‐second, 12‐lead surface ECG within 14 days of a transthoracic echocardiogram, performed between 2002 and 2022. LVSD was defined quantitatively by left ventricular ejection fraction (LVEF). RVSD was defined semiquantitatively. Novel pediatric models for LVEF ≤35% and LVEF <50% achieved excellent test areas under the curve of 0.93 (95% CI, 0.89–0.98) and 0.88 (95% CI, 0.83–0.94) respectively. The model to detect LVEF <50% had a sensitivity of 0.85, specificity of 0.80, positive predictive value of 0.095, and negative predictive value of 0.995. In comparison, the previously validated adult data‐derived model for LVEF <35% achieved an area under the curve of 0.87 (95% CI, 0.84–0.90) for LVEF ≤35% in children. A novel pediatric model for any RVSD detection reached a test area under the curve of 0.90 (0.87–0.94). Conclusions An artificial intelligence–enabled ECG demonstrates accurate detection of both LVSD and RVSD in pediatric patients. While adult‐trained models offer good performance, improvements are seen when training pediatric‐specific models.

Abstract Background Heart failure (HF) and atrial fibrillation (AF) are intertwined, sharing common risk factors and influencing each other in a reciprocal manner. While left ventricular diastolic dysfunction (LVDD) stands as a hallmark of HF's hemodynamic profile, its direct correlation with heightened AF risk remains uncertain. Our recent validation of an artificial intelligence-enabled electrocardiogram (AI-ECG) algorithm for LVDD detection has opened new avenues. Yet, whether AI-ECG-detected LVDD correlates with long-term AF risk remains unexplored. Methods We performed a retrospective study among all patients with a comprehensive LVDD assessment from the test population of the AI-ECG LVDD study between September 2001 and June 2022. Patients with prior documented AF according to health records or previous ECG were excluded. Patients were classified by AI-ECG as normal diastolic function (DF), grade 1 (G1), grade 2 (G2), or grade 3 (G3) LVDD. We assessed the risk of AF across AI-ECG LVDD grade, with normal DF as reference, with competing risk models. Results Of 97,023 patients free of AF at baseline, 63,533 (65.4%), 10,776 (11.1%), 18,940 (19.5%), and 3,774 (3.9%) were classified as normal DF, G1-LVDD, G2-LVDD, and G3-LVDD, respectively. Worse AI-ECG LVDD was associated with older age, lower left ventricular ejection fraction (LVEF), and higher rates of HF, diabetes, hypertension, chronic pulmonary disease, renal failure, and cardiovascular disease (p-of-trend&lt;0.001 for all). Over a median follow-up of 4.8 years (interquartile range 1.8-9.2 years), new-onset AF occurred in 6,795 (7.0%) patients. In multivariable survival analysis, adjusted to multiple risk factors including LVEF and clinical HF, G2 and G3 LVDD, representing patients with high filling pressures, were independently associated with a higher AF-risk [G2-LVDD: adjusted hazard ratio (aHR) 1.64 (95% confidence interval (95%CI) 1.52-1.73), G3-LVDD: aHR 2.24 (95%CI 2.02-2.48)]. Conclusion The AI-ECG LVDD algorithm independently predicts long-term AF risk, regardless of LVEF and clinical HF status, highlighting its potential for identifying high-risk populations and guiding proactive monitoring and management.