Aims:Hypertrophic cardiomyopathy (HCM) remains underdiagnosed due to limited access to expert imaging. We developed and validated a deep-learning (DL)-based echocardiographic model adaptable to point-of-care ultrasound (POCUS) for scalable HCM screening. Methods and results:We retrospectively analysed 134 956 expert transthoracic echocardiograms (TTE) from 73 598 patients at Sheba Medical Center (2007-2022). A TTE-trained DL model integrating structural features and temporal motion patterns from parasternal long-axis and apical four-chamber views estimated HCM probability. Performance was evaluated in an independent test cohort and clinical subgroups. External validation used bedside POCUS studies from non-cardiologists with handheld devices. The test cohort included 12 096 patients with 119 confirmed HCM cases (prevalence 0.98%; median age 75 years, 57% male). HCM-positive patients showed increased expert TTE-measured septal (1.67 [1.5, 2.0] vs. 1.01 [0.9, 1.19] cm) and posterior wall thickness (1.1 [1.0, 1.3] vs. 0.9 [0.8, 1.0] cm) (P < 0.001). The model achieved excellent discrimination with an area under the curve of 0.982 (95% CI 0.966-0.993), sensitivity 88.2%, and specificity 97.3%, robust across subgroups. The POCUS cohort (n = 1047, median age 73 years, 55% male) represented multimorbid inpatients with 65 (6.2%) classified as screen-positive by the algorithm. These showed higher expert TTE-measured septal thickness (1.26 [1.07, 1.46] vs. 1.06 [0.9, 1.2] cm; 22% vs. 4% with IVS ≥1.5 cm; P ≤ 0.01). Among 49 (75%) POCUS-flagged positive patients with formal TTE and clinical data, 8 (16%) were confirmed by expert adjudication to have HCM. Specificity is limited by occasional confounding amyloidosis detection (4% of POCUS-flagged patients). Conclusion:This DL-based model identifies HCM and demonstrates feasibility for POCUS screening, supporting earlier detection and broader diagnostic access.
Abstract Background Pulmonary hypertension (PH) is a progressive, under-recognised condition with substantial morbidity and mortality. Its non-specific presentation and multimodal diagnostic pathway contribute to delayed diagnosis. Artificial intelligence (AI)-enabled electrocardiography (ECG) may offer a scalable approach to case finding. Methods This PRISMA-DTA systematic review and meta-analysis was registered in PROSPERO (CRD420251241142). MEDLINE, Embase, Web of Science, Scopus and IEEE Xplore were searched to July 2026, with Google Scholar and citation searching used as supplementary sources. Eligible studies evaluated AI-enabled ECG in adults against RHC-confirmed PH or echocardiographic PH-related phenotypes. Two reviewers independently screened studies extracted data and assessed risk of bias using QUADAS-2. One ECG-only AUROC per study was pooled using a random-effects model. Results Seven retrospective studies were included; six contributed to meta-analysis. The pooled AUROC was 0.885 (95% CI 0.855-0.915), with very high heterogeneity (I2=98.68%; P<0.0001). The estimate combined heterogeneous RHC, echocardiographic and clinical-echocardiographic definitions and therefore did not represent accuracy for uniformly RHC-confirmed PH. Four studies performed external validation. Multimodal models outperformed ECG alone in two studies but were excluded from pooling. Conclusions AI-enabled ECG shows promising discrimination for PH-related targets, but current evidence is retrospective and heterogeneous. It should be considered an investigational case-finding or triage aid rather than a definitive diagnostic or stand-alone rule-out test.
Abstract Background Diagnosis of heart failure with preserved ejection fraction (HFpEF) using the HFA-PEFF and H2FPEF scores remains challenging in clinical practice, and relies on echocardiographic assessment. We aimed to determine whether diagnostic scoring based on automated deep learning interpretation of echocardiograms performs similar to manual measurements in diagnosing HFpEF. Methods We analyzed echocardiograms using an automated deep learning algorithm and manually in three cohorts: a test cohort (102 HFpEF patients diagnosed by right heart catheterization and echocardiography), an ambulatory validation cohort (129 HFpEF patients), and a diagnostic validation cohort (n = 427, of which 182 HFpEF and 245 non-HFpEF patients). We evaluated correlations between automated and manual HFA-PEFF and H2FPEF scores across cohorts, their correlation with pulmonary capillary wedge pressures (PCWP), and compared diagnostic accuracy using the area-under-the-curve (AUC). Results Automated and manual measurements showed good agreement across cohorts, with good correlations between HFA-PEFF (0.78-0.86) and H2FPEF (0.96-0.98) scores and similar correlations with PCWP. One in five patients with high-likelihood HFpEF based on manual HFA-PEFF scores were classified as intermediate-likelihood by automated scores due to lower estimated left atrial volumes, without consistent interaction with atrial fibrillation. AUCs for automated HFA-PEFF and H2FPEF scores did not consistently differ from manual scores (0.70 [95% confidence interval (CI): 0.66-0.74] vs. 0.71 [95% CI: 0.66-0.75], and 0.78 [95% CI: 0.73-0.82] vs. 0.75 [95% CI: 0.71-0.80], respectively). Conclusion HFA-PEFF and H2FPEF scores based on automated and manual echocardiographic analysis showed similar diagnostic accuracy, suggesting automated HFpEF diagnosis using deep learning analysis of echocardiograms is feasible.
Aims:Coronary artery disease (CAD) remains a leading cause of morbidity. Existing clinical likelihood models often lack specificity, contributing to unnecessary diagnostic testing. To develop, train, and validate the electro-mechanical risk (EMR) Score. This machine learning model uses cardiac mechanical information from resting seismocardiography (SCG) recordings and patient-level clinical risk factors to estimate obstructive CAD likelihood. Methods and results:This multi-centre clinical study included 2110 adults. Resting SCG was recorded using a sternum accelerometer. Obstructive CAD was defined as 50% or greater stenosis on coronary computed tomography angiography or invasive coronary angiography. A one-dimensional convolutional neural network was trained to compute the EMR Score. Performance was evaluated using repeated cross-validation and external-centre validation and compared with the 2024 ESC Risk-Factor-weighted Clinical Likelihood (RF-CL) model. Among 2110 participants (mean [SD] age, 57.8 [10.3] years; 801 women [38%]), 760 had obstructive CAD. In symptomatic individuals, the EMR Score achieved an AUC of 0.88, outperforming RF-CL (AUC, 0.85; P = 0.023), with higher specificity (53% vs. 35%) and comparable sensitivity (94% vs. 97%). In external-centre validation, the EMR Score achieved an AUC of 0.91, sensitivity of 97.8%, specificity of 52.0%, PPV of 78.0%, and NPV of 93.0%. In asymptomatic participants, the AUC was 0.89. The EMR Score classified 23% of the cohort as very low likelihood, with 2% CAD prevalence. Conclusion:The EMR Score non-invasively estimated obstructive CAD likelihood, with external-centre validation supporting generalizability beyond internal cross-validation. Clinical trial registration:ClinicalTrials.gov (NCT06880120, NCT06880133).
Aims:Heart failure (HF) in non-ST-segment elevation acute coronary syndrome (NSTE-ACS) is associated with poor prognosis but often under-recognized. Coronary computed tomography angiography (CCTA), increasingly used in NSTE-ACS, contains cardiopulmonary features not routinely assessed for HF. We evaluated whether an artificial intelligence (AI) algorithm applied to CCTA could identify HF likelihood in NSTE-ACS. Methods and results:In this retrospective external validation study, the AI algorithm was applied without retraining or recalibration to CCTA scans from 1009 patients with NSTE-ACS in the VERDICT trial. Using a pre-specified threshold, patients were classified as low or high AI likelihood of HF. The primary outcome was HF during index hospitalization. The secondary outcome was post-discharge HF hospitalization among patients discharged alive without HF, with analyses adjusted for global registry of acute coronary events score >140 and severe coronary artery disease. Death was treated as a competing risk. Overall, 838 patients (83%) were classified as low AI likelihood and 171 (17%) as high. During index hospitalization, HF was diagnosed in 10 patients (1%) with low AI likelihood and 12 (7%) with high. Sensitivity was 55%, specificity 84%, positive predictive value 7%, and negative predictive value 99%. High AI likelihood was associated with increased risk of index HF (subdistribution hazard ratio, 5.39, 95% confidence interval (CI) 2.32-12.50). After discharge, HF hospitalization occurred in 25 patients (3%) with low AI likelihood and 14 (8%) with high. High AI likelihood remained associated with HF hospitalization (subdistribution hazard ratio 2.56, 95% CI 1.34-4.90). Conclusion:AI-based CCTA analysis identified a large low-risk subgroup and a smaller subgroup at increased HF risk, supporting further evaluation of opportunistic HF assessment from CCTA.
Aims:Intravascular optical coherence tomography (OCT) enables high-resolution imaging of the coronary vessel wall, but manual image interpretation is time-consuming and existing automated approaches often require high computational resources and exhibit slow inference times, limiting clinical use. We developed OCT-AID-lite, a neural network for near-real-time multi-class OCT segmentation leveraging knowledge distillation and semi-supervised learning to accelerate inference while maintaining segmentation accuracy. Methods and results:A state-of-the-art model (OCT-AID) guided a compact U-Net-based student model (OCT-AID-lite) through knowledge distillation-based supervision. OCT-AID-lite was trained on 3466 manually annotated and 137 961 pseudo-labelled frames after automated quality control. On 389 internal test frames, OCT-AID-lite achieved a forward-pass time of 0.10 s for a 540-frame pullback, compared with 24.22 s for the OCT-AID model (P < 0.01). Including pre- and post-processing, total processing time was 5.50 s for OCT-AID-lite, compared with 30.62 s for OCT-AID (P < 0.01). For lipid and calcium plaque classification, OCT-AID-lite reached sensitivity/specificity of 98.1%/74.4% and 89.5%/87.5%, respectively. Pixel-wise segmentation performance on true-positive frames was high for guidewire, catheter, lumen, intima, and media (Dice: 0.79-0.99), moderate to high for sidebranch, lipid, and calcium (Dice: 0.76-0.78), and more variable for rare complex classes (Dice: 0.39-0.67). On an independent external test set, model predictions were in agreement with expert assessment. Conclusion:OCT-AID-lite enables accurate OCT segmentation in near real-time, allowing efficient quantitative characterization of plaque and vessel structures.
Early identification of acute myocardial infarction (AMI) remains challenging, particularly in non-ST-segment elevation presentations and occluded myocardial infarction, where conventional electrocardiogram (ECG) interpretation has limited sensitivity. Artificial intelligence-enabled ECG (AI-ECG) has emerged as a promising strategy to enhance early triage and diagnostic accuracy. To systematically evaluate the diagnostic performance of AI-enabled ECG algorithms for the detection of AMI, including ST-segment elevation myocardial infarction (STEMI) and non-ST-segment elevation myocardial infarction (NSTEMI), across diverse clinical settings. This diagnostic systematic review and meta-analysis was conducted in accordance with PRISMA guidelines and registered in PROSPERO (CRD420261292271). PubMed, Embase, and Cochrane CENTRAL were searched through January 2026. Studies evaluating AI-based ECG models for AMI detection and reporting sufficient data to reconstruct 2 × 2 contingency tables were included. Pooled sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were estimated using random-effects models (restricted maximum likelihood). Summary receiver operating characteristic (SROC) curves and area under the curve (AUC) were generated. Pre-specified subgroup analyses were performed for STEMI and NSTEMI Ten observational studies comprising 94 510 participants were included. For overall AMI detection,. AI-ECG demonstrated a pooled sensitivity of 89.4% (95% CI, 79.7-94.8) and specificity of 96% (95% CI, 91.2-98.2). The pooled NPV was 98.7% (95% CI, 94.1-99.7), and the pooled PPV was 73.3% (95% CI, 50.2-88.2). The SROC AUC was 0.97 (95% CI, 0.92-0.98). In STEMI, pooled sensitivity and specificity were 94.4% and 97.5%, respectively (AUC 0.98). In NSTEMI, pooled sensitivity was lower at 65.0%, with specificity of 87.5% and an AUC of 0.71. Heterogeneity was substantial, particularly among NSTEMI cohorts. AI-enabled ECG demonstrates high sensitivity and consistently excellent negative predictive value for AMI detection, supporting its role as a scalable, non-invasive triage adjunct at first medical contact. These findings highlight the potential of AI-ECG to facilitate early rule-out strategies and improve prioritization of patients requiring urgent ischaemic evaluation. Beyond diagnostic accuracy, AI-ECG may support probabilistic risk stratification and integration into early clinical decision-making pathways.
Aims:Continuous physiological monitoring outside clinical environments remains limited by usability, reproducibility, and user adherence. Vehicles offer a semi-controlled setting enabling unobtrusive multimodal sensing during everyday mobility. The Automotive Health Proof of Concept Trial (AutoHealth) investigates the real-world feasibility and accuracy of continuous in-vehicle cardiovascular monitoring using integrated optical, electrical, and acoustic sensors benchmarked against medical-grade reference standards. Methods and results:AutoHealth is a prospective, observational cohort study at Charité - Universitätsmedizin Berlin enrolling adults across five predefined cohorts: healthy individuals, patients with an elevated cardiometabolic risk, HFpEF, HFrEF, and persistent atrial fibrillation. Participants undergo comprehensive baseline phenotyping followed by a structured in-vehicle session comprising static and dynamic driving segments and predefined physical and cognitive stress tasks. Synchronized biosignals including rPPG, steering-wheel ECG, phonocardiography, and voice-derived features are compared with clinical reference measurements. The primary endpoint is the median absolute percentage error (MAPE) of in-vehicle vital sign estimates vs. reference measurements, with successful performance defined a priori as MAPE ≤10%. Secondary endpoints include the proportion of measurements meeting predefined clinical accuracy thresholds, arrhythmia detection performance, characterization of autonomic stress response, and correlations with functional mobility metrics. The study adheres to STROBE and SPIRIT-AI guidelines and is registered in the German Clinical Trials Register. Discussion:AutoHealth is designed to provide a prospective clinical validation of continuous multimodal cardiovascular monitoring inside a modified production vehicle under real-world driving conditions and reproducible closed-course testing. Study findings will characterize feasibility, performance, and translational potential of Automotive Health as a scalable prevention and remote physiological monitoring paradigm.
Aims:We aimed to develop and validate echocardiography-based prediction models for light chain (AL) and transthyretin (ATTR) cardiac amyloidosis and to quantify the incremental value of integrating artificial intelligence-derived electrocardiography (ECG) probabilities. Methods and results:We conducted a retrospective, multisite study within a single health system including patients with AL or ATTR cardiac amyloidosis and matched control subjects. AL and ATTR cohorts were handled independently. For each subtype, patients were randomly split into training and test cohorts, with an additional temporally distinct test cohort. Logistic regression models were developed using structured echocardiographic variables with parsimonious core and extended feature sets, as well as models integrating these features with a previously validated AI ECG probability. Model performance was assessed using receiver operating characteristic and precision-recall (PR) analyses. In AL amyloidosis, echocardiography-based models demonstrated moderate discrimination in the primary test cohort (area under the PR curve, AUPRC 0.674) but were inferior to ECG alone (AUPRC 0.824, P < 0.001), with no significant improvement from model combination (AUPRC 0.841 vs. ECG, P = 0.267). In contrast, for ATTR amyloidosis, echocardiography alone (AUPRC 0.687) performed worse than ECG (AUPRC 0.745, P = 0.140), while the combined model showed substantial improvement (AUROC 0.845, P < 0.001). Absolute performance declined in temporally distinct cohorts, but AUPRC comparisons were preserved. Conclusion:Subtype-specific echocardiography-based models demonstrate robust discrimination for cardiac amyloidosis, with divergent contributions of echocardiographic features by subtype. These results support tailored modelling strategies and prospective concurrent deployment of AL and ATTR models.
Aims:To develop an artificial intelligence (AI)-driven approach that analyses continuous-wave (CW) Doppler spectra from the aortic valve (AV) to detect reduced (≤40%) left ventricular ejection fraction (LVEF) without requiring dedicated two-dimensional left ventricular imaging or ECG-gated volumetric analysis for model inference. Current AI solutions for LVEF assessment rely on 2D imaging and ECG signals, limiting utility when data quality is poor. Methods and results:This retrospective study analysed 4231 aortic CW Doppler recordings from 3988 examinations (3580 patients). Preprocessing yielded 13 359 single-peak images. A CoAtNet-2 neural network was developed using patient-level training and validation cohorts and evaluated on an independent held-out test cohort. The network generated predictions for each single-peak image, and these probabilities were then averaged per examination. Maximum AV blood flow velocity and average CW Doppler pixel intensity were measured. LVEF ≤40% occurred in 20.8% of examinations, associating with lower maximum AV velocity and higher average pixel intensity. In the independent test cohort (782 examinations), the model achieved 85.2% accuracy, 79.0% sensitivity, 86.7% specificity, an AUC of 0.906, an NPV of 94.3%, and a PPV of 59.9%, with robust performance across subgroups. Conclusion:This proof-of-concept study demonstrates the feasibility of using AI to analyse CW Doppler spectra for rapid, non-invasive identification of reduced LVEF without requiring dedicated two-dimensional left ventricular imaging or ECG-gated volumetric analysis for model inference. By leveraging underutilized echocardiographic Doppler data, this signal-based approach may serve as an adjunctive screening or rule-out tool, particularly when standard imaging or ECG-gated analysis is limited or delayed.
Aims:Cardiovascular disease remains the leading driver of morbidity and mortality worldwide. As consumer smartwatches become ubiquitous, they offer a practical platform for continuous, real-world phenotyping, capturing passive and active biometrics that may detect occult disease and anticipate adverse cardiovascular events. Scaled deployment of these devices could enable remote screening, longitudinal monitoring, and earlier risk mitigation beyond traditional episodic care. Methods and results:The Health Electronic Assessment of Risks and Trends using Biometric Equipment and Technology (HEARTBEAT) study is a prospective, single-arm wearable study that has enrolled 897 participants since 10 October 2024. The primary objectives are to: (1) define associations between smartwatch-derived biometrics and incident cardiovascular events; (2) derive biometric profiles that support identification of underlying cardiometabolic and cardiovascular conditions; and (3) apply artificial intelligence to improve prediction of cardiovascular events. Participants are screened, consented, and enrolled in person or remotely. Data are captured through electronic medical record (EMR) extraction and a companion smartphone application. Participants are followed for 12 months, with outcomes adjudicated on an ongoing basis using EMR review and app-based surveys. Conclusion:The HEARTBEAT study will test whether scalable, consumer-grade wearables can move from wellness tracking to clinically meaningful signal, identifying comorbidities and predicting adverse cardiovascular outcomes in routine care settings. If successful, the study will help establish an evidence base for smartwatches as validated digital health tools for remote screening and continuous cardiovascular monitoring.
Aims:The use of cardiac signal-recording tools to monitor prognostic and diagnostic parameters in heart failure (HF) is an emerging field, constrained by usability and a low signal-to-noise environment. We evaluated a compact, user-friendly prototype for simultaneous acquisition of phonocardiography, photoplethysmography, electrocardiography, and subsequent processing to determine this strategy's usefulness in identifying parameters related to HF decompensation. Methods and results:We conducted a single-centre prospective study of 41 hospitalized patients; 36 included in the analysis. Autonomous recordings of phonocardiography, electrocardiography, and pulse-wave photoplethysmography were obtained. Signal processing employed a new supervised iterative filtering pipeline based on detection of coupled energy peaks, together with an independent analysis of the pulse-wave plethysmogram, enabling extraction of clinically relevant intervals even when one modality was of insufficient quality. In the cohort admitted with decompensated HF, measures at discharge showed shortening of electromechanical coupling time (0.148 ± 0.051 vs. 0.108 ± 0.039 s; P < 0.001), isovolumetric contraction time (0.152 ± 0.098 vs. 0.071 ± 0.023 s; P < 0.001), and ejection period (0.257 ± 0.078 vs. 0.344 ± 0.082 s; P < 0.0001) compared with admission. These findings were corroborated against the cardiology inpatient cohort without clinical congestion and contextualized using a publicly available cohort of 338 healthy individuals. Conclusion:This study presents initial evidence that a compact, easy-to-use device permits autonomous acquisition of diagnostic cardiac signals by non-expert users and that the proposed processing strategy reliably detects clinically relevant intervals in low-signal-to-noise ratio recordings, supporting its potential utility for monitoring patients with HF.
Aims:Early identification of obstructive coronary artery disease (ObCAD) is crucial because it is strongly associated with acute myocardial infarction. We developed a weighted average ensemble model integrating deep learning (DL) and machine learning (ML) to leverage imaging and clinical data for enhancing the detection of ObCAD. Methods and results:A retrospective cohort of 1054 patients was used to develop an ensemble model combining a 3D Vision Transformer with eXtreme Gradient Boosting and CatBoost for binary classification of ObCAD (>50% stenosis). Unstructured data comprised 3D cardiac non-contrast computed tomography (CT) scans, while structured data included 11 demographic and clinical features. Obstructive coronary artery disease labels were derived from corresponding coronary CT angiography. Model performance was evaluated using 10-fold cross-validation with fold-wise Wilcoxon signed-rank testing. The ensemble model achieved a mean receiver operating characteristic area under the curve (ROC AUC) of 0.81 ± 0.04 and an accuracy of 0.76 ± 0.04. It demonstrated a statistically significantly higher ROC AUC than individual component models. Feature importance analysis identified age, chest pain, and sex as the most influential predictors of ObCAD. Gradient-weighted class activation mapping visualization indicated that the 3D Vision Transformer primarily focused on cardiac regions containing coronary artery calcium deposits. Conclusion:Integrating DL-based imaging analysis with ML-based clinical modelling enhances the discriminative performance for ObCAD detection with complementary interpretability. This ensemble framework demonstrates potential to support clinical decision-making by identifying high-risk patients using routine cardiac CT combined with patient-level clinical data. Future studies using external validation and coronary artery calcium scores may further improve risk prediction.
Aims:Owing to the breadth of complex and highly dimensional clinical data associated with ageing, integration of multiple health domains is needed towards determining cardiac outcomes of older adults. We designed a machine learning (ML) approach to conglomerate multi-domain data and identify determinants of cardiac function in older adults. Methods and results:We applied a structured ML pipeline including data pre-processing, feature selection, and model development using Random Forest, Gradient Boosting, XGBoost, LightGBM, and support vector machine. Model performance was evaluated using stratified k-fold cross-validation and complementary discrimination metrics, including ROC-AUC, PR-AUC, balanced accuracy, sensitivity, and specificity. Feature importance was assessed using Random Forest (RF) importance and Shapley Additive exPlanations (SHAP), and the Tree-based Pipeline Optimization Tool (TPOT) was used for model optimization. The outcome was an impaired myocardial relaxation phenotype based on the mitral peak early-to-late diastolic filling velocity (E/A) ratio. The multi-domain dataset included demographic characteristics, clinical risk factors, physical activity, body composition, serum biomarkers, omics, and cardiac imaging, comprising 227 features from 984 older adults. Thirty key features were identified, mainly related to physical function and metabolomics. Using these features, the selected classifiers achieved ROC-AUC values above 0.79. XGBoost was retained as the primary tree-ensemble benchmark, with cross-validated ROC-AUC 0.8157 and test-set ROC-AUC 0.7658; TPOT was comparable (test-set ROC-AUC 0.7675). Higher XGBoost score was associated with death-or-admission events (HR 1.115, P = 0.029). Conclusion:Multi-domain ML identified clinically interpretable signals associated with impaired myocardial relaxation in ageing and with clinical events. Trial registration:ClinicalTrials.gov Identifier: NCT02791139.
Aims:Continuous positive airway pressure (CPAP) telemonitoring provides daily respiratory metrics, including Cheyne-Stokes breathing percentage (CSB%), which may reflect ventilatory-circulatory instability. However, inter-individual variability limits the clinical applicability of fixed thresholds. We developed a time-structured detection architecture using CPAP telemonitoring time-series data and evaluated its ability to identify temporal patterns preceding cardiovascular and cerebrovascular events. Methods and results:In this retrospective observational study, 1265 patients with obstructive sleep apnoea undergoing CPAP telemonitoring were analysed. Daily CSB% values were smoothed using a 3-day moving average and evaluated relative to individualized dynamic baselines. The detection framework consisted of two complementary components: identification of sustained deviation from baseline and detection of abrupt pre-event surges. Central apnoea predominance, assessed using the central apnoea index, was incorporated as a hierarchical escalation layer. Among 25 adjudicated cardiovascular and cerebrovascular events in 20 patients (heart failure 11, atrial fibrillation 7, cerebrovascular accident 7), the architecture detected 23 events (92.0%) within the predefined D-28 window. Median lead time was 12.0 days (IQR 2.8-24.0). False-positive alerts were concentrated within a subset of individuals. Hierarchical filtering reduced alerts by 94.9% relative to the baseline signal layer while preserving event enrichment. Distinct temporal phenotypes, including trajectory-dominant and spike-dominant patterns, were observed across disease categories, consistent with disease-specific pre-event dynamics. Conclusion:CPAP-derived CSB% may function as a time-structured digital biomarker reflecting evolving ventilatory-circulatory instability. A trajectory-based detection architecture may enable early identification of cardiovascular instability while maintaining operational feasibility in large-scale telemonitoring environments.
Aims:Current cardiovascular risk scores may underestimate the risk of future cardiovascular events, particularly in younger individuals with prolonged exposure to risk factors. In contrast, imaging-based detection of subclinical atherosclerosis provides a more accurate assessment of cardiovascular risk by identifying established vascular disease. We aimed to develop and prospectively validate an artificial intelligence-based tool using retinal fundus images to detect ultrasound-confirmed subclinical atherosclerosis. Methods and results:In this prospective observational study, 931 participants (mean age 52.6 years; 70.2% women) without prior cardiovascular disease underwent standardized clinical evaluation, non-mydriatic retinal imaging, and carotid and femoral ultrasound to detect subclinical atherosclerosis. A multimodal AI model integrating deep learning from retinal images with radiomic and clinical data was developed in a derivation cohort (n = 781) and evaluated in a held-out prospective test set (n = 150).Subclinical atherosclerosis was present in 50.8% of participants. In the prospective test set, the AI model demonstrated good discrimination. In image-only mode, the model achieved an area under the curve (AUC) of 0.80 (95% CI 0.73-0.87), with sensitivity of 88.2%. In the enhanced mode incorporating clinical variables, performance improved to an AUC of 0.86 (95% CI 0.80-0.92), with sensitivity of 93.4% and a negative predictive value of 90.6%. Discrimination was higher in younger individuals and those at low-to-intermediate cardiovascular risk. Conclusion:AI-based retinal image analysis enables non-invasive detection of systemic subclinical atherosclerosis. This scalable approach may enhance early identification of high cardiovascular-risk patients, particularly in populations in whom risk is underestimated.
Aims:To evaluate the feasibility and system-level performance of Heart2Miss, a decentralized community-based triage model deploying AI-powered point-of-care ultrasound (AI-POCUS) via a hub-intermediary-spoke approach in diabetes primary care for early heart-failure (HF) detection. Methods and results:In this prospective study, 1000 adults with diabetes and no known HF were screened over seven months across six primary care clinics (spokes); 985 with complete data were analysed. Novice biomedical and bioscience graduates underwent 4-week training to perform focused three-view handheld AI-POCUS. Images were AI-analysed and verified through the hub-intermediary-spoke pathway. The primary outcome was detection of previously undiagnosed HF. Secondary outcomes included reduction in tertiary-centre burden through the hub-intermediary-spoke pathway and novice sonographer performance. 11.1% (n = 109) had Stage B (pre-HF) and 1.0% (n = 10) Stage C HF (symptomatic HF). Rapid triage ruled out abnormality in 77.3% at the spoke and a further 12.6% after intermediary TTE confirmation, reducing tertiary diagnostic burden by 89.9%. Only 1.0% required tertiary referral. Regarding novice performance, >90% analysable scans were achieved for left-ventricular parameters and >85% for left-atrial volume. After 400 scans, scan time fell from 11.0 ± 5.3 min to 8.3 ± 4.4 min (Δ 2.31 min, 95% CI 1.52-3.11; P < 0.001), and complete three-view capture improved from 88.0% to 92.2% (P = 0.035). Conclusion:This decentralized hub-intermediary-spoke model combining AI-POCUS, telehealth verification, and a task-shifted bioscience workforce enabled early HF detection while substantially reducing specialist workload, supporting digital health-enabled workforce innovation and pathway redesign in resource-constrained settings.
Aims:Hypertension is a major contributor to cardiovascular morbidity and mortality, yet identifying individuals at risk before clinical diagnosis remains challenging. Here, we present a multi-horizon machine learning framework designed to model incident hypertension risk across multiple clinically meaningful time windows using data from 246 286 participants in the UK Biobank. The framework systematically compares predictive performance across five horizons under severe class imbalance, enabling analysis of how discrimination, precision, and risk drivers evolve as outcome prevalence changes over time. Methods and results:Seven classification algorithms were evaluated, including logistic regression, random forest, naïve Bayes, and four boosting-based ensemble methods. To enhance interpretability, we integrate SHapley Additive exPlanations (SHAP) with generative topographic mapping (GTM), combining feature-level attribution with population-level visualization of model predictions. Ensemble boosting models consistently achieved the strongest performance, with average precision increasing from 0.04 for the ≤2-year horizon to 0.22 for the ≤10-year horizon, while ROC-AUC remained relatively stable (∼0.75-0.79). Together, this framework reveals consistent predictors of hypertension risk (including baseline blood pressure, age, body mass index, medication burden, and cardiometabolic multimorbidity) and illustrates how combinations of risk factors organize hypertension risk across time horizons. Conclusions:Our results demonstrate how multi-horizon modelling and complementary explainability approaches can provide deeper insight into evolving disease risk patterns in large biomedical cohorts, supporting more interpretable and scalable strategies for population-level cardiovascular prevention. Such approaches may enable earlier identification of high-risk individuals and inform targeted screening and preventive interventions in routine care settings.
Aims:Accurate risk stratification before structural heart disease interventions is essential for clinical decision-making. Traditional risk models, such as the European System for Cardiac Operative Risk Evaluation II (EuroSCORE II) and Society of Thoracic Surgeons Predicted Risk of Mortality (STS-PROM), were designed for surgical patients and show inconsistent performance in transcatheter cohorts. Biological age, reflecting cumulative physiological decline, may offer prognostic value beyond chronological age and established risk scores. Methods and results:In this retrospective study of 1269 patients [non-transcatheter aortic valve implantation (TAVI) n = 751, TAVI n = 518] treated at the German Heart Center Munich, biological age was estimated from pre-operative chest radiographs using CXR-Age, a validated deep learning model. Analyses were conducted separately for surgical (non-TAVI) and transcatheter (TAVI) groups. For 30-day mortality, biological age outperformed EuroSCORE II in both subgroups [area under the receiver operating characteristic curve (AUC): non-TAVI 0.874 vs. 0.785, P < 0.001; TAVI 0.952 vs. 0.745, P = 0.004] and remained independently predictive after adjustment [TAVI OR 1.58 per year, 95% confidence interval (CI) 1.27-2.12]. While STS-PROM was the strongest single predictor for non-TAVI patients (AUC 0.949), it was similar to EuroSCORE II for TAVI patients (AUC 0.729). Notably, patients whose biological age exceeded their chronological age by more than 10 years faced higher major complication rates (17.3% vs. 9.2%; P = 0.016). Conclusion:Biological age distinguished risk across both populations, suggesting that deep learning-based biological age estimation from routine chest radiographs could serve as an automated, accessible complement to existing risk models.
