Publications

2025

Liang, Yixiu, Arunashis Sau, Boroumand Zeidaabadi, Joseph Barker, Konstantinos Patlatzoglou, Libor Pastika, Ewa Sieliwonczyk, et al. (2025) 2025. “Artificial Intelligence-Enhanced Electrocardiography to Predict Regurgitant Valvular Heart Diseases: An International Study.”. European Heart Journal. https://doi.org/10.1093/eurheartj/ehaf448.

BACKGROUND AND AIMS: Valvular heart disease (VHD) is a significant source of morbidity and mortality, though early intervention can improve outcomes. This study aims to develop artificial intelligence-enhanced electrocardiography (AI-ECG) models to diagnose and predict future moderate or severe regurgitant VHDs (rVHDs), including mitral regurgitation (MR), tricuspid regurgitation (TR), and aortic regurgitation (AR).

METHODS: The AI-ECG models were developed in a data set of 988 618 ECG and transthoracic echocardiogram pairs from 400 882 patients from Zhongshan Hospital, Shanghai, China. The AI-ECG models used a residual convolutional neural network with a discrete-time survival loss function. External evaluation was performed in outpatients from a secondary care data set from Beth Israel Deaconess Medical Center, Boston, USA, consisting of 34 214 patients with linked echocardiography.

RESULTS: In the internal test set, the AI-ECG models accurately predicted future significant MR [C-index 0.774, 95% confidence interval (CI) 0.753-0.792], AR (0.691, 95% CI 0.657-0.720), and TR (0.793, 95% CI 0.777-0.808). In age- and sex-adjusted Cox models, the highest risk quartile had a hazard ratio (HR) of 7.6 (95% CI 5.8-9.9, P < .0001) for risk of future significant MR, compared with the lowest risk quartile. For future AR and TR, the equivalent HRs were 3.8 (95% CI 2.7-5.5) and 9.9 (95% CI 7.5-13.0), respectively. These findings were confirmed in the transnational external test set. Imaging association analyses demonstrated AI-ECG predictions were associated with subclinical chamber remodelling.

CONCLUSIONS: This study developed AI-ECG models to diagnose and predict rVHDs and validated the models in a transnational and ethnically distinct cohort. The AI-ECG models could be utilized to guide surveillance echocardiography in patients at risk of future rVHDs, to facilitate early detection and intervention.

Gurnani, Mehak, Konstantinos Patlatzoglou, Joseph Barker, Derek Bivona, Libor Pastika, Ewa Sieliwonczyk, Boroumand Zeidaabadi, et al. (2025) 2025. “Revisiting Abnormalities of Ventricular Depolarization: Redefining Phenotypes and Associated Outcomes Using Tree-Based Dimensionality Reduction.”. Journal of the American Heart Association, e040814. https://doi.org/10.1161/JAHA.124.040814.

BACKGROUND: Abnormal ventricular depolarization, evident as a broad QRS complex on an ECG, is traditionally categorized into left bundle-branch block (LBBB) and right bundle-branch block or nonspecific intraventricular conduction delay. This categorization, although physiologically accurate, may fail to capture the nuances of diseases subtypes.

METHODS: We used unsupervised machine learning to identify and characterize novel broad QRS phenogroups. First, we trained a variational autoencoder on 1.1 million ECGs and discovered 51 latent features that showed high disentanglement and ECG reconstruction accuracy. We then extracted these features from 42 538 ECGs with QRS durations >120 milliseconds and employed a reversed graph embedding method to model population heterogeneity as a tree structure with different branches representing phenogroups.

RESULTS: Six phenogroups were identified, including phenogroups of right bundle-branch block and LBBB with varying risk of cardiovascular disease and mortality. The higher risk right bundle-branch block phenogroup exhibited increased risk of cardiovascular death (adjusted hazard ratio [aHR], 1.46 [1.30-1.63], P<0.0001) and all-cause mortality (aHR, 1.24 [1.16-1.33], P<0.0001) compared with the baseline phenogroup. Within LBBB ECGs, tree position predicted future cardiovascular disease risk differentially. Additionally, for subjects with LBBB undergoing cardiac resynchronization therapy, tree position predicted cardiac resynchronization therapy response independent of covariates, including QRS duration (adjusted odds ratio [aOR], 0.47 [0.25-0.86], P<0.05).

CONCLUSIONS: Our findings challenge the current paradigm, highlighting the potential for these phenogroups to enhance cardiac resynchronization therapy patient selection for subjects with LBBB and guide investigation and follow-up strategies for subjects with higher risk right bundle-branch block.

Macierzanka, Krzysztof, Arunashis Sau, Konstantinos Patlatzoglou, Libor Pastika, Ewa Sieliwonczyk, Mehak Gurnani, Nicholas S Peters, Jonathan W Waks, Daniel B Kramer, and Fu Siong Ng. (2025) 2025. “Siamese Neural Network-Enhanced Electrocardiography Can Re-Identify Anonymized Healthcare Data.”. European Heart Journal. Digital Health 6 (3): 417-26. https://doi.org/10.1093/ehjdh/ztaf011.

AIMS: Many research databases with anonymized patient data contain electrocardiograms (ECGs) from which traditional identifiers have been removed. We evaluated the ability of artificial intelligence (AI) methods to determine the similarity between ECGs and assessed whether they have the potential to be misused to re-identify individuals from anonymized datasets.

METHODS AND RESULTS: We utilized a convolutional Siamese neural network (SNN) architecture, which derives a Euclidean distance similarity metric between two input ECGs. A secondary care dataset of 864 283 ECGs (72 455 subjects) was used. Siamese neural network-electrocardiogram (SNN-ECG) achieves an accuracy of 91.68% when classifying between 2 689 124 same-subject pairs and 2 689 124 different-subject pairs. This performance increases to 93.61% and 95.97% in outpatient and normal ECG subsets. In a simulated 'motivated intruder' test, SNN-ECG can identify individuals from large datasets. In datasets of 100, 1000, 10 000, and 20 000 ECGs, where only one ECG is also from the reference individual, it achieves success rates of 79.2%, 62.6%, 45.0%, and 40.0%, respectively. If this was random, the success would be 1%, 0.1%, 0.01%, and 0.005%, respectively. Additional basic information, like subject sex or age-range, enhances performance further. We also found that, on the subject level, ECG pair similarity is clinically relevant; greater ECG dissimilarity associates with all-cause mortality [hazard ratio, 1.22 (1.21-1.23), P < 0.0001] and is additive to an AI-ECG model trained for mortality prediction.

CONCLUSION: Anonymized ECGs retain information that may facilitate subject re-identification, raising privacy and data protection concerns. However, SNN-ECG models also have positive uses and can enhance risk prediction of cardiovascular disease.

Ferro, Enrico G, Matthew R Reynolds, Jiaman Xu, Yang Song, David J Cohen, Rishi K Wadhera, Andre D’Avila, Peter J Zimetbaum, Robert W Yeh, and Daniel B Kramer. (2025) 2025. “Contemporary Administrative Codes to Identify Pulmonary Vein Isolation Procedures for Atrial Fibrillation.”. Journal of the American Heart Association 14 (2): e037003. https://doi.org/10.1161/JAHA.124.037003.

BACKGROUND: Use of pulmonary vein isolation (PVI) to treat atrial fibrillation continues to increase. Despite great interest in leveraging administrative data for real-world analyses, contemporary procedural codes for identifying PVI have not been evaluated.

METHODS AND RESULTS: In this observational retrospective cohort study, inpatient PVIs were identified among US Medicare fee-for-service beneficiaries using Current Procedural Terminology (CPT) code 93656 in Carrier Line Files. Each patient was matched with their claims from Medicare Provider Analysis and Review to compare CPT with International Classification of Diseases, Tenth Revision, Procedure Coding System (ICD-10-PCS) claims submitted by health care facilities to bill for PVIs. We performed the reverse for commonly matched ICD-10-PCS codes, to identify corresponding CPT-billed procedures. Finally, we reviewed institutional cases for additional comparison of CPT and ICD-10-PCS assignation for PVI. We identified 25 617 inpatient PVIs from January 2017 to December 2021, of which 18 165 (71%) were linked to Medicare Provider Analysis and Review. Of these, 16 672 (92%) were billed as ICD-10-PCS 02583ZZ: "Destruction of Conduction Mechanism, Percutaneous Approach." The reverse process yielded heterogeneous results: among 75 003 procedures billed as ICD-10-PCS 02583ZZ, only 15 691 (21%) matched with CPT 93656 (PVI), as several other unrelated procedures were billed under this ICD-10-PCS code. Institutional case review confirmed the greater specificity of CPT codes.

CONCLUSIONS: The ICD-10-PCS code associated with CPT-billed PVI procedures actually referred to ablation of the atrioventricular junction. Yet this ICD-10-PCS code also matched with a wide range of other procedures distinct from PVI. We conclude that ICD-10-PCS codes alone are not sensitive nor specific for identifying PVI in claims and cannot be reliably used in isolation for health services research on this important procedure.

2024

NPJ Digital Medicine Logo
Pastika, Libor, Arunashis Sau, Konstantinos Patlatzoglou, Ewa Sieliwonczyk, Antônio H Ribeiro, Kathryn A McGurk, Sadia Khan, et al. (2024) 2024. “Artificial Intelligence-Enhanced Electrocardiography Derived Body Mass Index As a Predictor of Future Cardiometabolic Disease.”. NPJ Digital Medicine 7 (1): 167. https://doi.org/10.1038/s41746-024-01170-0.

The electrocardiogram (ECG) can capture obesity-related cardiac changes. Artificial intelligence-enhanced ECG (AI-ECG) can identify subclinical disease. We trained an AI-ECG model to predict body mass index (BMI) from the ECG alone. Developed from 512,950 12-lead ECGs from the Beth Israel Deaconess Medical Center (BIDMC), a secondary care cohort, and validated on UK Biobank (UKB) (n = 42,386), the model achieved a Pearson correlation coefficient (r) of 0.65 and 0.62, and an R2 of 0.43 and 0.39 in the BIDMC cohort and UK Biobank, respectively for AI-ECG BMI vs. measured BMI. We found delta-BMI, the difference between measured BMI and AI-ECG-predicted BMI (AI-ECG-BMI), to be a biomarker of cardiometabolic health. The top tertile of delta-BMI showed increased risk of future cardiometabolic disease (BIDMC: HR 1.15, p < 0.001; UKB: HR 1.58, p < 0.001) and diabetes mellitus (BIDMC: HR 1.25, p < 0.001; UKB: HR 2.28, p < 0.001) after adjusting for covariates including measured BMI. Significant enhancements in model fit, reclassification and improvements in discriminatory power were observed with the inclusion of delta-BMI in both cohorts. Phenotypic profiling highlighted associations between delta-BMI and cardiometabolic diseases, anthropometric measures of truncal obesity, and pericardial fat mass. Metabolic and proteomic profiling associates delta-BMI positively with valine, lipids in small HDL, syntaxin-3, and carnosine dipeptidase 1, and inversely with glutamine, glycine, colipase, and adiponectin. A genome-wide association study revealed associations with regulators of cardiovascular/metabolic traits, including SCN10A, SCN5A, EXOG and RXRG. In summary, our AI-ECG-BMI model accurately predicts BMI and introduces delta-BMI as a non-invasive biomarker for cardiometabolic risk stratification.

Heart Rhythm Logo
Isaza, Nicolas, Hans F Stabenau, Daniel B Kramer, Arunashis Sau, Patricia Tung, Timothy R Maher, Andrew H Locke, et al. (2024) 2024. “The Spatial Ventricular Gradient Is Associated With Inducibility of Ventricular Arrhythmias During Electrophysiology Study.”. Heart Rhythm. https://doi.org/10.1016/j.hrthm.2024.05.005.

BACKGROUND: Myocardial electrical heterogeneity is critical for normal cardiac electromechanical function, but abnormal or excessive electrical heterogeneity is proarrhythmic. The spatial ventricular gradient (SVG), a vectorcardiographic measure of electrical heterogeneity, has been associated with arrhythmic events during long-term follow-up, but its relationship with short-term inducibility of ventricular arrhythmias (VAs) is unclear.

OBJECTIVE: This study was designed to determine associations between SVG and inducible VAs during electrophysiology study.

METHODS: A retrospective study was conducted of adults without prior sustained VA, cardiac arrest, or implantable cardioverter-defibrillator who underwent ventricular stimulation for evaluation of syncope and nonsustained ventricular tachycardia or for risk stratification before primary prevention implantable cardioverter-defibrillator implantation. The 12-lead electrocardiograms were converted into vectorcardiograms, and SVG magnitude (SVGmag) and direction (azimuth and elevation) were calculated. Odds of inducible VA were regressed by logistic models.

RESULTS: Of 143 patients (median age, 69 years; 80% male; median left ventricular ejection fraction [LVEF], 47%; 52% myocardial infarction), 34 (23.8%) had inducible VAs. Inducible patients had lower median LVEF (38% vs 50%; P < .0001), smaller SVGmag (29.5 vs 39.4 mV·ms; P = .0099), and smaller cosine SVG azimuth (cosSVGaz; 0.64 vs 0.89; P = .0007). When LVEF, SVGmag, and cosSVGaz were dichotomized at their medians, there was a 39-fold increase in adjusted odds (P = .002) between patients with all low LVEF, SVGmag, and cosSVGaz (65% inducible) compared with patients with all high LVEF, SVGmag, and cosSVGaz (4% [n = 1] inducible). After multivariable adjustment, SVGmag, cosSVGaz, and sex but not LVEF or other characteristics remained associated with inducible VAs.

CONCLUSION: Assessment of electrical heterogeneity by SVG, which reflects abnormal electrophysiologic substrate, adds to LVEF and identifies patients at high and low risk of inducible VA at electrophysiology study.

Heart Rhythm Logo
Serpa, Frans, André Rivera, Julia M Fernandes, Marcelo Antonio Pinheiro Braga, Beatriz Araújo, Iuri Ferreira Felix, Enrico Ferro, Peter J Zimetbaum, Andre D’Avila, and Daniel B Kramer. (2024) 2024. “Intracardiac Vs Transesophageal Echocardiography for Left Atrial Appendage Occlusion: An Updated Systematic Review and Meta-Analysis.”. Heart Rhythm. https://doi.org/10.1016/j.hrthm.2024.08.027.

BACKGROUND: Multiple studies continue to evaluate the use of intracardiac echocardiography (ICE) and transesophageal echocardiography (TEE) for guiding left atrial appendage occlusion (LAAO).

OBJECTIVE: To conduct an updated meta-analysis comparing the effectiveness and safety outcomes of both imaging modalities.

METHODS: PubMed, Cochrane, and Embase were searched for studies comparing ICE vs TEE to guide LAAO. Odds ratios (OR) with 95% confidence intervals (CI) were pooled using a random-effects model. The primary effectiveness endpoint was procedural success. The primary safety endpoint included the overall complications rate. Additional safety outcomes were assessed as secondary endpoints. Subgroup analysis of primary endpoints was conducted according to device type (Amulet, LAmbre, Watchman, Watchman FLX) and study region (American, Asia, Europe). We used R version 4.3.1 for all statistical analyses.

RESULTS: Our meta-analysis included 19 observational studies encompassing 42,474 patients, of whom 4,415 (10.4%) underwent ICE-guided LAAO. Compared with TEE, ICE was associated with a marginally higher procedural success (OR 1.33; 95% CI: 1.01-1.76; p=0.04; I2=0%). There was no significant difference in the overall complications rate (OR 1.02; 95% CI: 0.77-1.36; p=0.89; I2=5%). However, ICE showed higher rates of pericardial effusion (OR 2.11; 95% CI: 1.47-3.03; p<0.001; I2=0%) and residual iatrogenic atrial septal defect (iASD) (OR 1.52; 95% CI: 1.15-2.03; p<0.004; I2=0%). Subgroup analysis revealed variations in procedural success within the ICE group across study regions (p=0.02).

CONCLUSION: In this updated meta-analysis, the increasing adoption of ICE-guided LAAO demonstrated higher procedural success rates compared to TEE, although with limited statistical significance. Overall complication rates were similar; however, ICE showed higher rates of pericardial effusion and residual iASD.