Attia, Zachi I.

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.

Abstract Aims Artificial intelligence electrocardiography (AI-ECG) algorithms are emerging tools for identifying individuals at risk of atrial fibrillation (AF). We evaluated the predictive performance of a validated AI-ECG algorithm for incident AF in UK Biobank participants with AF risk factors, irrespective of prevalent cardiovascular disease, and its incremental value when added to clinical predictors. Methods and results The AI-ECG tool was applied to sinus rhythm ECGs from UK-Biobank participants with risk factors for AF but no AF. Model performance was evaluated using time-dependent ROC-AUC and Harrell’s C-index. Multivariable Cox regression was used to identify clinical risk factors associated with incident AF and to quantify the contribution of AI-ECG. A total of 21 842 participants (56% male) were included. The median follow-up time was 3.7 years (IQR 0.5–5.4) The ECG-AI tool achieved a ROC-AUC of 0.73 (95% CI 0.68–0.78) at 1 and 0.69 (95% CI 0.66–0.72) at 3 years. A multivariable Cox regression model using clinical parameters achieved a ROC-AUC of 0.71 (95% CI 0.66–0.75) at 1 and 0.71 (95% CI 0.68–0.74) at 3 years. By adding ECG-AI to the clinical Cox regression model, the ROC-AUC increased to 0.75 (95%CI 0.71–0.80) at 1 and 0.74 (95% CI 0.71–0.77) at 3 years. AI-ECG showed a hazard ratio of 1.23 per decile increase (95% CI 1.18–1.27). Conclusion An AI-ECG algorithm improved the prediction of incident AF when added to a clinical parameter-based model over a median follow-up time of 3.7 years among individuals with comorbidities predisposing to AF who may benefit from targeted screening and preventive strategies.

Abstract Introduction Atrial fibrillation (AF) is a major contributor to stroke risk, yet many cases remain undiagnosed due to its silent nature. Deep learning-based electrocardiogram (DL-ECG) analysis can streamline AF risk stratification, providing a more reliable and accessible method to stratify risk using only a standard resting ECG. However, this tool has been validated only on AF detected clinically or with short ECG event monitors. Purpose This study aimed to validate the predictive accuracy of a DL-ECG model for identifying AF in normal sinus rhythm (NSR) ECGs among patients undergoing continuous cardiac monitoring with implantable loop recorders (ILRs). Methods We conducted a historical cohort study across three major U.S. academic medical centers, enrolling patients with an ILR and no prior diagnosis or documented ECG evidence of AF. Pre-implantation 10-second, 12-lead ECGs in NSR were analyzed using the DL-ECG model. Performance was evaluated using the area under the receiver operating characteristic curve (AUC) and compared with the CHARGE-AF score. Kaplan-Meier survival analysis and Cox proportional hazards models assessed the association between DL-ECG risk categories and incident AF, defined as any episode lasting ≥6 minutes. Results We included 2,523 patients (mean age 66 ± 17 years, 52% male). AF was detected in 658 (26%) over a median follow-up of 19 months (IQR 8-37), Table 1. The DL-ECG model classified 1,634 patients (65%) as at low-risk and 889 (35%) as at high-risk. AF developed in 242 (14.8%) of the low-risk group and 416 (46.8%) of the high-risk group, yielding an AUC of 0.73 (95% CI 0.70-0.75), surpassing the CHARGE-AF AUC of 0.69 (95% CI 0.66-0.71). A combined DL-ECG and CHARGE-AF model yielded an AUC of 0.74 (95% CI 0.72-0.76), Figure 1 presents the AUC, Net Reclassification Index, and DeLong test comparisons among the three models. High-risk patients had a significantly higher incidence of AF, which remained significant after adjusting for CHARGE-AF variables, Figure 2. AF incidence increased as DL-ECG quartiles increased, with 61% of patients at the top quartile developing AF during follow-up, Figure 3. False negatives (patients initially classified as low-risk but in whom later AF was detected) had a longer median time to AF detection (127 days [IQR 28-425] vs. 101 days [IQR 20-318] for true positives, p = 0.01), and where found to have at least one positive ECG during follow-up in 36%, compared to 32% of true negatives, indicating eventual risk reclassification. Conclusions Among patients undergoing ILR monitoring, the DL-ECG model demonstrated superior predictive performance compared to CHARGE-AF, offering a promising and accessible alternative for early detection of silent AF. This approach could aid early AF detection, particularly in resource-limited settings where access to specialized rhythm monitoring services is limited.Table 1 and Figure 1 Figure 2 and 3

Abstract Background Ventricular systolic dysfunction (LVSD) is the main predictor of mortality in Chagas disease (ChD). Although LVSD can be treated with affordable medications that enhance both symptoms and survival, its diagnosis necessitates advanced imaging tests, which are often unavailable in resource-limited settings. Recently, artificial intelligence (AI) models applied to an electrocardiogram (ECG) have shown promise in detecting LVSD and assessing disease progression, acting as digital biomarkers and offering the potential for screening tools. Purpose This study aims to evaluate the performance of an AI-ECG model in detecting LVSD and predicting mortality in patients with ChD. Specifically, we will fine-tune the AI-ECG model to improve LVSD identification and update the previously published mortality score (JAHA) by replacing NT-proBNP with AI-ECG to assess mortality risk. Methods This study was conducted with patients from the SaMi-Trop project, a prospective cohort of individuals seropositive for ChD from 21 municipalities in Minas Gerais, an endemic region of Brazil. The baseline assessment occurred in 2013–2014, with follow-ups in 2015–2016 (FU1) and 2022–2023 (FU2). The AI-ECG model was fine-tuned using a dataset with an 8% low ejection fraction prevalence. The model was trained with a binary cross-entropy loss function, a batch size of 128, and an Adam optimizer with a learning rate of 0.001. The area under the curve (AUC) was calculated for validation. The fine-tuned model, parsed into 2-second ECG windows, was applied to two waves of the SaMi-Trop cohort: Baseline (4,701 ECGs from 1,906 patients) and FU1 (2,324 ECGs from 1,059 patients). The association between AI-detected LVSD (AI-LVSD) and mortality was assessed using Cox proportional hazard models. Results The fine-tuned AI-LVSD model achieved an AUC of 0.88 (95% CI: 0.78–0.97) in detecting LVSD. When updating the JAHA mortality score, AI-LVSD replaced NT-proBNP with a stronger association with mortality risk, yielding hazard ratios (HR) of 4.5 (95% CI: 3,0–6.8) for two-year mortality and 4.3 (95% CI: 3.4–5.5) for overall mortality. Additionally, when incorporated into a simplified hazard model alongside age and NYHA class, AI-LVSD emerged as a strong biomarker for mortality risk, with hazard ratios of 5.4 (95% CI: 3.7–7.8) for two-year mortality and 4.9 (95% CI: 4.0–6.1) for overall mortality (Figure 1). Conclusion The AI-ECG model accurately identified LVSD for screening purposes and effectively replaced NT-proBNP in a simplified mortality risk model for Chagas disease. This tool can potentially enhance early detection and risk stratification, particularly in remote areas with limited access to advanced diagnostic resources.Figure 01.Cox proportional hazard model

Abstract Background Left ventricular diastolic function (LVDF) and left bundle branch block (LBBB) are both linked with increased risk of LV systolic dysfunction, heart failure, and mortality. Assessing LVDF, which is typically done through echocardiography, becomes challenging in the presence of LBBB due to the intraventricular conduction delay that disrupts the reliability of tissue Doppler imaging. Recently, we validated an artificial intelligence-enabled electrocardiogram (AI-ECG) model to assess LVDF and filling pressures. Methods We performed a retrospective study among patients with ECG-confirmed LBBB who also underwent comprehensive echocardiography within 14 days of the index ECG between September 2001 and June 2022. Using multivariable survival models, we assessed the risk of all-cause mortality across AI-ECG-derived LVDF grades. Results Of 2,554 patients with LBBB (mean age 72.0+12.6, 54.6% women), 252 (9.8%), 656 (25.7%), and 1,436 (56.2%), and 210 (8.2%) had normal LVDF, G1 LVDF, G2 LVDF, and G3 LVDF, respectively. The median LV ejection fraction (LVEF) was 54% and 49.3% of the patients had heart failure. Worse LVDF was associated with older age, lower LVEF, and higher rates of diabetes, hypertension, chronic pulmonary disease, renal failure, cardiovascular disease, and heart failure (p-of-trend<0.001 for all). During a median follow-up of 2.6 years (interquartile range 0.5-4.0 years), 847 (33.2%) patients died. In multivariable survival analysis, adjusted to multiple risk factors including baseline LVEF and heart failure, worse LVDF was independently associated with a higher risk of death [G1 LVDF: adjusted hazard ratio (aHR) 2.15 (95% confidence interval (95%CI) 0.75-1.77), G2 LVDF: aHR 2.25 (95%CI 1.50-3.38) G3 LVDF: aHR 2.96 (95%CI 1.88-4.65); p<0.001 for both)]. These findings were consistent also in a subgroup of patients with baseline LVEF > 50%. Conclusion Among patients with LBBB, worse LVDF, as derived by a dedicated AI-ECG algorithm is independently associated with all-cause mortality beyond baseline LVEF and clinical heart failure.survival across AI-ECG LVDF grades

Abstract Background Artificial intelligence (AI) models trained on 12-lead ECGs effectively detect left ventricular systolic dysfunction (LVSD; left ventricular ejection fraction [LVEF] <=40%). Continuous ECG monitoring via Holter recordings provides an opportunity for opportunistic screening for structural heart disease beyond rhythm disorders. We hypothesized that a lead-invariant version of the 12-lead AI model would enable a Holter monitor to screen for both arrhythmias and ventricular dysfunction. Methods We retrospectively analyzed continuous Holter ECGs from 17,665 patients who underwent a Holter and transthoracic echocardiogram (TTE) within 30 days of each other at Mayo Clinic. From each Holter, a random 20-minute of valid (non-flatline/lead disconnect) ECG segment was extracted and analyzed for LVSD detection using the adapted lead-invariant AI model. To evaluate stability, we examined model performance across different time points of the day, presenting results as area under the receiver operating characteristic curve (AUC) over time. Moreover, we illustrated the model’s robustness to noisy data by comparing its performance on raw ECG signals with that on bandpass-filtered inputs. Results Among 17,665 patients (mean age 59 years, 48.57% female), 4.96% had an LVEF <=40%. The AI model demonstrated strong predictive performance (20-minute segment AUC 0.90, mean prediction of 24-hour AUC 0.92). Analysis of results over time (Figure) revealed temporal patterns in predictive accuracy, with specific time periods showing greater stability. Despite modest variability, model performance remained consistently high throughout the day, confirming robustness across different physiological states. The predictions remained robust with noisy input. We did not observe performance improvement when the baseline wander and high frequency noise are removed by the bandpass filter. Conclusion Applying a 12-lead AI ECG model with a lead-invariant framework to a continuous Holter ECG enables effective screening for left ventricular dysfunction. This suggests that AI-based analysis of Holter-monitors can facilitate opportunistic screening of ventricular dysfunction and may enable assessment of an arrhythmia’s impact on LVEF, as well as the relationship between arrhythmia burden and LVEF.Figure 1.Mean prediction AUC of the day Figure 2.AUC for different time point

Abstract Background Left ventricular ejection fraction (LVEF) and diastolic dysfunction (DD) are key prognostic factors in patients with ST-segment elevation myocardial infarction (STEMI). AI-enabled ECG (AI-ECG) has been shown to have excellent agreement with transthoracic echocardiography (TTE) assessment and incremental prognostic value to LVEF. Methods We retrospectively analyzed 1,005 adult acute STEMI patients referred for echocardiography (10/2007–01/2025) who had ECGs within seven days before and after the reference echocardiogram. The ECG closest to the TTE date was designated as the pre-STEMI and post-STEMI ECG, respectively. AI-ECG was used to classify diastolic function grades and grades 2 and 3 were considered as DD. Patients were stratified by LVEF (≤50% vs. >50%) at the time of TTE and the association between AI-ECG-determined DD and 10-year mortality was assessed using Cox regression and Kaplan-Meier curves. Results Among 1,005 patients (mean age 66.4 ± 13.4 years, 29.9% females), AI-ECG-determined DD prior to STEMI was present in 47.9% of patients with LVEF ≤50% and 30.3% of those with preserved LVEF (>50%). DD status remained unchanged post-STEMI in 79.3% of cases, while 10.2% showed worsening diastolic function. Regardless of LVEF, AI-ECG-determined DD prior to STEMI was associated with higher mortality risk (LVEF ≤50%: HR 3.03, 95% CI 2.16–4.24, p < 0.001; LVEF >50%: HR 2.32, 95% CI 1.58–3.41, p < 0.001). Post STEMI AI-ECG-determined DD remained a significant predictor of mortality (LVEF ≤50%: HR 2.54, 95% CI 1.82–3.53, p < 0.001; LVEF >50%: HR 2.43, 95% CI 1.66–3.58, p < 0.001). Echocardiographically determined DD also correlated with higher mortality (LVEF ≤50%: HR 4.91, 95% CI 2.95–8.15, p < 0.001; LVEF >50%: HR 2.72, 95% CI 1.68–4.41, p < 0.001). In patients with LVEF ≤50%, AI-ECG-determined DD remained associated with higher mortality even after adjusting for echocardiographically determined DD (pre-STEMI: HR 1.85, 95% CI 1.23–2.80, p < 0.001; post-STEMI: HR 1.90, 95% CI 1.23–2.91, p < 0.001). Conclusion AI-ECG-determined DD is a significant predictor of mortality in STEMI, independent of LVEF. The presence of normal filling pressure may indicate a better clinical outcome in patients with reduced LVEF patients. These findings underscore the clinical utility of AI-ECG in refining risk stratification beyond conventional LVEF assessment.

Abstract Background Right ventricular (RV) systolic function is an independent predictor of outcome, but a robust automated method of quantitation is lacking. Purpose To develop and validate a deep learning model for quantifying RV systolic function from 2D echo clips and assess its prognostic value in various loading states and left-heart disease. Methods A deep learning convolutional neural network (CNN) was developed to quantitate RV systolic function using all available apical 4 chamber (A4C) views from 13,039 patients (61,187 A4C video clips) in a US institution. Videos were converted to frames of 256x256 grayscale pixels, with graphics, including the ECG, removed. The model used the S3D architecture, a 3D-CNN with many of the spatial and temporal convolutions separated. RV free wall longitudinal strain (RVFWLS) was used to define RV function: normal (≤-25%); mild (-20% to –24%); moderate (-15% to –19%) and severe RV dysfunction (0% to –14%). The AI-estimate was the average of multiple 32-frame segments for each clip identified as an A4C, including both standard and RV-focused variations. The model was validated in four distinct clinical cohorts from independent data sources: 1) 1365 patients referred to an external diagnostic echocardiography lab in a different US state; 2) 569 patients recruited prospectively with known or suspected pulmonary hypertension (PH); 3) 416 patients with moderate or greater tricuspid valve regurgitation (TR); 4) 872 patients imaged at an Eastern European University with left-sided cardiac diseases (valvular disease, cardiomyopathies, etc.), athletes, or healthy volunteers. Results Model output was strongly correlated with the presence of RV dysfunction by manual measurement [AUC 0.907 (95% CI: 0.888-0.926)] in the internal validation set. Results were similar in an external validation US cohort [N=1365; AUC 0.919 (95% CI: 0.899-0.939)]. When the model was applied to 569 patients with known or suspected PH, presence and severity of the AI measures of RV dysfunction were strongly associated with functional class, 6 min walk distance, NT-proBNP levels and right heart failure (p<0.001). The presence and severity of RV dysfunction (as categorized by the AI model) was strongly associated with all-cause mortality in patients with PH (Figure 1A) and in patients with significant TR (Figure 1B). The European cohort was a geographically and institutionally distinct cohort with a low prevalence of >mild TR (10.8%) or PH (6.4%). In this cohort the AI model predicted a reduced RVEF (< 45%) with an AUC of 0.864 (95% CI: 0.831-0.898) and associated with both all-cause mortality (Figure 2A) and a composite endpoint of heart failure hospitalization or death (Figure 2B). Conclusions A fully automated, AI-based assessment of RV systolic function, using 2D apical 4 chamber echo views alone, is highly feasible and can accurately categorize disease severity and predict outcome across a wide spectrum of cardiovascular conditions.Figure 1 Figure 2

Abstract Aims Coronary angiography might contain clinically relevant information, beyond its traditional role in delineating coronary artery disease. We sought to explore the role of deep learning (DL) in detecting right ventricular dysfunction from routinely acquired cine images during coronary angiography. Methods and results 3D-convolutional neural networks (CNN) were trained to identify right ventricle (RV) dysfunction from diagnostic angiography of the right coronary artery (RCA). The model’s input was two cine angiograms from LAO and RAO projections. The model’s output was normal vs. abnormal RV function. Ground truth on RV function was obtained from the nearest transthoracic echocardiogram to the coronary angiogram. Two analyses were performed to identify (i) any RV dysfunction (≥mild) and (ii) significant RV dysfunction (≥mild–moderate). A total of 10 336 coronary angiograms from 9849 patients (mean age 66, 36% women) were included. The cohort was split into training (70%), validation (15%), and testing (15%) sets. In the testing sets, the models attained an area under the ROC curve (AUC), sensitivity, and specificity of 0.82 (95% CI: 0.80–0.84), 0.75, and 0.74, for detecting any RV dysfunction, and 0.83 (95% CI: 0.80–0.86), 0.82, and 0.70 for detecting significant (≥mild–moderate) RV dysfunction. Combining an ECG-driven AI model with the angio-driven model improved the AUC values to 0.83 and 0.87 for detecting any or advanced RV dysfunction, respectively. Conclusion A novel deep learning algorithm yielded acceptable classification accuracy in detecting RV dysfunction from routine cine angiography of the RCA. The model’s performance improved further by adding ECG to its inputs.

Abstract Background Elevated left ventricular filling pressure (LVFP) is associated with an increased risk of heart failure and mortality. Pulmonary artery wedge pressure (PAWP) is considered the gold standard for assessing LVFP. Recently, we validated an artificial intelligence-enabled electrocardiogram (AI-ECG) to assess LV diastolic function (LVDF) and LVFP. Herein, we compare AI-ECG with PAWP in predicting long-term mortality. Methods We included adults who had comprehensive echocardiography with available ECG within 14-days and right-heart-catheterization (RHC) within 30-days of echocardiography. Elevated-LVFP was defined as PAWP>15 mmHg by RHC, grades 2-3 LVDF by echocardiography, or as adjudicated by the AI-ECG algorithm. Prognosis by AI-ECG was compared to that by echocardiography and PAWP. Results Among 4,168 patients (median age 64 years, 47.7% female), 1,841 had normal-LVFP, and 2,327 had elevated-LVFP by RHC. Echocardiography classified 1,502 patients as normal-LVFP, 1,643 as elevated-LVFP, and 1,023 as indeterminate-LVFP. AI-ECG demonstrated an area-under-the-curve of 0.879 (95%-confidence-interval 0.867-0.891) for predicting elevated-LVFP as determined by echocardiography and 0.742 (95%-confidence-interval 0.724-0.759) for predicting elevated-LVFP as determined by RHC. Over a median follow-up of 7.7 years, 2,735 (34.4%) patients died. In multivariable models, elevated-LVFP by any modality was associated with an increased risk of death (AI-ECG: adjusted-hazard-ratio 1.80, 95% confidence-interval 1.55-2.09; echocardiography: adjusted-hazard-ratio 1.74, 95% confidence-interval 1.51-2.02; PAWP: adjusted-hazard-ratio 1.26, 95% confidence-interval 1.11-1.43). AI-ECG-derived LVFP provided incremental prognostic discrimination compared to echocardiography- or RHC-derived LVFP. Conclusions AI-ECG-derived LVFP offers an accessible, affordable, and noninvasive alternative for detecting elevated LVFP, with potential independent prognostic value beyond traditional echocardiographic and invasive assessments.Survival across LVFP concordance