Mitral Valve Regurgitation is a common valvular condition representing a hidden burden of disease within the population. Valvular disease increases strain upon the heart muscles which can lead to heart failure. The purpose of the study was to develop and validate a CNN-based deep learning algorithm to develop a diagnostic tool to assist in clinical decision making for early detection of moderate-to-severe Mitral Valve regurgitation. For development Mount Sinai used 577,190 ECGs paired to transthoracic echocardiograms from 116,612 patients as training inputs to the model for detection of Mitral Valve Regurgitation. The Area Under Receiver Operating Characteristic curve (AUROC) for the external test set was 0.81 (95% CI:0.80–0.82) using 6860 subjects. Performance was evaluated for the MR classifier with respect to performance in concomitant Aortic Stenosis + MR, and compared this to performance for isolated valvular lesions. It was found that for MR classification, AUROC was higher for patients with concomitant AS + MR using external validation (AS + MR: 0.88; Isolated MR: 0.81).

Multi-center retrospective cohort study applying deep learning to electrocardiograms to identify left heart valvular dysfunction | Communications Medicine (nature.com)

Aortic stenosis is a common valvular condition that occurs when the aortic valve narrows and blood cannot flow normally. The smaller opening causes your heart's left ventricle to pump harder to push blood through the narrowed aortic valve. Mount Sinai developed a CNN-based deep learning algorithm developed to detect the presence of moderate-to-severe Aortic Stenosis. 585,098 Echo-ECG pairs from 120,564 patients were used for training and 8,111 cohorts for external validation. An AUROC for the external test set achieved 0.86 (95% CI:0.85–0.87). Performance was evaluated for the AS classifier with respect to performance in concomitant AS + Mitral Regurgitation, compared to performance for isolated valvular lesions. We found that for AS classification, AUROC was higher for patients with concomitant AS + MR using external validation (AS + MR: 0.89; Isolated AS: 0.85).

Multi-center retrospective cohort study applying deep learning to electrocardiograms to identify left heart valvular dysfunction | Communications Medicine (nature.com)

A pulmonary embolism (PE) is life-threatening condition caused by a sudden blockage in your pulmonary arteries, the blood vessels that send blood to your lungs. It may occur when a blood clot in the deep veins of a patient’s leg breaks off and travels to their lungs. A PE is a medical emergency that needs prompt diagnosis and treatment. Current clinical scoring systems for pulmonary embolism (PE) screening have low specificity. Given the current lack of specificity in scoring systems it requires clinicians to use thoracic imaging modalities, most commonly computed tomography pulmonary angiogram (CTPA), to confirm diagnosis. A method for immediate, bedside diagnosis or exclusion of PE is not currently available. A CNN based deep learning algorithm integrating ECG waveforms was developed by Mount Sinai using 23793 CTPAs (10.0% PE-positive), 320746 ECGs and clinical data (demographics, comorbidities, vital signs, and labs) from 21183 patients. A deep learning model AI-ECG model that incorporated limited clinical data performed at an AUROC of 0.81 on the 10% or n=2118 test set. The AI-ECG PE model combined with clinical data may be used to recommend computed tomography pulmonary angiogram or exclude pulmonary embolism in patients with moderate to high likelihood of pulmonary embolism after clinical stratification or those at low suspicion with an abnormal D-dimer. Synergistic deep learning of ECG waveforms with traditional clinical variables can increase the specificity of PE detection in patients with moderate or greater suspicion for PE.

Development of a machine learning model using electrocardiogram signals to improve acute pulmonary embolism screening | European Heart Journal - Digital Health | Oxford Academic (oup.com)

Right ventricular (RV) size and functional metrics have important prognostic implications for many diseases, including cardiomyopathy, pulmonary hypertension, and structural heart disease. Right Ventricle Systolic Dysfunction and Right Ventricle Dilation are not as easily assessed as the left ventricle. Mount Sinai trained a CNN-based AI-ECG model to predict RV dilation (RVEDV >120 mL/m2), RV dysfunction (RVEF ≤40%), and numerical RVEDV and RVEF from an ECG paired with MRI volumetric measurements in the UK Biobank (UKBB; n=42 938). We then fine-tuned the model in a multicenter health system Mount Sinai Hospital n=3019 followed up by prospective validation over 4 months in the same center (MSHvalidation; n=115). The prevalence of RV dysfunction for UKBB/MSHoriginal/MSHvalidation cohorts was 1.0%/18.0%/15.7%, respectively. RV dysfunction model achieved an AUROC performance in the UKBB/ MSHoriginal/MSHvalidation cohorts at 0.86/0.81/0.77, respectively. Deep learning–ECG analysis can identify significant cardiac magnetic resonance imaging RV dysfunction and dilation with good performance. RVEF model AUROC performance was consistent in all tested subgroups, including those with and without LV dysfunction.

Quantitative Prediction of Right Ventricular Size and Function From the ECG | Journal of the American Heart Association (ahajournals.org)

Heart failure affects millions globally. The right ventricle plays a crucial role in maintaining overall heart function, and any dysfunction can impact circulation and oxygenation. Right Ventricle Systolic Dysfunction and Right Ventricle Dilation are not as easily assessed as the left ventricle, therefore integrating, and combining multiple parameters may be a more reliable way to determine normal or abnormal function. Development of a CNN-based AI-ECG algorithm was developed by Mount Sinai to analyze ECGs and predict a composite outcome of either RV systolic dysfunction or RV dilation. To develop the algorithm a data set with 761,510 ECGs paired with Echo for 148,227 patients were used. The external test set included 13432 patients for Right Ventricle Dysfunction. The model developed achieved an AUROC of 0.84 (95% CI: 0.84-0.84) during external validation testing. AI-ECG algorithms developed to detect RVD may enhance the usefulness of ECG screening and management for right sided heart failure.

Using Deep-Learning Algorithms to Simultaneously Identify Right and Left Ventricular Dysfunction From the Electrocardiogram (sciencedirectassets.com)

Hypertrophic cardiomyopathy (HCM) is a cardiovascular disease characterized by an abnormally thick heart muscle. Subjects with HCM may not exhibit symptoms but may experience progressive symptoms that can lead to serious heart issues. The HCM algorithm was developed by Mount Sinai using innovative vision-based transformer data science methods.  The HCM AI-ECG analysis for diagnosis of hypertrophic cardiomyopathy can aid in early detection of HCM. A total of 78,831 ECGs from four hospitals were used for training and 3,859 ECGs from a holdout set of patients were used for external validation (AUROC: 0.80).

A foundational vision transformer improves diagnostic performance for electrocardiograms | npj Digital Medicine (nature.com)

An ST-elevation Myocardial Infarction (STEMI) is a type of severe heart attack caused by an occlusion of one or more of the coronary arteries that supply the heart with blood. A vision transformer-based AI-ECG algorithm was developed to detect STEMI using ECG waveforms that are typically used for detection and diagnosis of STEMI. The STEMI algorithm was developed by Mount Sinai using 21,799 total ECGs for fine-tuning and 4352 ECGs to test the model. The prevalence of STEMI was around 5.4% in the testing set, performance was AUROC: 0.95.

A foundational vision transformer improves diagnostic performance for electrocardiograms | npj Digital Medicine (nature.com)

Premature ventricular complexes (PVCs) are prevalent and, although often benign, may lead to PVC-induced cardiomyopathy.  Frequent PVCs can lead to changes in the heart’s structure and function causing an irregular heartbeat and/or weakening of the heart muscle. However, PVC induced cardiomyopathy is potentially reversible. A novel CNN-based deep learning ECG algorithm was developed by Mount Sinai using N = 14,241 subjects with 27,370 ECGs in the training set to predict cardiomyopathy in patients with premature ventricular complexes (PVCs). The primary outcome was first diagnosis of LVEF <=40% within 6 months. The model predicted reduction for LVEF <=40% in the external test set (n= 688 and 1,375 ECGs, prevalence 39%) with area under the receiver operating curve of 0.79 (95% CI: 0.77-0.81). This CNN-based algorithm using 12-lead ECG information alone can predict new-onset cardiomyopathy in patients with PVCs, independent of PVC burden, contributing to better patient management.

A Novel ECG-Based Deep Learning Algorithm to Predict Cardiomyopathy in Patients With Premature Ventricular Complexes — Icahn School of Medicine at Mount Sinai (mssm.edu)