Detection of Left Ventricular Systolic Dysfunction From Electrocardiographic Images.

BACKGROUND: Left ventricular (LV) systolic dysfunction is associated with a >8-fold increased risk of heart failure and a 2-fold risk of premature death. The use of ECG signals in screening for LV systolic dysfunction is limited by their availability to clinicians. We developed a novel deep learning-based approach that can use ECG images for the screening of LV systolic dysfunction. METHODS: Using 12-lead ECGs plotted in multiple different formats, and corresponding echocardiographic data recorded within 15 days from the Yale New Haven Hospital between 2015 and 2021, we developed a convolutional neural network algorithm to detect an LV ejection fraction <40%. The model was validated within clinical settings at Yale New Haven Hospital and externally on ECG images from Cedars Sinai Medical Center in Los Angeles, CA; Lake Regional Hospital in Osage Beach, MO; Memorial Hermann Southeast Hospital in Houston, TX; and Methodist Cardiology Clinic of San Antonio, TX. In addition, it was validated in the prospective Brazilian Longitudinal Study of Adult Health. Gradient-weighted class activation mapping was used to localize class-discriminating signals on ECG images. RESULTS: Overall, 385 601 ECGs with paired echocardiograms were used for model development. The model demonstrated high discrimination across various ECG image formats and calibrations in internal validation (area under receiving operation characteristics [AUROCs], 0.91; area under precision-recall curve [AUPRC], 0.55); and external sets of ECG images from Cedars Sinai (AUROC, 0.90 and AUPRC, 0.53), outpatient Yale New Haven Hospital clinics (AUROC, 0.94 and AUPRC, 0.77), Lake Regional Hospital (AUROC, 0.90 and AUPRC, 0.88), Memorial Hermann Southeast Hospital (AUROC, 0.91 and AUPRC 0.88), Methodist Cardiology Clinic (AUROC, 0.90 and AUPRC, 0.74), and Brazilian Longitudinal Study of Adult Health cohort (AUROC, 0.95 and AUPRC, 0.45). An ECG suggestive of LV systolic dysfunction portended >27-fold higher odds of LV systolic dysfunction on transthoracic echocardiogram (odds ratio, 27.5 [95% CI, 22.3-33.9] in the held-out set). Class-discriminative patterns localized to the anterior and anteroseptal leads (V2 and V3), corresponding to the left ventricle regardless of the ECG layout. A positive ECG screen in individuals with an LV ejection fraction ≥40% at the time of initial assessment was associated with a 3.9-fold increased risk of developing incident LV systolic dysfunction in the future (hazard ratio, 3.9 [95% CI, 3.3-4.7]; median follow-up, 3.2 years). CONCLUSIONS: We developed and externally validated a deep learning model that identifies LV systolic dysfunction from ECG images. This approach represents an automated and accessible screening strategy for LV systolic dysfunction, particularly in low-resource settings.

Retrieval-Augmented Generation for Extracting CHA2DS2-VASc Risk Factors from Unstructured Clinical Notes in Patients with Atrial Fibrillation.

Background: Assessment of stroke risk in patients with atrial fibrillation (AF) is crucial for guiding anticoagulation therapy. CHA2DS2-VASc is a widely used score for defining this risk, but current assessments rely on manual calculation by clinicians or approximations from structured EHR data elements. Unstructured clinical notes contain rich information that could enhance risk assessment. We developed and validated a Retrieval-Augmented Generation (RAG) approach to extract CHA2DS2-VASc risk factors from unstructured notes in patients with AF. Methods: We employed a RAG architecture paired with the large language model, Llama3.1, to extract features relevant to CHA2DS2-VASc scores from unstructured notes. The model was deployed on a random set of 1,000 clinical notes (934 AF patients) from Yale New Haven Health System (YNHHS). To establish a gold standard, 2 clinicians manually reviewed and labeled CHA2DS2-VASc risk factors in a random subset of 200 notes. The CHA2DS2-VASc scores were calculated for each patient using structured data alone and by incorporating risk factors identified with RAG. We assessed performance across risk factors using macro-averaged area under the receiver operating characteristic (AUROC). For external validation, we utilized 100 manually labeled clinical notes from the MIMIC-IV database. Results: The RAG model demonstrated robust performance in extracting risk factors from clinical notes. In the 1000 clinical notes, RAG identified several risk factors more frequently than structured elements, including hypertension (82.4% vs 26.2%), stroke/TIA (62.9% vs 45.5%), vascular disease (83.4% vs 56.6%), and diabetes (84.1% vs 47.2%). In the 200 expert-annotated notes, the RAG approach achieved high performance for various risk factors, with AUROCs ranging from 0.96 to 0.98 for hypertension, diabetes, and age ≥75 years. Incorporating risk factors identified by RAG increased CHA2DS2-VASc scores compared with using structured data alone. Conclusion: An LLM-optimized RAG can accurately extract CHA2DS2-VASc risk factors from unstructured clinical notes in AF patients. This approach can enable computable risk assessment and guide appropriate anticoagulation therapy.

Artificial Intelligence-Enhanced Risk Stratification of Cancer Therapeutics-Related Cardiac Dysfunction Using Electrocardiographic Images.

Background: Risk stratification strategies for cancer therapeutics-related cardiac dysfunction (CTRCD) rely on serial monitoring by specialized imaging, limiting their scalability. We aimed to examine an application of artificial intelligence (AI) to electrocardiographic (ECG) images as a surrogate for imaging risk biomarkers, and its association with early CTRCD. Methods: Across a U.S.-based health system (2013-2023), we identified 1,550 patients (age 60 [IQR:51-69] years, 1223 [78.9%] women) without cardiomyopathy who received anthracyclines and/or trastuzumab for breast cancer or non-Hodgkin lymphoma and had ECG performed ≤12 months before treatment. We deployed a validated AI model of left ventricular systolic dysfunction (LVSD) to baseline ECG images and defined low, intermediate, and high-risk groups based on AI-ECG LVSD probabilities of <0.01, 0.01 to 0.1, and ≥0.1 (positive screen), respectively. We explored the association with early CTRCD (new cardiomyopathy, heart failure, or left ventricular ejection fraction [LVEF]<50%), or LVEF<40%, up to 12 months post-treatment. In a mechanistic analysis, we assessed the association between global longitudinal strain (GLS) and AI-ECG LVSD probabilities in studies performed within 15 days of each other. Results: Among 1,550 patients without known cardiomyopathy (median follow-up: 14.1 [IQR:13.4-17.1] months), 83 (5.4%), 562 (36.3%) and 905 (58.4%) were classified as high, intermediate, and low risk by baseline AI-ECG. A high- vs low-risk AI-ECG screen (≥0.1 vs <0.01) was associated with a 3.4-fold and 13.5-fold higher incidence of CTRCD (adj.HR 3.35 [95%CI:2.25-4.99]) and LVEF<40% (adj.HR 13.52 [95%CI:5.06-36.10]), respectively. Post-hoc analyses supported longitudinal increases in AI-ECG probabilities within 6-to-12 months of a CTRCD event. Among 1,428 temporally-linked echocardiograms and ECGs, AI-ECG LVSD probabilities were associated with worse GLS (GLS -19% [IQR:-21 to -17%] for probabilities <0.1, to -15% [IQR:-15 to -9%] for ≥0.5 [p<0.001]). Conclusions: AI applied to baseline ECG images can stratify the risk of early CTRCD associated with anthracycline or trastuzumab exposure in the setting of breast cancer or non-Hodgkin lymphoma therapy.

Artificial intelligence-enhanced risk stratification of cancer therapeutics-related cardiac dysfunction using electrocardiographic images.

Background: Risk stratification strategies for cancer therapeutics-related cardiac dysfunction (CTRCD) rely on serial monitoring by specialized imaging, limiting their scalability. Objectives: To examine an artificial intelligence (AI)-enhanced electrocardiographic (AI-ECG) surrogate for imaging risk biomarkers, and its association with CTRCD. Methods: Across a five-hospital U.S.-based health system (2013-2023), we identified patients with breast cancer or non-Hodgkin lymphoma (NHL) who received anthracyclines (AC) and/or trastuzumab (TZM), and a control cohort receiving immune checkpoint inhibitors (ICI). We deployed a validated AI model of left ventricular systolic dysfunction (LVSD) to ECG images (≥0.1, positive screen) and explored its association with i) global longitudinal strain (GLS) measured within 15 days (n=7,271 pairs); ii) future CTRCD (new cardiomyopathy, heart failure, or left ventricular ejection fraction [LVEF]<50%), and LVEF<40%. In the ICI cohort we correlated baseline AI-ECG-LVSD predictions with downstream myocarditis. Results: Higher AI-ECG LVSD predictions were associated with worse GLS (-18% [IQR:-20 to -17%] for predictions<0.1, to -12% [IQR:-15 to -9%] for ≥0.5 (p<0.001)). In 1,308 patients receiving AC/TZM (age 59 [IQR:49-67] years, 999 [76.4%] women, 80 [IQR:42-115] follow-up months) a positive baseline AI-ECG LVSD screen was associated with ~2-fold and ~4.8-fold increase in the incidence of the composite CTRCD endpoint (adj.HR 2.22 [95%CI:1.63-3.02]), and LVEF<40% (adj.HR 4.76 [95%CI:2.62-8.66]), respectively. Among 2,056 patients receiving ICI (age 65 [IQR:57-73] years, 913 [44.4%] women, follow-up 63 [IQR:28-99] months) AI-ECG predictions were not associated with ICI myocarditis (adj.HR 1.36 [95%CI:0.47-3.93]). Conclusion: AI applied to baseline ECG images can stratify the risk of CTRCD associated with anthracycline or trastuzumab exposure.

Development and multinational validation of an algorithmic strategy for high Lp(a) screening.

Elevated lipoprotein (a) (Lp(a)) is associated with premature atherosclerotic cardiovascular disease. However, fewer than 0.5% of individuals undergo Lp(a) testing, limiting the evaluation and use of novel targeted therapeutics currently under development. Here we describe the development of a machine learning model for targeted screening for elevated Lp(a) (≥150 nmol l-1) in the UK Biobank (N = 456,815), the largest cohort with protocolized Lp(a) testing. We externally validated the model in 3 large cohort studies, ARIC (N = 14,484), CARDIA (N = 4,124) and MESA (N = 4,672). The model, Algorithmic Risk Inspection for Screening Elevated Lp(a) (ARISE), reduced the number needed to test to find one individual with elevated Lp(a) by up to 67.3%, based on the probability threshold, with consistent performance across external validation cohorts. ARISE could be used to optimize screening for elevated Lp(a) using commonly available clinical features, with the potential for its deployment in electronic health records to enhance the yield of Lp(a) testing in real-world settings.

Automated Transformation of Unstructured Cardiovascular Diagnostic Reports into Structured Datasets Using Sequentially Deployed Large Language Models.

Background: Rich data in cardiovascular diagnostic testing are often sequestered in unstructured reports, with the necessity of manual abstraction limiting their use in real-time applications in patient care and research. Methods: We developed a two-step process that sequentially deploys generative and interpretative large language models (LLMs; Llama2 70b and Llama2 13b). Using a Llama2 70b model, we generated varying formats of transthoracic echocardiogram (TTE) reports from 3,000 real-world echo reports with paired structured elements, leveraging temporal changes in reporting formats to define the variations. Subsequently, we fine-tuned Llama2 13b using sequentially larger batches of generated echo reports as inputs, to extract data from free-text narratives across 18 clinically relevant echocardiographic fields. This was set up as a prompt-based supervised training task. We evaluated the fine-tuned Llama2 13b model, HeartDx-LM, on several distinct echocardiographic datasets: (i) reports across the different time periods and formats at Yale New Haven Health System (YNHHS), (ii) the Medical Information Mart for Intensive Care (MIMIC) III dataset, and (iii) the MIMIC IV dataset. We used the accuracy of extracted fields and Cohen's Kappa as the metrics and have publicly released the HeartDX-LM model. Results: The HeartDX-LM model was trained on randomly selected 2,000 synthetic echo reports with varying formats and paired structured labels, with a wide range of clinical findings. We identified a lower threshold of 500 annotated reports required for fine-tuning Llama2 13b to achieve stable and consistent performance. At YNHHS, the HeartDx-LM model accurately extracted 69,144 out of 70,032 values (98.7%) across 18 clinical fields from unstructured reports in the test set from contemporary records where paired structured data were also available. In older echo reports where only unstructured reports were available, the model achieved 87.1% accuracy against expert annotations for the same 18 fields for a random sample of 100 reports. Similarly, in expert-annotated external validation sets from MIMIC-IV and MIMIC-III, HeartDx-LM correctly extracted 201 out of 220 available values (91.3%) and 615 out of 707 available values (87.9%), respectively, from 100 randomly chosen and expert annotated echo reports from each set. Conclusion: We developed a novel method using paired large and moderate-sized LLMs to automate the extraction of unstructured echocardiographic reports into tabular datasets. Our approach represents a scalable strategy that transforms unstructured reports into computable elements that can be leveraged to improve cardiovascular care quality and enable research.

Performance of contemporary cardiovascular risk stratification scores in Brazil: an evaluation in the ELSA-Brasil study.

AIMS: Despite notable population differences in high-income and low- and middle-income countries (LMICs), national guidelines in LMICs often recommend using US-based cardiovascular disease (CVD) risk scores for treatment decisions. We examined the performance of widely used international CVD risk scores within the largest Brazilian community-based cohort study (Brazilian Longitudinal Study of Adult Health, ELSA-Brasil). METHODS: All adults 40-75 years from ELSA-Brasil (2008-2013) without prior CVD who were followed for incident, adjudicated CVD events (fatal and non-fatal MI, stroke, or coronary heart disease death). We evaluated 5 scores-Framingham General Risk (FGR), Pooled Cohort Equations (PCEs), WHO CVD score, Globorisk-LAC and the Systematic Coronary Risk Evaluation 2 score (SCORE-2). We assessed their discrimination using the area under the receiver operating characteristic curve (AUC) and calibration with predicted-to-observed risk (P/O) ratios-overall and by sex/race groups. RESULTS: There were 12 155 individuals (53.0±8.2 years, 55.3% female) who suffered 149 incident CVD events. All scores had a model AUC>0.7 overall and for most age/sex groups, except for white women, where AUC was <0.6 for all scores, with higher overestimation in this subgroup. All risk scores overestimated CVD risk with 32%-170% overestimation across scores. PCE and FGR had the highest overestimation (P/O ratio: 2.74 (95% CI 2.42 to 3.06)) and 2.61 (95% CI 1.79 to 3.43)) and the recalibrated WHO score had the best calibration (P/O ratio: 1.32 (95% CI 1.12 to 1.48)). CONCLUSION: In a large prospective cohort from Brazil, we found that widely accepted CVD risk scores overestimate risk by over twofold, and have poor risk discrimination particularly among Brazilian women. Our work highlights the value of risk stratification strategies tailored to the unique populations and risks of LMICs.

Real-world evaluation of an algorithmic machine-learning-guided testing approach in stable chest pain: a multinational, multicohort study.

Aims: An algorithmic strategy for anatomical vs. functional testing in suspected coronary artery disease (CAD) (Anatomical vs. Stress teSting decIsion Support Tool; ASSIST) is associated with better outcomes than random selection. However, in the real world, this decision is rarely random. We explored the agreement between a provider-driven vs. simulated algorithmic approach to cardiac testing and its association with outcomes across multinational cohorts. Methods and results: In two cohorts of functional vs. anatomical testing in a US hospital health system [Yale; 2013-2023; n = 130 196 (97.0%) vs. n = 4020 (3.0%), respectively], and the UK Biobank [n = 3320 (85.1%) vs. n = 581 (14.9%), respectively], we examined outcomes stratified by agreement between the real-world and ASSIST-recommended strategies. Younger age, female sex, Black race, and diabetes history were independently associated with lower odds of ASSIST-aligned testing. Over a median of 4.9 (interquartile range [IQR]: 2.4-7.1) and 5.4 (IQR: 2.6-8.8) years, referral to the ASSIST-recommended strategy was associated with a lower risk of acute myocardial infarction or death (hazard ratioadjusted: 0.81, 95% confidence interval [CI] 0.77-0.85, P < 0.001 and 0.74 [95% CI 0.60-0.90], P = 0.003, respectively), an effect that remained significant across years, test types, and risk profiles. In post hoc analyses of anatomical-first testing in the Prospective Multicentre Imaging Study for Evaluation of Chest Pain (PROMISE) trial, alignment with ASSIST was independently associated with a 17% and 30% higher risk of detecting CAD in any vessel or the left main artery/proximal left anterior descending coronary artery, respectively. Conclusion: In cohorts where historical practices largely favour functional testing, alignment with an algorithmic approach to cardiac testing defined by ASSIST was associated with a lower risk of adverse outcomes. This highlights the potential utility of a data-driven approach in the diagnostic management of CAD.

Computational Phenomapping of Randomized Clinical Trials to Enable Assessment of their Real-world Representativeness and Personalized Inference.

Importance: Randomized clinical trials (RCTs) are the standard for defining an evidence-based approach to managing disease, but their generalizability to real-world patients remains challenging to quantify. Objective: To develop a multidimensional patient variable mapping algorithm to quantify the similarity and representation of electronic health record (EHR) patients corresponding to an RCT and estimate the putative treatment effects in real-world settings based on individual treatment effects observed in an RCT. Design: A retrospective analysis of the Treatment of Preserved Cardiac Function Heart Failure with an Aldosterone Antagonist Trial (TOPCAT; 2006-2012) and a multi-hospital patient cohort from the electronic health record (EHR) in the Yale New Haven Hospital System (YNHHS; 2015-2023). Setting: A multicenter international RCT (TOPCAT) and multi-hospital patient cohort (YNHHS). Participants: All TOPCAT participants and patients with heart failure with preserved ejection fraction (HFpEF) and ≥1 hospitalization within YNHHS. Exposures: 63 pre-randomization characteristics measured across the TOPCAT and YNNHS cohorts. Main Outcomes and Measures: Real-world generalizability of the RCT TOPCAT using a multidimensional phenotypic distance metric between TOPCAT and YNHHS cohorts. Estimation of the individualized treatment effect of spironolactone use on all-cause mortality within the YNHHS cohort based on phenotypic distance from the TOPCAT cohort. Results: There were 3,445 patients in TOPCAT and 11,712 HFpEF patients across five hospital sites. Across the 63 TOPCAT variables mapped by clinicians to the EHR, there were larger differences between TOPCAT and each of the 5 EHR sites (median SMD 0.200, IQR 0.037-0.410) than between the 5 EHR sites (median SMD 0.062, IQR 0.010-0.130). The synthesis of these differences across covariates using our multidimensional similarity score also suggested substantial phenotypic dissimilarity between the TOPCAT and EHR cohorts. By phenotypic distance, a majority (55%) of TOPCAT participants were closer to each other than any individual EHR patient. Using a TOPCAT-derived model of individualized treatment benefit from spironolactone, those predicted to derive benefit and receiving spironolactone in the EHR cohorts had substantially better outcomes compared with predicted benefit and not receiving the medication (HR 0.74, 95% CI 0.62-0.89). Conclusions and Relevance: We propose a novel approach to evaluating the real-world representativeness of RCT participants against corresponding patients in the EHR across the full multidimensional spectrum of the represented phenotypes. This enables the evaluation of the implications of RCTs for real-world patients. KEY POINTS: Question: How can we examine the multi-dimensional generalizability of randomized clinical trials (RCT) to real-world patient populations?Findings: We demonstrate a novel phenotypic distance metric comparing an RCT to real-world populations in a large multicenter RCT of heart failure patients and the corresponding patients in multisite electronic health records (EHRs). Across 63 pre-randomization characteristics, pairwise assessments of members of the RCT and EHR cohorts were more discordant from each other than between members of the EHR cohort (median standardized mean difference 0.200 [0.037-0.410] vs 0.062 [0.010-0.130]), with a majority (55%) of RCT participants closer to each other than any individual EHR patient. The approach also enabled the quantification of expected real world outcomes based on effects observed in the RCT.Meaning: A multidimensional phenotypic distance metric quantifies the generalizability of RCTs to a given population while also offering an avenue to examine expected real-world patient outcomes based on treatment effects observed in the RCT.

A Deep Learning Approach for Automated Extraction of Functional Status and New York Heart Association Class for Heart Failure Patients During Clinical Encounters.

Introduction: Serial functional status assessments are critical to heart failure (HF) management but are often described narratively in documentation, limiting their use in quality improvement or patient selection for clinical trials. We developed and validated a deep learning-based natural language processing (NLP) strategy to extract functional status assessments from unstructured clinical notes. Methods: We identified 26,577 HF patients across outpatient services at Yale New Haven Hospital (YNHH), Greenwich Hospital (GH), and Northeast Medical Group (NMG) (mean age 76.1 years; 52.0% women). We used expert annotated notes from YNHH for model development/internal testing and from GH and NMG for external validation. The primary outcomes were NLP models to detect (a) explicit New York Heart Association (NYHA) classification, (b) HF symptoms during activity or rest, and (c) functional status assessment frequency. Results: Among 3,000 expert-annotated notes, 13.6% mentioned NYHA class, and 26.5% described HF symptoms. The model to detect NYHA classes achieved a class-weighted AUROC of 0.99 (95% CI: 0.98-1.00) at YNHH, 0.98 (0.96-1.00) at NMG, and 0.98 (0.92-1.00) at GH. The activity-related HF symptom model achieved an AUROC of 0.94 (0.89-0.98) at YNHH, 0.94 (0.91-0.97) at NMG, and 0.95 (0.92-0.99) at GH. Deploying the NYHA model among 166,655 unannotated notes from YNHH identified 21,528 (12.9%) with NYHA mentions and 17,642 encounters (10.5%) classifiable into functional status groups based on activity-related symptoms. Conclusions: We developed and validated an NLP approach to extract NYHA classification and activity-related HF symptoms from clinical notes, enhancing the ability to track optimal care and identify trial-eligible patients.

An Ensemble Deep Learning Algorithm for Structural Heart Disease Screening Using Electrocardiographic Images: PRESENT SHD.

Background: Identifying structural heart diseases (SHDs) early can change the course of the disease, but their diagnosis requires cardiac imaging, which is limited in accessibility. Objective: To leverage 12-lead ECG images for automated detection and prediction of multiple SHDs using a novel deep learning model. Methods: We developed a series of convolutional neural network models for detecting a range of individual SHDs from images of ECGs with SHDs defined by transthoracic echocardiograms (TTEs) performed within 30 days of the ECG at the Yale New Haven Hospital (YNHH). SHDs were defined based on TTEs with LV ejection fraction <40%, moderate-to-severe left-sided valvular disease (aortic/mitral stenosis or regurgitation), or severe left ventricular hypertrophy (IVSd > 1.5cm and diastolic dysfunction). We developed an ensemble XGBoost model, PRESENT-SHD, as a composite screen across all SHDs. We validated PRESENT-SHD at 4 US hospitals and a prospective population-based cohort study, the Brazilian Longitudinal Study of Adult Health (ELSA-Brasil), with concurrent protocolized ECGs and TTEs. We also used PRESENT-SHD for risk stratification of new-onset SHD or heart failure (HF) in clinical cohorts and the population-based UK Biobank (UKB). Results: The models were developed using 261,228 ECGs from 93,693 YNHH patients and evaluated on a single ECG from 11,023 individuals at YNHH (19% with SHD), 44,591 across external hospitals (20-27% with SHD), and 3,014 in the ELSA-Brasil (3% with SHD). In the held-out test set, PRESENT-SHD demonstrated an AUROC of 0.886 (0.877-894), sensitivity of 89%, and specificity of 66%. At hospital-based sites, PRESENT-SHD had AUROCs ranging from 0.854-0.900, with sensitivities and specificities of 93-96% and 51-56%, respectively. The model generalized well to ELSA-Brasil (AUROC, 0.853 [0.811-0.897], sensitivity 88%, specificity 62%). PRESENT-SHD performance was consistent across demographic subgroups. A positive PRESENT-SHD screen portended a 2- to 4-fold higher risk of new-onset SHD/HF, independent of demographics, comorbidities, and the competing risk of death across clinical sites and UKB, with high predictive discrimination. Conclusion: We developed and validated PRESENT-SHD, an AI-ECG tool identifying a range of SHD using images of 12-lead ECGs, representing a robust, scalable, and accessible modality for automated SHD screening and risk stratification. CONDENSED ABSTRACT: Screening for structural heart disorders (SHDs) requires cardiac imaging, which has limited accessibility. To leverage 12-lead ECG images for automated detection and prediction of multiple SHDs, we developed PRESENT-SHD, an ensemble deep learning-based model. PRESENT-SHD demonstrated excellent performance for detecting SHDs across 5 US hospitals and a population-based cohort in Brazil. The model successfully predicted the risk of new-onset SHD or heart failure in both US clinical cohorts and the community-based UK Biobank. By using ubiquitous ECG images to predict a composite outcome of multiple SHDs, PRESENT-SHD establishes a scalable paradigm for cardiovascular screening and risk stratification.

Development and Multinational Validation of an Ensemble Deep Learning Algorithm for Detecting and Predicting Structural Heart Disease Using Noisy Single-lead Electrocardiograms.

Background and Aims: AI-enhanced 12-lead ECG can detect a range of structural heart diseases (SHDs) but has a limited role in community-based screening. We developed and externally validated a noise-resilient single-lead AI-ECG algorithm that can detect SHD and predict the risk of their development using wearable/portable devices. Methods: Using 266,740 ECGs from 99,205 patients with paired echocardiographic data at Yale New Haven Hospital, we developed ADAPT-HEART, a noise-resilient, deep-learning algorithm, to detect SHD using lead I ECG. SHD was defined as a composite of LVEF<40%, moderate or severe left-sided valvular disease, and severe LVH. ADAPT-HEART was validated in four community hospitals in the US, and the population-based cohort of ELSA-Brasil. We assessed the model's performance as a predictive biomarker among those without baseline SHD across hospital-based sites and the UK Biobank. Results: The development population had a median age of 66 [IQR, 54-77] years and included 49,947 (50.3%) women, with 18,896 (19.0%) having any SHD. ADAPT-HEART had an AUROC of 0.879 (95% CI, 0.870-0.888) with good calibration for detecting SHD in the test set, and consistent performance in hospital-based external sites (AUROC: 0.852-0.891) and ELSA-Brasil (AUROC: 0.859). Among those without baseline SHD, high vs. low ADAPT-HEART probability conferred a 2.8- to 5.7-fold increase in the risk of future SHD across data sources (all P<0.05). Conclusions: We propose a novel model that detects and predicts a range of SHDs from noisy single-lead ECGs obtainable on portable/wearable devices, providing a scalable strategy for community-based screening and risk stratification for SHD.

Using Artificial Intelligence to Predict Heart Failure Risk from Single-lead Electrocardiographic Signals: A Multinational Assessment.

Importance: Despite the availability of disease-modifying therapies, scalable strategies for heart failure (HF) risk stratification remain elusive. Portable devices capable of recording single-lead electrocardiograms (ECGs) can enable large-scale community-based risk assessment. Objective: To evaluate an artificial intelligence (AI) algorithm to predict HF risk from noisy single-lead ECGs. Design: Multicohort study. Setting: Retrospective cohort of individuals with outpatient ECGs in the integrated Yale New Haven Health System (YNHHS) and prospective population-based cohorts of UK Biobank (UKB) and Brazilian Longitudinal Study of Adult Health (ELSA-Brasil). Participants: Individuals without HF at baseline. Exposures: AI-ECG-defined risk of left ventricular systolic dysfunction (LVSD). Main Outcomes and Measures: Among individuals with ECGs, we isolated lead I ECGs and deployed a noise-adapted AI-ECG model trained to identify LVSD. We evaluated the association of the model probability with new-onset HF, defined as the first HF hospitalization. We compared the discrimination of AI-ECG against the pooled cohort equations to prevent HF (PCP-HF) score for new-onset HF using Harrel's C-statistic, integrated discrimination improvement (IDI), and net reclassification improvement (NRI). Results: There were 194,340 YNHHS patients (age 56 years [IQR, 41-69], 112,082 women [58%]), 42,741 UKB participants (65 years [59-71], 21,795 women [52%]), and 13,454 ELSA-Brasil participants (56 years [41-69], 7,348 women [55%]) with baseline ECGs. A total of 3,929 developed HF in YNHHS over 4.5 years (2.6-6.6), 46 in UKB over 3.1 years (2.1-4.5), and 31 in ELSA-Brasil over 4.2 years (3.7-4.5). A positive AI-ECG screen was associated with a 3- to 7-fold higher risk for HF, and each 0.1 increment in the model probability portended a 27-65% higher hazard across cohorts, independent of age, sex, comorbidities, and competing risk of death. AI-ECG's discrimination for new-onset HF was 0.725 in YNHHS, 0.792 in UKB, and 0.833 in ELSA-Brasil. Across cohorts, incorporating AI-ECG predictions in addition to PCP-HF resulted in improved Harrel's C-statistic (Δ=0.112-0.114), with an IDI of 0.078-0.238 and an NRI of 20.1%-48.8% for AI-ECG vs. PCP-HF. Conclusions and Relevance: Across multinational cohorts, a noise-adapted AI model with lead I ECGs as the sole input defined HF risk, representing a scalable portable and wearable device-based HF risk-stratification strategy.

Study Protocol for the Artificial Intelligence-Driven Evaluation of Structural Heart Diseases Using Wearable Electrocardiogram (ID-SHD).

Introduction: Portable devices capable of electrocardiogram (ECG) acquisition have the potential to enhance structural heart disease (SHD) management by enabling early detection through artificial intelligence-ECG (AI-ECG) algorithms. However, the performance of these AI algorithms for identifying SHD in a real-world screening setting is unknown. To address this gap, we aim to evaluate the validity of our wearable-adapted AI algorithm, which has been previously developed and validated for detecting SHD from single-lead portable ECGs in patients undergoing routine echocardiograms in the Yale New Haven Hospital (YNHH). Research Methods and Analysis: This is the protocol for a cross-sectional study in the echocardiographic laboratories of YNHH. The study will enroll 585 patients referred for outpatient transthoracic echocardiogram (TTE) as part of their routine clinical care. Patients expressing interest in participating in the study will undergo a screening interview, followed by enrollment upon meeting eligibility criteria and providing informed consent. During their routine visit, patients will undergo a 1-lead ECG with two devices - one with an Apple Watch and the second with another portable 1-lead ECG device. With participant consent, these 1-lead ECG data will be linked to participant demographic and clinical data recorded in the YNHH electronic health records (EHR). The study will assess the performance of the AI-ECG algorithm in identifying SHD, including left ventricular systolic dysfunction (LVSD), valvular disease and severe left ventricular hypertrophy (LVH), by comparing the algorithm's results with data obtained from TTE, which is the established gold standard for diagnosing SHD. Ethics and Dissemination: All patient EHR data required for assessing eligibility and conducting the AI-ECG will be accessed through secure servers approved for protected health information. Data will be maintained on secure, encrypted servers for a minimum of five years after the publication of our findings in a peer-reviewed journal, and any unanticipated adverse events or risks will be reported by the principal investigator to the Yale Institutional Review Board, which has reviewed and approved this protocol (Protocol Number: 2000035532).

A Multicenter Evaluation of the Impact of Therapies on Deep Learning-based Electrocardiographic Hypertrophic Cardiomyopathy Markers.

Background: Artificial intelligence-enhanced electrocardiography (AI-ECG) can identify hypertrophic cardiomyopathy (HCM) on 12-lead ECGs and offers a novel way to monitor treatment response. While the surgical or percutaneous reduction of the interventricular septum (SRT) represented initial HCM therapies, mavacamten offers an oral alternative. Objective: To evaluate biological response to SRT and mavacamten. Methods: We applied an AI-ECG model for HCM detection to ECG images from patients who underwent SRT across three sites: Yale New Haven Health System (YNHHS), Cleveland Clinic Foundation (CCF), and Atlantic Health System (AHS); and to ECG images from patients receiving mavacamten at YNHHS. Results: A total of 70 patients underwent SRT at YNHHS, 100 at CCF, and 145 at AHS. At YNHHS, there was no significant change in the AI-ECG HCM score before versus after SRT (pre-SRT: median 0.55 [IQR 0.24-0.77] vs post-SRT: 0.59 [0.40-0.75]). The AI-ECG HCM scores also did not improve post SRT at CCF (0.61 [0.32-0.79] vs 0.69 [0.52-0.79]) and AHS (0.52 [0.35-0.69] vs 0.61 [0.49-0.70]). Among 36 YNHHS patients on mavacamten therapy, the median AI-ECG score before starting mavacamten was 0.41 (0.22-0.77), which decreased significantly to 0.28 (0.11-0.50, p <0.001 by Wilcoxon signed-rank test) at the end of a median follow-up period of 237 days. Conclusions: The lack of improvement in AI-based HCM score with SRT, in contrast to a significant decrease with mavacamten, suggests the potential role of AI-ECG for serial monitoring of pathophysiological improvement in HCM at the point-of-care using ECG images.

Automated Diagnostic Reports from Images of Electrocardiograms at the Point-of-Care.

Timely and accurate assessment of electrocardiograms (ECGs) is crucial for diagnosing, triaging, and clinically managing patients. Current workflows rely on a computerized ECG interpretation using rule-based tools built into the ECG signal acquisition systems with limited accuracy and flexibility. In low-resource settings, specialists must review every single ECG for such decisions, as these computerized interpretations are not available. Additionally, high-quality interpretations are even more essential in such low-resource settings as there is a higher burden of accuracy for automated reads when access to experts is limited. Artificial Intelligence (AI)-based systems have the prospect of greater accuracy yet are frequently limited to a narrow range of conditions and do not replicate the full diagnostic range. Moreover, these models often require raw signal data, which are unavailable to physicians and necessitate costly technical integrations that are currently limited. To overcome these challenges, we developed and validated a format-independent vision encoder-decoder model - ECG-GPT - that can generate free-text, expert-level diagnosis statements directly from ECG images. The model shows robust performance, validated on 2.6 million ECGs across 6 geographically distinct health settings: (1) 2 large and diverse US health systems- Yale-New Haven and Mount Sinai Health Systems, (2) a consecutive ECG dataset from a central ECG repository from Minas Gerais, Brazil, (3) the prospective cohort study, UK Biobank, (4) a Germany-based, publicly available repository, PTB-XL, and (5) a community hospital in Missouri. The model demonstrated consistently high performance (AUROC≥0.81) across a wide range of rhythm and conduction disorders. This can be easily accessed via a web-based application capable of receiving ECG images and represents a scalable and accessible strategy for generating accurate, expert-level reports from images of ECGs, enabling accurate triage of patients globally, especially in low-resource settings.

Scalable Risk Stratification for Heart Failure Using Artificial Intelligence applied to 12-lead Electrocardiographic Images: A Multinational Study.

Background: Current risk stratification strategies for heart failure (HF) risk require either specific blood-based biomarkers or comprehensive clinical evaluation. In this study, we evaluated the use of artificial intelligence (AI) applied to images of electrocardiograms (ECGs) to predict HF risk. Methods: Across multinational longitudinal cohorts in the integrated Yale New Haven Health System (YNHHS) and in population-based UK Biobank (UKB) and Brazilian Longitudinal Study of Adult Health (ELSA-Brasil), we identified individuals without HF at baseline. Incident HF was defined based on the first occurrence of an HF hospitalization. We evaluated an AI-ECG model that defines the cross-sectional probability of left ventricular dysfunction from a single image of a 12-lead ECG and its association with incident HF. We accounted for the competing risk of death using the Fine-Gray subdistribution model and evaluated the discrimination using Harrel's c-statistic. The pooled cohort equations to prevent HF (PCP-HF) were used as a comparator for estimating incident HF risk. Results: Among 231,285 individuals at YNHHS, 4472 had a primary HF hospitalization over 4.5 years (IQR 2.5-6.6) of follow-up. In UKB and ELSA-Brasil, among 42,741 and 13,454 people, 46 and 31 developed HF over a follow-up of 3.1 (2.1-4.5) and 4.2 (3.7-4.5) years, respectively. A positive AI-ECG screen portended a 4-fold higher risk of incident HF among YNHHS patients (age-, sex-adjusted HR [aHR] 3.88 [95% CI, 3.63-4.14]). In UKB and ELSA-Brasil, a positive-screen ECG portended 13- and 24-fold higher hazard of incident HF, respectively (aHR: UKBB, 12.85 [6.87-24.02]; ELSA-Brasil, 23.50 [11.09-49.81]). The association was consistent after accounting for comorbidities and the competing risk of death. Higher model output probabilities were progressively associated with a higher risk for HF. The model's discrimination for incident HF was 0.718 in YNHHS, 0.769 in UKB, and 0.810 in ELSA-Brasil. Across cohorts, incorporating model probability with PCP-HF yielded a significant improvement in discrimination over PCP-HF alone. Conclusions: An AI model applied to images of 12-lead ECGs can identify those at elevated risk of HF across multinational cohorts. As a digital biomarker of HF risk that requires just an ECG image, this AI-ECG approach can enable scalable and efficient screening for HF risk.

Comparative Effectiveness of Second-line Antihyperglycemic Agents for Cardiovascular Outcomes: A Large-scale, Multinational, Federated Analysis of the LEGEND-T2DM Study.

Background: SGLT2 inhibitors (SGLT2is) and GLP-1 receptor agonists (GLP1-RAs) reduce major adverse cardiovascular events (MACE) in patients with type 2 diabetes mellitus (T2DM). However, their effectiveness relative to each other and other second-line antihyperglycemic agents is unknown, without any major ongoing head-to-head trials. Methods: Across the LEGEND-T2DM network, we included ten federated international data sources, spanning 1992-2021. We identified 1,492,855 patients with T2DM and established cardiovascular disease (CVD) on metformin monotherapy who initiated one of four second-line agents (SGLT2is, GLP1-RAs, dipeptidyl peptidase 4 inhibitor [DPP4is], sulfonylureas [SUs]). We used large-scale propensity score models to conduct an active comparator, target trial emulation for pairwise comparisons. After evaluating empirical equipoise and population generalizability, we fit on-treatment Cox proportional hazard models for 3-point MACE (myocardial infarction, stroke, death) and 4-point MACE (3-point MACE + heart failure hospitalization) risk, and combined hazard ratio (HR) estimates in a random-effects meta-analysis. Findings: Across cohorts, 16·4%, 8·3%, 27·7%, and 47·6% of individuals with T2DM initiated SGLT2is, GLP1-RAs, DPP4is, and SUs, respectively. Over 5·2 million patient-years of follow-up and 489 million patient-days of time at-risk, there were 25,982 3-point MACE and 41,447 4-point MACE events. SGLT2is and GLP1-RAs were associated with a lower risk for 3-point MACE compared with DPP4is (HR 0·89 [95% CI, 0·79-1·00] and 0·83 [0·70-0·98]), and SUs (HR 0·76 [0·65-0·89] and 0·71 [0·59-0·86]). DPP4is were associated with a lower 3-point MACE risk versus SUs (HR 0·87 [0·79-0·95]). The pattern was consistent for 4-point MACE for the comparisons above. There were no significant differences between SGLT2is and GLP1-RAs for 3-point or 4-point MACE (HR 1·06 [0·96-1·17] and 1·05 [0·97-1·13]). Interpretation: In patients with T2DM and established CVD, we found comparable cardiovascular risk reduction with SGLT2is and GLP1-RAs, with both agents more effective than DPP4is, which in turn were more effective than SUs. These findings suggest that the use of GLP1-RAs and SGLT2is should be prioritized as second-line agents in those with established CVD. Funding: National Institutes of Health, United States Department of Veterans Affairs.

Comparative Effectiveness of Second-Line Antihyperglycemic Agents for Cardiovascular Outcomes: A Multinational, Federated Analysis of LEGEND-T2DM.

BACKGROUND: Sodium-glucose cotransporter 2 inhibitors (SGLT2is) and glucagon-like peptide-1 receptor agonists (GLP-1 RAs) reduce the risk of major adverse cardiovascular events (MACE) in patients with type 2 diabetes mellitus (T2DM). However, their effectiveness relative to each other and other second-line antihyperglycemic agents is unknown, without any major ongoing head-to-head clinical trials. OBJECTIVES: The aim of this study was to compare the cardiovascular effectiveness of SGLT2is, GLP-1 RAs, dipeptidyl peptidase-4 inhibitors (DPP4is), and clinical sulfonylureas (SUs) as second-line antihyperglycemic agents in T2DM. METHODS: Across the LEGEND-T2DM (Large-Scale Evidence Generation and Evaluation Across a Network of Databases for Type 2 Diabetes Mellitus) network, 10 federated international data sources were included, spanning 1992 to 2021. In total, 1,492,855 patients with T2DM and cardiovascular disease (CVD) on metformin monotherapy were identified who initiated 1 of 4 second-line agents (SGLT2is, GLP-1 RAs, DPP4is, or SUs). Large-scale propensity score models were used to conduct an active-comparator target trial emulation for pairwise comparisons. After evaluating empirical equipoise and population generalizability, on-treatment Cox proportional hazards models were fit for 3-point MACE (myocardial infarction, stroke, and death) and 4-point MACE (3-point MACE plus heart failure hospitalization) risk and HR estimates were combined using random-effects meta-analysis. RESULTS: Over 5.2 million patient-years of follow-up and 489 million patient-days of time at risk, patients experienced 25,982 3-point MACE and 41,447 4-point MACE. SGLT2is and GLP-1 RAs were associated with lower 3-point MACE risk than DPP4is (HR: 0.89 [95% CI: 0.79-1.00] and 0.83 [95% CI: 0.70-0.98]) and SUs (HR: 0.76 [95% CI: 0.65-0.89] and 0.72 [95% CI: 0.58-0.88]). DPP4is were associated with lower 3-point MACE risk than SUs (HR: 0.87; 95% CI: 0.79-0.95). The pattern for 3-point MACE was also observed for the 4-point MACE outcome. There were no significant differences between SGLT2is and GLP-1 RAs for 3-point or 4-point MACE (HR: 1.06 [95% CI: 0.96-1.17] and 1.05 [95% CI: 0.97-1.13]). CONCLUSIONS: In patients with T2DM and CVD, comparable cardiovascular risk reduction was found with SGLT2is and GLP-1 RAs, with both agents more effective than DPP4is, which in turn were more effective than SUs. These findings suggest that the use of SGLT2is and GLP-1 RAs should be prioritized as second-line agents in those with established CVD.