Artificial intelligence: revolutionizing cardiology with large language models.

Natural language processing techniques are having an increasing impact on clinical care from patient, clinician, administrator, and research perspective. Among others are automated generation of clinical notes and discharge letters, medical term coding for billing, medical chatbots both for patients and clinicians, data enrichment in the identification of disease symptoms or diagnosis, cohort selection for clinical trial, and auditing purposes. In the review, an overview of the history in natural language processing techniques developed with brief technical background is presented. Subsequently, the review will discuss implementation strategies of natural language processing tools, thereby specifically focusing on large language models, and conclude with future opportunities in the application of such techniques in the field of cardiology.

Electrocardiographic temporo-spatial assessment of depolarization and repolarization changes after epicardial arrhythmogenic substrate ablation in Brugada syndrome.

Aims: In Brugada syndrome (BrS), with spontaneous or ajmaline-induced coved ST elevation, epicardial electro-anatomic potential duration maps (epi-PDMs) were detected on a right ventricle (RV) outflow tract (RVOT), an arrhythmogenic substrate area (AS area), abolished by epicardial-radiofrequency ablation (EPI-AS-RFA). Novel CineECG, projecting 12-lead electrocardiogram (ECG) waveforms on a 3D heart model, previously localized depolarization forces in RV/RVOT in BrS patients. We evaluate 12-lead ECG and CineECG depolarization/repolarization changes in spontaneous type-1 BrS patients before/after EPI-AS-RFA, compared with normal controls. Methods and results: In 30 high-risk BrS patients (93% males, age 37 + 9 years), 12-lead ECGs and epi-PDMs were obtained at baseline, early after EPI-AS-RFA, and late follow-up (FU) (2.7-16.1 months). CineECG estimates temporo-spatial localization during depolarization (Early-QRS and Terminal-QRS) and repolarization (ST-Tpeak, Tpeak-Tend). Differences within BrS patients (baseline vs. early after EPI-AS-RFA vs. late FU) were analysed by Wilcoxon signed-rank test, while differences between BrS patients and 60 age-sex-matched normal controls were analysed by the Mann-Whitney test. In BrS patients, baseline QRS and QTc durations were longer and normalized after EPI-AS-ATC (151 ± 15 vs. 102 ± 13 ms, P < 0.001; 454 ± 40 vs. 421 ± 27 ms, P < 0.000). Baseline QRS amplitude was lower and increased at late FU (0.63 ± 0.26 vs. 0.84 ± 13 ms, P < 0.000), while Terminal-QRS amplitude decreased (0.24 ± 0.07 vs. 0.08 ± 0.03 ms, P < 0.000). At baseline, CineECG depolarization/repolarization wavefront prevalently localized in RV/RVOT (Terminal-QRS, 57%; ST-Tpeak, 100%; and Tpeak-Tend, 61%), congruent with the AS area on epi-PDM. Early after EPI-AS-RFA, RV/RVOT localization during depolarization disappeared, as Terminal-QRS prevalently localized in the left ventricle (LV, 76%), while repolarization still localized on RV/RVOT [ST-Tpeak (44%) and Tpeak-Tend (98%)]. At late FU, depolarization/repolarization forces prevalently localized in the LV (Terminal-QRS, 94%; ST-Tpeak, 63%; Tpeak-Tend, 86%), like normal controls. Conclusion: CineECG and 12-lead ECG showed a complex temporo-spatial perturbation of both depolarization and repolarization in BrS patients, prevalently localized in RV/RVOT, progressively normalizing after epicardial ablation.

Body surface potential mapping detects early disease onset in plakophilin-2-pathogenic variant carriers.

AIMS: Arrhythmogenic right ventricular cardiomyopathy (ARVC) is a progressive inherited cardiac disease. Early detection of disease and risk stratification remain challenging due to heterogeneous phenotypic expression. The standard configuration of the 12 lead electrocardiogram (ECG) might be insensitive to identify subtle ECG abnormalities. We hypothesized that body surface potential mapping (BSPM) may be more sensitive to detect subtle ECG abnormalities. METHODS AND RESULTS: We obtained 67 electrode BSPM in plakophilin-2 (PKP2)-pathogenic variant carriers and control subjects. Subject-specific computed tomography/magnetic resonance imaging based models of the heart/torso and electrode positions were created. Cardiac activation and recovery patterns were visualized with QRS- and STT-isopotential map series on subject-specific geometries to relate QRS-/STT-patterns to cardiac anatomy and electrode positions. To detect early signs of functional/structural heart disease, we also obtained right ventricular (RV) echocardiographic deformation imaging. Body surface potential mapping was obtained in 25 controls and 42 PKP2-pathogenic variant carriers. We identified five distinct abnormal QRS-patterns and four distinct abnormal STT-patterns in the isopotential map series of 31/42 variant carriers. Of these 31 variant carriers, 17 showed no depolarization or repolarization abnormalities in the 12 lead ECG. Of the 19 pre-clinical variant carriers, 12 had normal RV-deformation patterns, while 7/12 showed abnormal QRS- and/or STT-patterns. CONCLUSION: Assessing depolarization and repolarization by BSPM may help in the quest for early detection of disease in variant carriers since abnormal QRS- and/or STT-patterns were found in variant carriers with a normal 12 lead ECG. Because electrical abnormalities were observed in subjects with normal RV-deformation patterns, we hypothesize that electrical abnormalities develop prior to functional/structural abnormalities in ARVC.

CineECG illustrating the ventricular activation sequence in progressive AV-junctional conduction block.

We present the use of CineECG in visualizing abnormal ventricular activation in a case of a complex conduction disorder. CineECG combines the standard 12­lead surface ECG with a 3D anatomical model of the heart. It projects the location and direction of the average ventricular activation and recovery on the heart model over time. In this case, CineECG was able to visualize the different type of fascicular conduction in this progressive conduction block. This novel imaging technique was able to provide additional insight in this complex case, and might be of use in other complex ECG patterns.

ECG-based techniques to enhance clinical practice in cardiac genetic disease management.

INTRODUCTION: Inherited cardiomyopathies are associated with a broad spectrum of potentially lethal phenotypes characterized by structural and electrical myocardial remodeling. Increased awareness and genetic cascade screening lead to more genotype-positive, yet phenotype-negative individuals to be evaluated and followed up. The predictive value of genetic testing is hampered by incomplete penetrance and high variability in disease onset, progression and severity. CLINICAL CHALLENGES: Dilated cardiomyopathy usually manifests with symptoms of heart failure and ventricular arrhythmias (VA) develop in advanced disease. In arrhythmogenic cardiomyopathy (ACM), electrical remodeling can precede structural and functional changes and life-threatening VA can be the first disease manifestation. Early signs and symptoms may be subtle and go unnoticed. Physicians are in great need of appropriate screening and risk-stratification strategies. Task Force Criteria (TFC) were established to standardize the clinical diagnosis of ACM but risk-stratification remains challenging. Accurate prediction of disease progression in variation carriers is currently beyond the capabilities of diagnostic tests. PROPOSED DIAGNOSTIC TECHNIQUES: We propose three ECG-based techniques; isopotential mapping, inverse ECG and CineECG, to enhance risk-stratification in ACM. With the use of isopotential mapping abnormal spatio-temporal activation and repolarization may be identified. Furthermore, by combining subject specific ≥12­lead ECG data with cardiothoracic imaging using inverse ECG techniques, the direct link between ECG and cardiac anatomy can be obtained. CONCLUSION: New ECG techniques may prove more sensitive to detect early de- and repolarization abnormalities in yet asymptomatic variation carriers. Early electrical signs of disease progression may be identified prior to symptoms. Furthermore, individualized risk-stratification may be enhanced.

CineECG: A novel method to image the average activation sequence in the heart from the 12-lead ECG.

The standard 12-lead electrocardiogram (ECG) is a diagnostic tool to asses cardiac electrical activity. The vectorcardiogram is a related tool that represents that activity as the direction of a vector. In this work we investigate CineECG, a new 12-lead ECG based analysis method designed to directly estimate the average cardiac anatomical location of activation over time. We describe CineECG calculation and a novel comparison parameter, the average isochrone position (AIP). In a model study, fourteen different activation sequences were simulated and corresponding 12-lead ECGs were computed. The CineECG was compared to AIP in terms of location and direction. In addition, 67-lead body surface potential maps from ten patients were used to study the sensitivity of CineECG to electrode mispositioning and anatomical model selection. Epicardial activation maps from four patients were used for further evaluation. The average distance between CineECG and AIP across the fourteen sequences was 23.7 ± 2.4 mm, with significantly better agreement in the terminal (27.3 ± 5.7 mm) versus the initial QRS segment (34.2 ± 6.1 mm). Up to four cm variation in electrode positioning produced an average distance of 6.5 ± 4.5 mm between CineECG trajectories, while substituting a generic heart/torso model for a patient-specific one produced an average difference of 6.1 ± 4.8 mm. Dominant epicardial activation map features were recovered. Qualitatively, CineECG captured significant features of activation sequences and was robust to electrode misplacement. CineECG provides a realistic representation of the average cardiac activation in normal and diseased hearts. In particular, the terminal segment of the CineECG might be useful to detect pathology.

Modeling the His-Purkinje Effect in Non-invasive Estimation of Endocardial and Epicardial Ventricular Activation.

Inverse electrocardiography (iECG) estimates epi- and endocardial electrical activity from body surface potentials maps (BSPM). In individuals at risk for cardiomyopathy, non-invasive estimation of normal ventricular activation may provide valuable information to aid risk stratification to prevent sudden cardiac death. However, multiple simultaneous activation wavefronts initiated by the His-Purkinje system, severely complicate iECG. To improve the estimation of normal ventricular activation, the iECG method should accurately mimic the effect of the His-Purkinje system, which is not taken into account in the previously published multi-focal iECG. Therefore, we introduce the novel multi-wave iECG method and report on its performance. Multi-wave iECG and multi-focal iECG were tested in four patients undergoing invasive electro-anatomical mapping during normal ventricular activation. In each subject, 67-electrode BSPM were recorded and used as input for both iECG methods. The iECG and invasive local activation timing (LAT) maps were compared. Median epicardial inter-map correlation coefficient (CC) between invasive LAT maps and estimated multi-wave iECG versus multi-focal iECG was 0.61 versus 0.31. Endocardial inter-map CC was 0.54 respectively 0.22. Modeling the His-Purkinje system resulted in a physiologically realistic and robust non-invasive estimation of normal ventricular activation, which might enable the early detection of cardiac disease during normal sinus rhythm.

Incorporating structural abnormalities in equivalent dipole layer based ECG simulations.

Introduction: Electrical activity of the myocardium is recorded with the 12-lead ECG. ECG simulations can improve our understanding of the relation between abnormal ventricular activation in diseased myocardium and body surface potentials (BSP). However, in equivalent dipole layer (EDL)-based ECG simulations, the presence of diseased myocardium breaks the equivalence of the dipole layer. To simulate diseased myocardium, patches with altered electrophysiological characteristics were incorporated within the model. The relation between diseased myocardium and corresponding BSP was investigated in a simulation study. Methods: Activation sequences in normal and diseased myocardium were simulated and corresponding 64-lead BSP were computed in four models with distinct patch locations. QRS-complexes were compared using correlation coefficient (CC). The effect of different types of patch activation was assessed. Of one patient, simulated electrograms were compared to electrograms recorded during invasive electro-anatomical mapping. Results: Hundred-fifty-three abnormal activation sequences were simulated. Median QRS-CC of delayed versus dyssynchronous were significantly different (1.00 vs. 0.97, p < 0.001). Depending on the location of the patch, BSP leads were affected differently. Within diseased regions, fragmentation, low bipolar voltages and late potentials were observed in both recorded and simulated electrograms. Discussion: A novel method to simulate cardiomyopathy in EDL-based ECG simulations was established and evaluated. The new patch-based approach created a realistic relation between ECG waveforms and underlying activation sequences. Findings in the simulated cases were in agreement with clinical observations. With this method, our understanding of disease progression in cardiomyopathies may be further improved and used in advanced inverse ECG procedures.

Effect of Segmentation Uncertainty on the ECGI Inverse Problem Solution and Source Localization.

Electrocardiographic Imaging (ECGI) is a promising tool to non-invasively map the electrical activity of the heart using body surface potentials (BSPs) and the patient specific anatomical data. One of the first steps of ECGI is the segmentation of the heart and torso geometries. In the clinical practice, the segmentation procedure is not fully-automated yet and is in consequence operator-dependent. We expect that the inter-operator variation in cardiac segmentation would influence the ECGI solution. This effect remains however non quantified. In the present work, we study the effect of segmentation variability on the ECGI estimation of the cardiac activity with 262 shape models generated from fifteen different segmentations. Therefore, we designed two test cases: with and without shape model uncertainty. Moreover, we used four cases for ectopic ventricular excitation and compared the ECGI results in terms of reconstructed activation times and excitation origins. The preliminary results indicate that a small variation of the activation maps can be observed with a model uncertainty but no significant effect on the source localization is observed.

The Effect of Segmentation Variability in Forward ECG Simulation.

Segmentation of patient-specific anatomical models is one of the first steps in Electrocardiographic imaging (ECGI). However, the effect of segmentation variability on ECGI remains unexplored. In this study, we assess the effect of heart segmentation variability on ECG simulation. We generated a statistical shape model from segmentations of the same patient and generated 262 cardiac geometries to run in an ECG forward computation of body surface potentials (BSPs) using an equivalent dipole layer cardiac source model and 5 ventricular stimulation protocols. Variability between simulated BSPs for all models and protocols was assessed using Pearson's correlation coefficient (CC). Compared to the BSPs of the mean cardiac shape model, the lowest variability (average CC = 0.98 ± 0.03) was found for apical pacing whereas the highest variability (average CC = 0.90 ± 0.23) was found for right ventricular free wall pacing. Furthermore, low amplitude BSPs show a larger variation in QRS morphology compared to high amplitude signals. The results indicate that the uncertainty in cardiac shape has a significant impact on ECGI.

Comparing Non-invasive Inverse Electrocardiography With Invasive Endocardial and Epicardial Electroanatomical Mapping During Sinus Rhythm.

This study presents a novel non-invasive equivalent dipole layer (EDL) based inverse electrocardiography (iECG) technique which estimates both endocardial and epicardial ventricular activation sequences. We aimed to quantitatively compare our iECG approach with invasive electro-anatomical mapping (EAM) during sinus rhythm with the objective of enabling functional substrate imaging and sudden cardiac death risk stratification in patients with cardiomyopathy. Thirteen patients (77% males, 48 ± 20 years old) referred for endocardial and epicardial EAM underwent 67-electrode body surface potential mapping and CT imaging. The EDL-based iECG approach was improved by mimicking the effects of the His-Purkinje system on ventricular activation. EAM local activation timing (LAT) maps were compared with iECG-LAT maps using absolute differences and Pearson's correlation coefficient, reported as mean ± standard deviation [95% confidence interval]. The correlation coefficient between iECG-LAT maps and EAM was 0.54 ± 0.19 [0.49-0.59] for epicardial activation, 0.50 ± 0.27 [0.41-0.58] for right ventricular endocardial activation and 0.44 ± 0.29 [0.32-0.56] for left ventricular endocardial activation. The absolute difference in timing between iECG maps and EAM was 17.4 ± 7.2 ms for epicardial maps, 19.5 ± 7.7 ms for right ventricular endocardial maps, 27.9 ± 8.7 ms for left ventricular endocardial maps. The absolute distance between right ventricular endocardial breakthrough sites was 30 ± 16 mm and 31 ± 17 mm for the left ventricle. The absolute distance for latest epicardial activation was median 12.8 [IQR: 2.9-29.3] mm. This first in-human quantitative comparison of iECG and invasive LAT-maps on both the endocardial and epicardial surface during sinus rhythm showed improved agreement, although with considerable absolute difference and moderate correlation coefficient. Non-invasive iECG requires further refinements to facilitate clinical implementation and risk stratification.

The relation of 12 lead ECG to the cardiac anatomy: The normal CineECG.

BACKGROUND: The interpretation of the 12­lead ECG is notoriously difficult and requires experts to distinguish normal from abnormal ECG waveforms. ECG waveforms depend on body build and electrode positions, both often different in males and females. To relate the ECG waveforms to cardiac anatomical structures is even more difficult. The novel CineECG algorithm enables a direct projection of the 12­lead ECG to the cardiac anatomy by computing the mean location of cardiac activity over time. The aim of this study is to investigate the cardiac locations of the CineECG derived from standard 12­lead ECGs of normal subjects. METHODS: In this study we used 6525 12­lead ECG tracings labelled as normal obtained from the certified Physionet PTB XL Diagnostic ECG Database to construct the CineECG. All 12 lead ECGs were analyzed, and then divided by age groups (18-29,30-39,40-49,50-59,60-69,70-100 years) and by gender (male/female). For each ECG, we computed the CineECG within a generic 3D heart/torso model. Based on these CineECG's, the average normal cardiac location and direction for QRS, STpeak, and TpeakTend segments were determined. RESULTS: The CineECG direction for the QRS segment showed large variation towards the left free wall, whereas the STT segments were homogeneously directed towards the septal/apical region. The differences in the CineECG location for the QRS, STpeak, and TpeakTend between the age and gender groups were relatively small (maximally 10 mm at end T-wave), although between the gender groups minor differences were found in the 4 chamber direction angles (QRS 4°, STpeak 5°, and TpeakTend 8°) and LAO (QRS 1°, STpeak 13°, and TpeakTend 30°). CONCLUSION: CineECG demonstrated to be a feasible and pragmatic solution for ECG waveform interpretation, relating the ECG directly to the cardiac anatomy. The variations in depolarization and repolarization CineECG were small within this group of normal healthy controls, both in cardiac location as well as in direction. CineECG may enable an easier discrimination between normal and abnormal QRS and T-wave morphologies, reducing the amount of expert training. Further studies are needed to prove whether novel CineECG can significantly contribute to the discrimination of normal versus abnormal ECG tracings.

Feasibility study of a 3D camera to reduce electrode repositioning errors during longitudinal ECG acquisition.

INTRODUCTION: Longitudinal monitoring of sometimes subtle waveform changes of the 12­lead electrocardiogram (ECG) is complicated by patient-specific and technical factors, such as the inaccuracy of electrode repositioning. This feasibility study uses a 3D camera to reduce electrode repositioning errors, reduce ECG waveform variability and enable detailed longitudinal ECG monitoring. METHODS: Per subject, three clinical ECGs were obtained during routine clinical follow-up. Additionally, two ECGs were recorded guided by two 3D cameras, which were used to capture the precordial electrode locations and direct electrode repositioning. ECG waveforms and parameters were quantitatively compared between 3D camera guided ECGs and clinical ECGs. Euclidian distances between original and repositioned precordial electrodes from 3D guided ECGs were measured. RESULTS: Twenty subjects (mean age 65.1 ± 8.2 years, 35% females) were included. The ECG waveform variation between routine ECGs was significantly higher compared to 3D guided ECGs, for both the QRS complex (correlation coefficient = 0.90 vs 0.98, p < 0.001) and the STT segment (correlation coefficient = 0.88 vs. 0.96, p < 0.001). QTc interval variation was reduced for 3D camera guided ECGs compared to routine clinical ECGs (5.6 ms vs. 9.6 ms, p = 0.030). The median distance between 3D guided repositioned electrodes was 10.0 [6.4-15.2] mm, and did differ between males and females (p = 0.076). CONCLUSIONS: 3D guided repositioning of precordial electrodes resulted in, a low repositioning error, higher agreement between waveforms of consecutive ECGs and a reduction of QTc variation. These findings suggest that longitudinal monitoring of disease progression using 12­lead ECG waveforms is feasible in clinical practice.

Novel CineECG enables anatomical 3D localization and classification of bundle branch blocks.

AIMS: Ventricular conduction disorders can induce arrhythmias and impair cardiac function. Bundle branch blocks (BBBs) are diagnosed by 12-lead electrocardiogram (ECG), but discrimination between BBBs and normal tracings can be challenging. CineECG computes the temporo-spatial trajectory of activation waveforms in a 3D heart model from 12-lead ECGs. Recently, in Brugada patients, CineECG has localized the terminal components of ventricular depolarization to right ventricle outflow tract (RVOT), coincident with arrhythmogenic substrate localization detected by epicardial electro-anatomical maps. This abnormality was not found in normal or right BBB (RBBB) patients. This study aimed at exploring whether CineECG can improve the discrimination between left BBB (LBBB)/RBBB, and incomplete RBBB (iRBBB). METHODS AND RESULTS: We utilized 500 12-lead ECGs from the online Physionet-XL-PTB-Diagnostic ECG Database with a certified ECG diagnosis. The mean temporo-spatial isochrone trajectory was calculated and projected into the anatomical 3D heart model. We established five CineECG classes: 'Normal', 'iRBBB', 'RBBB', 'LBBB', and 'Undetermined', to which each tracing was allocated. We determined the accuracy of CineECG classification with the gold standard diagnosis. A total of 391 ECGs were analysed (9 ECGs were excluded for noise) and 240/266 were correctly classified as 'normal', 14/17 as 'iRBBB', 55/55 as 'RBBB', 51/51 as 'LBBB', and 31 as 'undetermined'. The terminal mean temporal spatial isochrone contained most information about the BBB localization. CONCLUSION: CineECG provided the anatomical localization of different BBBs and accurately differentiated between normal, LBBB and RBBB, and iRBBB. CineECG may aid clinical diagnostic work-up, potentially contributing to the difficult discrimination between normal, iRBBB, and Brugada patients.

Uncertainty Quantification of the Effects of Segmentation Variability in ECGI.

Despite advances in many of the techniques used in Electrocardiographic Imaging (ECGI), uncertainty remains insufficiently quantified for many aspects of the pipeline. The effect of geometric uncertainty, particularly due to segmentation variability, may be the least explored to date. We use statistical shape modeling and uncertainty quantification (UQ) to compute the effect of segmentation variability on ECGI solutions. The shape model was made with Shapeworks from nine segmentations of the same patient and incorporated into an ECGI pipeline. We computed uncertainty of the pericardial potentials and local activation times (LATs) using polynomial chaos expansion (PCE) implemented in UncertainSCI. Uncertainty in pericardial potentials from segmentation variation mirrored areas of high variability in the shape model, near the base of the heart and the right ventricular outflow tract, and that ECGI was less sensitive to uncertainty in the posterior region of the heart. Subsequently LAT calculations could vary dramatically due to segmentation variability, with a standard deviation as high as 126ms, yet mainly in regions with low conduction velocity. Our shape modeling and UQ pipeline presented possible uncertainty in ECGI due to segmentation variability and can be used by researchers to reduce said uncertainty or mitigate its effects. The demonstrated use of statistical shape modeling and UQ can also be extended to other types of modeling pipelines.

Big Data and Artificial Intelligence: Opportunities and Threats in Electrophysiology.

The combination of big data and artificial intelligence (AI) is having an increasing impact on the field of electrophysiology. Algorithms are created to improve the automated diagnosis of clinical ECGs or ambulatory rhythm devices. Furthermore, the use of AI during invasive electrophysiological studies or combining several diagnostic modalities into AI algorithms to aid diagnostics are being investigated. However, the clinical performance and applicability of created algorithms are yet unknown. In this narrative review, opportunities and threats of AI in the field of electrophysiology are described, mainly focusing on ECGs. Current opportunities are discussed with their potential clinical benefits as well as the challenges. Challenges in data acquisition, model performance, (external) validity, clinical implementation, algorithm interpretation as well as the ethical aspects of AI research are discussed. This article aims to guide clinicians in the evaluation of new AI applications for electrophysiology before their clinical implementation.