Citations

Background: Atrial septal defects (ASD) and ventricular septal defects (VSD) are among the most common forms of congenital heart disease (CHD). Early detection and intervention for these defects are crucial to prevent complications. Given advancements in AI technology, the development of AI-enabled ECG models for detecting these defects is of significant interest. Aims: This study aimed to create AI models using pediatric ECG data for the detection of ASD, VSD, and a combined model detecting either ASD or VSD. Methods: We analyzed a dataset of 7,795 pediatric patients from Mayo Clinic who had undergone echocardiography between 2002 and 2022, with corresponding 10-second, 12-lead surface ECGs taken within 14 days of echocardiography. This dataset included 259 patients with ASDs and 139 patients with VSDs (single or multiple). The control group was comprised of 7,397 patients without significant or complex CHD. For cases and controls, we included patients with all forms of electrocardiographic abnormalities and those with valvular lesions who had less than moderate stenosis or regurgitation. The dataset was divided into training, validation, and testing subsets to develop convolutional neural network (CNN) models for ASD, VSD, and either ASD or VSD detection. Results: Our ASD and VSD detection models achieved test AUCs of 0.82 (95% CI 0.76, 0.88) and 0.80 (95% CI 0.70, 0.90), respectively. Meanwhile, a combined model that detected either ASD or VSD demonstrated a test AUC of 0.81 (95% CI, 0.76, 0.86), sensitivity of 0.73, specificity of 0.76, PPV of 0.14, and NPV of 0.98 (Figure 1). Conclusions: AI-enabled ECG models demonstrate promising performance in detecting ASD and VSD in pediatric patients. A combined model that detects either ASD or VSD also shows robust results. Future research should focus on multi-center validation of these models, refining their diagnostic capacities, and investigating their potential for early intervention in CHDs.

Introduction: Left and right ventricular systolic dysfunction (LVSD and RVSD respectively) contribute to considerable pediatric morbidity. Both LVSD and RVSD can occur in children before symptom onset, emphasizing the need for early detection. Existing deep-learning algorithms can identify LVSD in adults using 12-lead ECG; however, their efficacy in children is uncertain. Aims: We aimed to create a deep-learning model for LVSD and RVSD detection in pediatric patients using an AI-enabled ECG and compare novel pediatric models to a previously validated adult LVSD model. Methods: We identified 10,214 pediatric patients at the Mayo Clinic with a transthoracic echocardiogram (TTE) and a 10-second, 12-lead surface ECG within 14 days of TTE, performed from 2002 -2022. Our cohort comprised the first TTE-ECG pair from each patient, totaling 138 with EF ≤ 35%, 312 with EF ≤ 50%, 482 with RVSD, and 9,567 without any LVSD or RVSD. The cohort was split into training, validation, and testing datasets to develop neural networks for systolic dysfunction detection. Results: The cohort was 52.3% male, with an average age of 10.3 years. Novel models generated in the pediatric cohort demonstrated excellent performance in systolic dysfunction detection (Figure 1). Pediatric models for EF ≤ 35% and EF ≤ 50% achieved test AUCs of 0.92 (95% CI 0.89, 0.95) and 0.89 (95% CI 0.85, 0.94) respectively. The RVSD model reached a test AUC of 0.91 (95% CI 0.88, 0.94). A combined model for EF ≤ 35% or RVSD exhibited a test AUC of 0.90 (95% CI 0.87, 0.93), sensitivity 0.88, specificity 0.82, PPV 0.20, and NPV 0.95. A combined model for EF ≤ 50% or RVSD exhibited a test AUC of 0.88 (95% CI 0.85, 0.91). The previously validated adult data-derived model achieved an AUC of 0.86 (95% CI 0.83, 0.89) for EF ≤ 50% in children. Conclusions: AI-enabled ECGs demonstrate efficient detection of both LVSD and RVSD in pediatric patients. Future multi-site validation is needed to further establish the reliability of these models.

Background: Our team has previously developed convolutional neural networks (CNNs) to estimate age and sex from a 10-second, 12-lead ECG as indicators of patient wellness. Here, we develop an additional wellness network to estimate body mass index (BMI) from the ECG input signal. Aims: To evaluate the performance of neural networks trained to classify and directly estimate BMI. Methods: We identified all adult patients in the Mayo Clinic system who had a 10-second, 12-lead ECG within 15 days of a BMI measurement. Of the&nbsp;174,863 patients in the cohort, 2,895 were underweight (BMI < 18.5), 42,387 were of normal weight (BMI 18.5-24.9), and 57,939 were overweight only (BMI 25.0-29.9). Among patients categorized as obese, 39,788 had class I obesity (BMI 30.0-34.9), 18,978 had class II obesity (BMI 35.0-39.9), and 12,876 had class III obesity (BMI ≥ 40.0). The cohort was split into training, validation, and testing datasets in an 8:1:1 ratio to develop networks for BMI classification and a direct&nbsp;estimation of BMI. Results: In detecting overweight in the testing set, the model achieved an AUC of 0.86 (95% CI 0.85, 0.87). In detecting class I, class II, and class III obesity, the model yielded AUCs of 0.84 (95% CI 0.83, 0.85), 0.84 (95% CI 0.83, 0.85), and 0.85 (95% CI 0.85, 0.86), respectively. The model to estimate BMI achieved a mean absolute error of 3.70. Conclusions: AI-enabled ECGs demonstrate efficient detection of overweight patients and all classes of obesity. Further research is needed to determine if discrepancies between the traditionally measured BMI and the AI-ECG BMI reflect the presence or lack of metabolic abnormalities.

Background: Enlargement or dilation of the left ventricle (LV) is associated with increased cardiovascular morbidity and mortality, and is an established precursor of ventricular dysfunction and heart failure in almost all major cardiovascular conditions. Although there are existing artificial intelligence frameworks that make use of deep learning to detect LV enlargement, many entail LV segmentation, which can be error-prone especially for point-of-care ultrasound (POCUS) images. Goal: To develop a fully end-to-end deep learning framework that takes as input either transthoracic echocardiographic (TTE) or POCUS 2D image clips, and classifies directly (i.e., without first segmenting the LV) whether a patient has an enlarged LV. Methods: Using a sample of retrospective TTE data enriched for LV enlargement (out of 13,868 patients total, 3,351 had an enlarged LV defined using sex-specific guideline-directed criteria), we trained a classifier that takes as input apical 2-chamber and parasternal long axis videos, and provides as output a binary label denoting whether a patient has an enlarged LV. Then, we evaluated the trained model using a holdout testing dataset of retrospective TTE data (2,087 patients) as well as data from a prospective cohort of 428 patients from whom TTE and POCUS were collected simultaneously. Results: For the retrospective TTE testing dataset, the AUC was 0.851 (95% confidence interval (CI) of 0.830-871). For the prospective cohort, the model performed comparably for both TTE and POCUS video clips (TTE: AUC of 0.865, 95% CI of 0.823-0.906; POCUS: AUC of 0.840, 95% CI of 0.797-0.883). Conclusions: The current deep learning model performed comparably well for both TTE and POCUS echocardiograms. We regard these results as a proof-of-concept of the potential utility of this model in helping point-of-care clinicians make decisions regarding whether patients should receive further evaluation for LV enlargement and associated conditions.

Introduction: Artificial intelligence applied to electrocardiograms (ECG-AI) enables early and efficient detection of cardiovascular disease. The technical and clinical impact of different approaches to designing ECG-AI remains under-explored. We consider left ventricular hypertrophy (LVH) and examine the impact of disease-specific labeling of minority etiologies of left ventricular wall thickening- hypertrophic cardiomyopathy (HCM) and cardiac amyloidosis (CA) - on ECG-AI performance. Hypothesis: ECG-AI designed to detect HCM or CA outperforms ECG-AI broadly trained to identify LVH due to differences in attention to key ECG features associated with HCM or CA. Methods: Independent neural networks were trained to identify HCM, CA, and LVH. HCM and CA patients were identified using the full depth of EHR data and validated via clinician adjudication (N = 3,025 and 4,251 patients, respectively). Moderate and severe LVH patients (N = 98,144) were identified using echocardiogram results and international LVH criteria. We compared the ability of disease-specific ECG-AI vs. LVH ECG-AI to detect HCM (or CA) by examining model performance within each of the three test sets. Results: 52.6% of HCM patients and 49.5% of CA patients met LVH criteria. Within the HCM test set, the AUROC, sensitivity, specificity, and positive predictive value (PPV) of the HCM ECG-AI were 0.94, 0.88, 0.85, and 60%, respectively, whereas the same for LVH ECG-AI were 0.83, 0.83, 0.64, and 37%. Within the CA test set, CA ECG-AI achieved AUROC, sensitivity, specificity, and PPV of 0.94, 0.81, 0.91, and 69%, respectively, whereas the same for LVH ECG-AI were 0.75, 0.73, 0.68, and 31%. ECGs positive for HCM exhibited higher P wave area, QRS deflection and R wave amplitude compared to LVH-positive ECGs (Cohen's D > 0.75). ECGs positive for CA exhibited lower R wave amplitude/area, QRS deflection and QRS areas (Cohen's D > 0.6). Conclusion: Models trained using carefully labeled HCM or CA identify distinct signals associated with the diseases, resulting in improved disease discrimination compared to a model trained to identify echocardiographic left ventricular thickening. Labeling choices should be communicated transparently to clinicians to ensure safe and effective use of ECG-AI.

Introduction: Cardiotoxicity is a serious side effect of anthracycline treatment, most commonly manifesting as a reduction in left ventricular ejection fraction (LVEF) on cardiac imaging. Cost-effective alternatives to cardiac imaging are currently not available. Artificial intelligence (AI) models have been developed to detect a reduced LVEF from 12-lead electrocardiograms (ECGs) in the general population. Hypothesis: AI-ECG model can detect a reduced LVEF after anthracycline therapy (i.e. cardiotoxicity). Methods: We included an unbiased oncological population of 989 female breast cancer patients receiving (neo)adjuvant anthracycline chemotherapy. Average follow-up time was 9.83±4.2 years. Patients with incomplete transthoracic echocardiography (TTE) data, HER-2-directed therapy, significant comorbidities, upfront metastatic disease, second primary malignancy or pre-existing cardiovascular disease were excluded from the analyses. Patients with LVEF decline for reasons other than anthracycline-induced cardiotoxicity were also excluded. Primary readout was the diagnostic performance of AI-ECG by area under the curve (AUC) for various levels of LVEF <50%. Results: Using 1,877 ECGs from 703 unique patients, the AI-ECG model detected an LVEF <50% and ≤35% after anthracycline therapy with an AUC of 0.93 and 0.94, respectively. Conclusions: These data support the use of AI-ECG for cardiomyopathy (cardiotoxicity) screening after anthracycline-based chemotherapy. This technology could serve as a gatekeeper to more costly testing and could enable patients to monitor themselves over long periods of time.

Introduction: Biologic sex and hormonal concentrations shape ECG parameters due to sex hormone effects on cardiac function. AI-enabled ECGs have previously been shown to accurately predict patient sex in adults and correlate with sex hormone levels. Aims: We aimed to test the ability of AI-enabled ECGs to predict pediatric patient sex and explore the influence of pubertal development. Methods: Two AI-enabled ECG models were created using a convolutional neural network trained on pediatric 10-second, 12-lead surface ECGs. The first model was trained de novo using pediatric data. The second model used transfer learning from a previously validated adult data-derived algorithm. We analyzed the first ECG from 89,063 pediatric patients (aged <19 years) recorded from 1988-2022, and divided the cohort into training, validation, and testing datasets. Subgroup analysis was performed on prepubertal (0-7 years), pubertal (8-14 years), and post-pubertal (15-18 years) patients. Results: The cohort consisted of 46.6% males, with 21,434 prepubertal, 26,300 pubertal, and 41,329 post-pubertal children. No significant performance difference was observed between the two models. Both models demonstrated 81% accuracy and an overall AUC of 0.91 in the testing cohorts, with high discriminatory ability in post-pubertal teenagers (AUC = 0.98). (Figure 1) Conclusions: AI-ECG predicts pediatric patient sex with varying performance across pubertal stages. The highest discriminatory ability is observed post-puberty, decreasing in pubertal and prepubertal children. The results suggest hormonal and physiological changes during puberty may influence cardiac electrophysiology measurably. More research is required to interpret CNN-predicted sex discordance with actual sex and its implications for future cardiovascular risk.

Introduction: Over 100 FDA-approved medications, electrolyte perturbations, and many disease states can prolong the QT interval in up to 10% of patients. In contrast, approximately 1 in 2,000 people have congenital long QT syndrome (LQTS) hallmarked by pathological QT prolongation secondary to LQTS-causative mutations. Hypothesis: an artificial intelligence (AI) deep neural network (DNN) analysis of the 12-lead ECG can distinguish patients with LQTS from those with acquired QT prolongation. Methods: The study cohort included 1599 patients with genetically confirmed LQTS and over 2.5 million controls from Mayo Clinic’s ECG data vault. Every patient/control with ≥ 1 ECG above age- and sex- specific 99 th percentile values for QTc [> 460 ms for all patients (male/female) < 13 years of age, or > 470 ms for men and > 480 ms for women above this age] was included. An AI-DNN involving a multi-layer convolutional neural network was developed. To simulate screening conditions, patients were matched at a ratio of 1:2,000 (incidence of LQTS) or 1:200 (balance of LQTS vs acquired QT prolongation in a tertiary referral center). Results: Among the 1,599 patients with LQTS, 808 (> 50%) had ≥ 1 ECG with QTc above the aforementioned QTc thresholds (2,987 ECGs) compared to 361,069/2.5M controls (14% of Mayo Clinic patients getting an ECG, 989,313 ECGs). Following age- and sex- matching and splitting, the model successfully distinguished LQTS from those with acquired QT prolongation at 1:2,000 matching with an AUC of 0.937 (accuracy 89%, sensitivity 81%, specificity 90%, PPV 0.05, NPV 0.99). Furthermore, when matching at a rate that genetically-mediated QT prolongation would be encountered in clinic (1:200), the model still successfully separated the two groups (AUC 0.912, accuracy 88%, sensitivity 78%, specificity 89%, PPV 0.1, NPV 0.99). Conclusion: For patients with a QTc exceeding its 99 th percentile values seen in health, this novel AI-DNN 12-lead ECG distinguishes abnormal QT prolongation stemming from LQTS versus acquired QT prolongation with high performance characteristics (AUC > 0.93). Even when scaled to referral center and LQTS incidence ratios, a negative AI-DNN signal rules out the presence of a LQTS disease-causative mutation with 99% negative predictive value.

Introduction: Artificial Intelligence-enabled electrocardiogram (AI-ECG) can estimate the probability of having atrial fibrillation (AFib) while in sinus rhythm. Recurrent stroke in subjects without known AFib may be due to undiagnosed AFib. We evaluated the association between AI-ECG AFib and recurrent ischemic stroke and with incident AFib Methods: Consecutive adults admitted to a Mayo Clinic or Mayo Clinic Health System hospital or to Olmsted Community Hospital for ischemic stroke between January,1,2000 and June,30,2022 were identified using the Rochester Epidemiology Project. Recurrent ischemic strokes were identified after reviewing medical records and imaging studies. We excluded those with known AFib or without ECGs within 3 years prior to the first ischemic stroke. The AI-ECG network was previously developed using 649,931 ECGs from > 180,000 subjects. The primary endpoint was recurrent ischemic stroke; secondary endpoint was incident AFib. Cox model adjusted for age, sex, and CHA2DS2-VASc2 score variables. Results: We included 4,816 subjects, mean age 70±15 yr, 54% male, 92% white, median CHA2DS2-VASc2 score 4 [3-5], median follow-up 3 [1-7] yr. 280 subjects (6%) had a second stroke and 366 (8%) were diagnosed with incident AFib over follow-up. Those with a positive AI-ECG AFib had an increased risk for recurrent ischemic stroke(HR 1.54, 95% CI 1.21-1.96, P<0.001 Figure A) This association remained significant after adjustment for CHA2DS2-VASc2 score(HR 1.37 95% CI 1.07-1.76, P=0.01). These associations were not affected by anticoagulation use over follow-up. AI-ECG AFib was also associated with incident AFib(HR 2.42 95% CI 1.96-2.98, P<0.001 Figure B) Conclusions: AI-ECG AFib predicts recurrent ischemic stroke in the community, possibly reflecting undiagnosed AFib that can be detected by AI-ECG, as it also predicts incident AFib. These results underscore the potential utility of AI-ECG for risk stratification of patients with ischemic stroke.

Background: Recently, our team developed an end-to-end artificial intelligence (AI) framework that can successfully estimate left ventricular ejection fraction (LVEF) from echocardiogram videos. However, this framework requires specific views (A2C and PLAX), and cannot estimate LVEF for studies from which these views are missing. Goal: To develop a deep learning model to estimate LVEF from any type and number of echocardiogram video(s). Methods: We built a deep learning model for B-mode echocardiogram videos of the left ventricle (A2C, A3C, A4C, or PLAX views), using retrospective transthoracic echocardiography (TTE) data. The model was evaluated with two test datasets: one with random valid (i.e., showing the left ventricle) views, and one with preselected views (both having 433 patients). Later, the model was tested with a prospective patient cohort that had both TTE and point-of-care ultrasound (POCUS) data collected simultaneously (393 patients). Results: On the retrospective TTE test set, we observed excellent performance (five random valid views per study: RMSE of 6.87% and Pearson correlation (PC) of 0.82 on LVEF estimation, and an AUC of 0.96 (95% CI: of 0.94-0.98) on low LVEF classification (<=40%); preselected views: RMSE of 6.75%, PC of 0.83, and an AUC of 0.96 (95% CI: 0.94-0.98)). For both TTE and POCUS data from the prospective cohort, we also observed strong performance (TTE: RMSE of 5.62%, PC of 0.87, and an AUC of 0.97; POCUS: RMSE of 6.12%, PC of 0.84, and an AUC of 0.96). For the prospective test sets as well as for the retrospective TTE test set with preselected views, results reflect the estimated LVEF average of A2C, A3C, A4C, and PLAX videos from each patient. Conclusions: For both TTE and POCUS data, the current model showed strong performance which was comparable to that of our previously established view-specific model. Alongside this comparable performance, the current model surpasses our previous model in terms of flexibility, as it isn't dependent on view classification and allows a broader selection in terms of the number and types of videos per study or patient. This innovative approach could prove particularly beneficial in point-of-care situations in which it may not be feasible to obtain high-quality, view-specific videos.