International Journal of Circumpolar Health

Abstract Current risk stratification tools can limit the optimal implementation of new and emerging therapies for patients with heart rhythm disorders. For example, stroke prevention treatments have outpaced means for stroke risk stratification for patients with atrial fibrillation (AF). Artificial intelligence (AI) techniques have shown promise for improving various tasks in cardiovascular medicine. Here we explain key concepts in AI that are central to using these technologies for better risk stratification, highlighting one approach particularly well-suited to the task of portable, personalized risk stratification – probabilistic graphical models (PGMs). Probabilistic graphical models can empower physicians to ask and answer a variety of clinical questions, which we demonstrate using a preliminary model of AF-related stroke risk among 1.6 million patients within the University of Utah Health System. This example also highlights the ability of PGMs to combine social determinants of health (SDoH) and other non-traditional variables with standard clinical and demographic ones to improve personalized risk predictions and address risk factor interactions. When combined with electronic health data, these computational technologies hold great promise to empower personalized, explainable, and equitable risk assessment.

Abstract Aims Little is known about the role of electrocardiography (ECG) in the community population independent of physical and laboratory examinations. Thus, this study developed and validated several ECG-based models for cardiovascular disease (CVD) risk assessment, with or without simple questionnaire-based variables. Methods Using a derivation cohort of 3,734 Chinese participants aged ≥40 years, we developed the ECG-based models to predict the risk of developing CVD (comprising fatal and non-fatal coronary heart disease, unstable angina, stroke, and heart failure). Candidate predictors associated with CVD were screened from hundreds of ECG characteristics using a hybrid algorithm. By incorporating the questionnaire-based predictors, we constructed the ECG–questionnaire model. All models were tested in an external validation cohort (n = 1,224) to determine their discrimination and calibration. Results Over a maximum follow-up of 7 years, 433 CVD events occurred in the derivation cohort. The ECG model with 37 selected features achieved comparable performance to the clinical model using traditional cardiovascular risk factors (C-statistic: 0.690, 95% confidence interval [CI]: 0.638–0.743) in the external validation cohort. Such performance significantly improved when the questionnaire-based predictors were added (C-statistic: 0.734, 95% CI 0.685–0.784; calibration χ2: 3.334, P = 0.950). Compared with the clinical model, 17.4% of the participants were correctly assigned to the corresponding risk groups, with an absolute integrated discrimination index of 0.048 (95% CI 0.016–0.080). Conclusions The ECG model with/without questionnaire-based variables can accurately predict future CVD risk independent of physical and laboratory examinations, suggesting its great potential in routine clinical practice.

Abstract Aims Coronary artery disease (CAD) incidence continues to rise with an increasing burden of chronic coronary disease (CCD). Current probability-based risk assessment for obstructive CAD (oCAD) lacks sufficient diagnostic accuracy. We aimed to develop and validate a deep learning (DL) algorithm utilizing electrocardiogram (ECG) waveforms and clinical features to predict oCAD in patients with suspected CCD. Methods and Results The study includes subjects undergoing invasive angiography for evaluation of CCD over a 4-year period at a quaternary care center. oCAD was defined as performance of percutaneous coronary intervention (PCI) based on assessment by interventional cardiologists during elective angiography. DL models were developed for ECG waveforms alone (DL-ECG), clinical features from standard risk scores (DL-Clinical), and the combination of ECG waveforms and clinical features (DL-MM); a commonly used pre-test probability estimation tool from the CAD Consortium study was used for comparison (CAD2) [3]. The CAD2 model (AUC 0.733 [0.717-0.750]) had similar performance as the DL-Clinical model (AUC 0.762 [0.746-0.778]). The DL-ECG model (AUC 0.741 [0.726-0.758]) had similar performance as both the clinical feature models. The DL-MM model (AUC 0.807 [0.793-0.822]) had a superior performance. Validation in an external cohort demonstrated similar performance in the DL-MM (AUC 0.716 [0.707-0.726]) and CAD2 risk score (AUC 0.715 [0.705-0.724]). Conclusion A multimodality DL model utilizing ECG waveforms and clinical risk factors can improve prediction of oCAD in CCD compared to risk-factor based models. Prospective research is warranted to determine whether incorporating DL methods in ECG analysis improves diagnosis of oCAD and outcomes in CCD.

Abstract Background Aortic stenosis (AS) is a condition marked by high morbidity and mortality in severe, symptomatic cases without intervention via transcatheter aortic valve implantation (TAVI) or surgical aortic valve replacement (SAVR). Racial and ethnic disparities in access to these treatments have been documented, particularly in North America, where socioeconomic factors like health insurance confound analyses. This study evaluates disparities in AS management across racial and ethnic groups, accounting for socioeconomic deprivation, using an artificial intelligence framework. Methods We conducted a retrospective cohort study using a natural language processing (NLP) pipeline to analyse both structured and unstructured data from > 1 million patients at a London hospital. Key variables included age, sex, self-reported race and ethnicity, AS severity, and socioeconomic status. The primary outcomes were rates of valvular intervention and all-cause mortality. Results Among 6,967 AS patients, Black patients were younger, more symptomatic and more comorbid than White patients. Black patients with objective evidence of AS on echocardiography were less likely to receive a clinical diagnosis than White patients. In severe AS, TAVI and SAVR procedures were performed at lower rates among Black patients than White patients, with a longer time to SAVR. In multivariate analysis of severe AS, controlling for socioeconomic status, Black patients experienced higher mortality (HR=1.42, 95% CI=1.05-1.92, P=0.02). Conclusions An artificial intelligence framework characterises racial and ethnic disparities in AS management, which persist in a universal healthcare system, highlighting targets for future healthcare interventions.

Abstract Aims Clinical consequences of coronary artery disease (CAD) are varied (e.g., atrial fibrillation [AF] and heart failure [HF]) and current risk stratification tools are ineffective. We aimed to identify clusters of individuals with CAD exhibiting unique patterns on the electrocardiogram (ECG) in an unsupervised manner and assess their association with cardiovascular risk. Methods Twenty-one ECG markers were derived from single-lead median-beat ECGs of 1,928 individuals with CAD without a previous diagnosis of AF, HF or ventricular arrhythmia (VA) from the imaging study in UK Biobank (CAD-IMG-UKB). An unsupervised clustering algorithm was used to group these markers into distinct clusters. We characterized each cluster according to their demographic and ECG characteristics, as well as their prevalent and incident risk of AF, HF and VA (4-year median follow-up). Validation and association with prevalent diagnoses was performed in an independent cohort of 1,644 individuals. Results The model identified two clusters within the CAD-IMG-UKB cohort. Cluster 1 (N=359) exhibited prolonged QRS duration and QT intervals, along with greater morphological variations in QRS and T waves, compared to cluster 2 (N=1,569). Cluster 1, relative to cluster 2, had a significantly higher risk of incident HF (hazard ratio [HR]:2.40, 95%-confidence interval [CI]:1.51-3.83), confirmed by independent validation (HR:1.77, CI:1.31-2.41). It also showed a higher association with prevalent HF (odds ratio:4.10, CI:2.02-8.29), independent of clinical risk factors. Conclusions Our approach identified a cluster of individuals with CAD sharing ECG characteristics indicating HF risk, holding significant implications for targeted treatment and prevention enabling accessible large-scale screening.

Abstract Background Intracoronary optical coherence tomography (OCT) provides detailed information on coronary lesions, but interpretation of OCT images is time-consuming and subject to interobserver variability. Objectives To develop and validate a deep learning-based multiclass semantic segmentation algorithm for OCT (OCT-AID). Methods A reference standard was obtained through manual multiclass annotation (guidewire artefact, lumen, side branch, intima, media, lipid plaque, calcified plaque, thrombus, plaque rupture, and background) of OCT images from a representative subset of pullbacks from the PECTUS-obs study. Pullbacks were randomly divided into a training and internal test set. An additional independent dataset was used for external testing. Results In total, 2808 frames were used for training and 218 for internal testing. The external test set comprised 392 frames. On the internal test set, the mean Dice score across nine classes was 0.659 overall and 0.757 on the true-positive frames, ranging from 0.281 to 0.989 per class. Substantial to almost perfect agreement was achieved for frame-wise identification of both lipid (κ=0.817, 95% CI 0.743-0.891) and calcified plaques (κ=0.795, 95% CI 0.703-0.887). For plaque quantification (e.g. lipid arc, calcium thickness), intraclass correlations of 0.664-0.884 were achieved. In the external test set, κ-values for lipid and calcified plaques were 0.720 (95% CI 0.640-0.800) and 0.851 (95% CI 0.794-0.908), respectively. Conclusions The developed multiclass semantic segmentation method for intracoronary OCT images demonstrated promising capabilities for various classes, while having included difficult frames, such as those containing artefacts or destabilized plaques. This algorithm is an important step towards comprehensive and standardized OCT image interpretation.

Abstract Aims Remote monitoring of cardiac implantable electronic devices enables pre-emptive management of heart failure (HF) without additional patient engagement. The HeartLogic™ algorithm in implantable cardioverter defibrillators (ICDs) combines physiological parameters to predict HF events, facilitating earlier interventions. This study evaluated its diagnostic performance and resource implications within an HF management service. Methods and results In a single-centre study, 212 patients with cardiac resynchronization therapy ICDs (CRT-Ds) were monitored for 12-months. During follow-up, 18 (8%) patients died, and 15 HF hospitalizations occurred in 13 (6%) patients. Outpatient visits totalled 37 in 34 (16%) patients. HeartLogic™ alerts occurred in 58% of patients, with 100% sensitivity for HF-related hospitalizations. The positive predictive value was 29% including only alerts associated with HF events, while it was 51% including HF events and explained alerts. Unexplained alert rate was 0.46 per patient-year. Clinical interventions, mainly medication adjustments, followed 82 alerts. Total management time was 257 h/year, equivalent to 0.57 full-time equivalents for managing 1000 CRT-D patients. Conclusion The integration of HeartLogic™ into routine care demonstrated its utility in optimizing HF management, improving healthcare resource allocation. The algorithm can enhance proactive patient management and provide holistic care within the existing healthcare infrastructure.

Abstract Aims Heart failure (HF) hospitalizations are associated with poor survival outcomes, emphasizing the need for early intervention. Deep learning algorithms have shown promise in HF detection through electrocardiogram (ECG), However, their utility in ongoing HF monitoring remains uncertain. This study developed a deep learning model using 12-lead ECGs collected at two different time points to evaluate HF status changes, aiming to enhance early intervention and continuous monitoring in various healthcare settings. Methods and Results We analyzed 30,171 ECGs from 6,531 adult patients at Kobe University Hospital. The participants were randomly assigned to training, validation, and test datasets. A Transformer-based model was developed to classify HF status into deteriorated, improved, and no-change classes based on ECG waveform signals at two different time points. Performance metrics, such as the area under the receiver operating characteristic curve (AUROC) and accuracy, were calculated, and attention mapping via gradient-weighted class activation mapping was utilized to interpret the model's decision-making ability. The patients had an average age of 64.6 years (±15.4 years), brain natriuretic peptide of 66.3 pg/mL (24.6 to 175.1 pg/mL). For HF status classification, the model achieved an AUROC of 0.889 (95% CI: 0.879–0.898) and an accuracy of 0.871 (95% CI: 0.864–0.878). Conclusions Transformer-based deep learning model demonstrated high accuracy in detecting HF status changes, highlighting its potential as a non-invasive, efficient tool for HF monitoring. The reliance of the model on ECGs reduces the need for invasive, costly diagnostics, aligning with clinical needs for accessible HF management.

Abstract Aims Near-infrared spectroscopy-intravascular ultrasound (NIRS-IVUS) and optical coherence tomography (OCT) can assess coronary plaque pathology but are limited by time-consuming and expertise-driven image analysis. Recently introduced machine learning(ML)-classifiers have expedited image processing, but their performance in assessing plaque pathology against histological standards remains unclear. The aim of this study is to assess the performance of NIRS-IVUS-ML-based and OCT-ML-based plaque characterization against a histological standard. Methods and Results Matched histological and NIRS-IVUS/OCT frames from human cadaveric hearts were manually annotated and fibrotic(FT), calcific(Ca) and necrotic core(NC) regions-of-interest(ROIs) were identified. NIRS-IVUS and OCT frames were processed by their respective ML-classifiers to segment and characterize plaque components. The histologically-defined ROIs were overlaid onto corresponding NIRS-IVUS/OCT frames and the ML-classifier estimations compared with histology. In total 131 pairs of NIRS-IVUS/histology and 184 pairs of OCT/histology were included. The agreement of NIRS-IVUS-ML with histology (concordance correlation coefficient(CCC) 0.81 and 0.88) was superior to OCT-ML (CCC 0.64 and 0.73) for the plaque area and burden. Plaque compositional analysis showed a substantial agreement of the NIRS-IVUS-ML with histology for FT, Ca and NC ROIs (CCC: 0.73, 0.75 and 0.66, respectively) while the agreement of the OCT-ML with histology was 0.42, 0.62 and 0.13, respectively. The overall accuracy of NIRS-IVUS-ML and OCT-ML for characterizing atheroma types was 83% and 72%, respectively. Conclusions NIRS-IVUS-ML plaque compositional analysis has a good performance in assessing plaque components while OCT-ML performs well for the FT, moderately for the Ca and has weak performance in detecting NC tissue. This may be attributable to limitations of standalone intravascular imaging and to the fact that the OCT-ML classifier was trained on human experts rather than histological standards.

Abstract Background Cardiac telerehabilitation addresses common barriers for attendance at exercise-based cardiac rehabilitation (EBCR). Pragmatic real-world studies are however lacking, limiting generalisability of available evidence. We aimed to evaluate feasibility, safety, and patient perceptions of remotely delivered EBCR in a multicentre clinical practice setting after myocardial infarction (MI). Methods This study included 232 post-MI patients (63.7 years, 77.5% men) from 23 cardiac rehabilitation centres in Sweden (2020-2022). Exercise was delivered twice per week for three months through a real-time group-based video meeting connecting a physiotherapist to patients exercising at home. Outcomes were assessed before and after remote EBCR completion and comprised assessment of physical fitness, self-reported physical activity and exercise, physical capacity, kinesiophobia, health-related quality of life (HRQoL), self-efficacy for exercise, exercise adherence, patient acceptance. Safety monitoring in terms of adverse events (AE) and serious adverse events (SAE) was recorded. Results A total of 67.2% of the patients attended ≥ 75% of prescribed exercise sessions. Significant improvements in physical fitness, self-reported exercise, physical capacity, kinesiophobia, and HRQoL were observed. Patients agreed that remote EBCR improved health care access (83%), was easy to use (94%) and found exercise performance and interaction acceptable (95%). Sixteen exercise-related AEs (most commonly dizziness and musculoskeletal symptoms) were registered, all of which were resolved. Two SAEs requiring hospitalization were reported, both unrelated to exercise. Conclusions This multicentre study supports remote EBCR post-MI as feasible and safe with a high patient acceptance in a real-world setting. The clinical effectiveness needs to be confirmed in a randomised controlled trial. Trial registration number NCT04260958.

Abstract Aims Digital health technologies are considered promising innovations to reduce healthcare’s environmental footprint. However, this assumption remains largely unstudied. We compared the environmental impact of telemonitoring and care on site (CoS) in post-myocardial infarction (MI) follow-up and explored how it influenced patients’ and healthcare professionals’ (HPs) perceptions of using telemonitoring. Methods We conducted a mixed-method study; a standardised life cycle assessment and qualitative interviews and focus groups. We studied the environmental impact of resource use per patient for one-year post-MI follow-up in a Dutch academic hospital, as CoS or partially via telemonitoring. We used the Environmental Footprint 3.1 method. Qualitative data were analysed using Thematic Analysis. Results The environmental impact of telemonitoring was larger than CoS for all impact categories, including global warming (+480%) and mineral/metal resource use (+4,390%). Production of telemonitoring devices contributed most of the environmental burden (89%). Telemonitoring and CoS achieved parity in most impact categories at 65 km one-way patient car commute. HPs and patients did not consider the environmental impact in their preference for telemonitoring, as the patient’s individual health was their primary concern – especially after a cardiac event. However, patients and HPs were generally positive towards sustainable healthcare and willing to use telemonitoring more sustainably. Conclusion Telemonitoring had a substantially bigger environmental impact than CoS in the studied setting. Patient commute distance, reuse of devices, and tailored use of devices should be considered when implementing telemonitoring for clinical follow-up. Patients and HPs supported these solutions to enhance sustainability-informed cardiovascular care as the default option.

Abstract Aims Sudden cardiac arrest (SCA) is a commonly fatal event that often occurs without prior indications. To improve outcomes and enable preventative strategies, the electrocardiogram (ECG) in conjunction with deep learning was explored as a potential screening tool. Methods and results A publicly available data set containing 10 s of 12-lead ECGs from individuals who did and did not have an SCA, information about time from ECG to arrest, and age and sex was utilized for analysis to individually predict SCA or not using deep convolution neural network models. The base model that included age and sex, ECGs within 1 day prior to arrest, and data sampled from windows of 720 ms around the R-waves from 221 individuals with SCA and 1046 controls had an area under the receiver operating characteristic curve of 0.77. With sensitivity set at 95%, base model specificity was 31%, which is not clinically applicable. Gradient-weighted class activation mapping showed that the model mostly relied on the QRS complex to make predictions. However, models with ECGs recorded between 1 day to 1 month and 1 month to 1 year prior to arrest demonstrated predictive capabilities. Conclusion Deep learning models processing ECG data are a promising means of screening for SCA, and this method explains differences in SCAs due to age and sex. Model performance improved when ECGs were nearer in time to SCAs, although ECG data up to a year prior had predictive value. Sudden cardiac arrest prediction was more dependent upon QRS complex data compared to other ECG segments.

Abstract Aims Many research databases with anonymised 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 anonymised datasets. Methods and results We utilised 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. 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, 1,000, 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 Anonymised 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.