Science

Abstract:

Using data from 14,613 participants in the Atherosclerosis Risk in Communities (ARIC) study, this research developed a deep residual convolutional neural network (CNN) to predict heart failure from standard 12-lead ECGs. The ECG-AI model achieved an AUC of 0.756 using solely ECG data, comparable to established ARIC (AUC 0.802) and Framingham Heart Study risk calculators (AUC 0.780). When combined with clinical variables in a light gradient boosting machine model, the highest AUC of 0.818 was achieved, with the ECG-AI model output serving as the most important predictor of heart failure within 10 years.DOI: 10.1093/ehjdh/ztab080

Abstract:

This study developed a single-lead ECG-based deep learning model for fatal coronary heart disease (FCHD) risk prediction using 167,662 ECGs from 50,132 patients. The ECG-AI model achieved AUC = 0.76, while the combined Cox proportional hazards model reached AUC = 0.87 on external validation data. Testing on 243 paired clinical and Apple Watch ECGs showed strong concordance (R = 0.74), with 99% agreement in low/high-risk FCHD classification, demonstrating the feasibility of remote monitoring for fatal coronary heart disease using wearable devices.DOI: 10.1016/S2666-6936(24)00030-6

Abstract:

This study applied AI methods to baseline ECGs from 1,217 adult survivors of childhood cancer to predict late-onset cardiomyopathy. The ECG-based AI model achieved an impressive AUC of 0.87, outperforming clinical features alone (AUC 0.69), while the combined ECG and clinical model reached the highest AUC of 0.89 with 78% sensitivity and 81% specificity. AI using ECG data may assist in identifying childhood cancer survivors at increased risk for developing future cardiomyopathy.DOI: 10.1200/CCI.20.00176

Abstract:

ECG-AIR is an innovative iOS platform enabling remote execution of AI models on smartwatch ECG data for cardiovascular disease detection and prediction. The platform consists of two sub-applications: Gather (to collect smartwatch ECGs from users) and Run (to execute AI models either embedded within the iPhone or on cloud infrastructure). Testing with Apple Watch demonstrated remarkably fast execution times—less than 500ms for on-device processing and less than 5 seconds for cloud-based analysis. The platform was validated using a previously developed deep learning model (ECG-AI) for heart failure risk prediction, with the TensorFlow-Light smartphone version providing identical predictions to the original TensorFlow model. To the best of our knowledge, ECG-AIR is the first remote AI platform enabling retrieval and analysis of digital smartwatch ECG for CVD detection and prediction, potentially facilitating early therapeutic interventions through preventive cardiology and telemedicine efforts.DOI: 10.1016/j.cvdhj.2022.09.002

Abstract:

This study developed and validated multi-task deep learning models that simultaneously estimate natriuretic peptide values (BNP and NT-proBNP) and classify clinically relevant strata directly from raw 12-lead and Lead I ECG waveforms. Using 102,311 paired ECG-BNP samples from Wake Forest Baptist Health and external validation with 88,179 ECG-NT-proBNP pairs from UTHSC, the 12-lead model achieved AUCs of 0.88-0.89 and Spearman correlations of 0.75 for BNP classification, with similar performance in the external NT-proBNP cohort. This approach offers a rapid, non-invasive tool for heart failure biomarker assessment that may improve clinical workflows, enable ambulatory monitoring, and enhance timely clinical decision-making.DOI: 10.1161/circ.152.suppl_3.4369685

Abstract:

This study externally validated the ECG-AI convolutional neural network model using 6,736 participants from the Multi-Ethnic Study of Atherosclerosis (MESA) who were free of heart failure at enrollment. The ECG-AI model demonstrated comparable predictive ability to the original ARIC-developed model and performed similarly to established heart failure prediction scores. This external validation confirms the model's robustness and potential for identifying patients at elevated risk for heart failure across diverse populations and large databases.DOI: 10.1016/S0735-1097(22)01319-5

Abstract:

This retrospective study developed ECG-AI models to predict fatal coronary heart disease (FCHD) using data from 50,132 patients. The single-lead ECG-AI model combined with demographics achieved AUC = 0.87 on external validation, with exceptional 2-year FCHD prediction accuracy of AUC = 0.91. The study demonstrated strong correlation (R = 0.74) between 12-lead and single-lead ECG predictions, with 86% agreement in risk stratification. These findings support the feasibility of remote FCHD monitoring using consumer wearable devices with single-lead ECG capabilities.DOI: 10.3390/jcdd11120395

Abstract:

This study developed and validated ECG-AI models for peripartum cardiomyopathy (PPCM) detection using 12-lead and single-lead ECGs from 50,373 patients. The 12-lead ECG-AI model achieved AUC = 0.88 for PPCM detection, while the single-lead model reached AUC = 0.86. External validation on 2,305 patients demonstrated AUCs of 0.85 and 0.83 respectively. The models showed strong correlation (R = 0.74) and high concordance in risk stratification. This breakthrough enables early PPCM detection during pregnancy and postpartum periods, potentially preventing maternal mortality through timely intervention.DOI: 10.1016/j.ajogmf.2024.101337

Abstract:

This study developed AI models to detect and predict preeclampsia from 12-lead ECGs using data from University of Tennessee Health Science Center (904 ECGs from 759 females, 78.8% African American) and Atrium Health Wake Forest Baptist (817 ECGs from 141 females, 45.4% African American). The ECG-AI model achieved AUC of 0.85 (0.77-0.93) on UTHSC holdout data and 0.81 (0.77-0.84) on AHWFB external validation. Predictive accuracy for preeclampsia within 30, 60, and 90 days before diagnosis yielded AUCs of 0.92, 0.89, and 0.90 respectively. For early-onset preeclampsia (<34 weeks), the model achieved exceptional AUC of 0.98 (0.89-1.00), demonstrating high accuracy for identifying this life-threatening condition affecting over 76,000 women yearly.DOI: 10.3389/fcvm.2024.1360238