Abstract Background Artificial intelligence (AI) holds considerable promise for improving cardiovascular care, but prior to widespread adoption AI-powered tools require evaluation in randomised trials. We systematically reviewed the current evidence from randomised trials of AI tools in cardiovascular medicine. Methods We conducted a systematic review of randomised trials comparing AI tools vs. standard care in cardiovascular medicine. PubMed, the Cochrane Central Register of Controlled Trials, Scopus, and the International Clinical Trials Registry Platform were searched from January 1, 2018, to April 10, 2026. Screening and full-text review were performed by three independent reviewers and risk of bias was assessed using the RoB 2 criteria. Results Of 11,151 identified records, 8,599 remained after removal of duplicates. Following screening and full-text review, 25 trials met the eligibility criteria. These trials evaluated AI tools in acute cardiovascular care (n = 3), cardiac imaging (n = 6), decision support for medical therapy (n = 6), electrophysiology (n = 8), and interventional cardiology (n = 2). Nine trials were single-centre, and two were international. Seventeen trials (68%) reported positive results, defined by non-inferiority or superiority of the AI-based strategy. Overall risk of bias was rated as high or with some concerns in 84% of the included trials, and only 36% explicitly reported following the CONSORT-/SPIRIT-AI reporting guidelines. Discussion Evidence from randomised trials evaluating AI tools in cardiovascular medicine is limited and most trials have a high risk of bias. Wider implementation of AI tools into practice requires robust, international, multi-centre randomised trials that evaluate patient-relevant outcomes, ensure external validity, and adhere to reporting guidelines. Trial registration number: PROSPERO CRD420251019264
Abstract Aims Identifying the infarct-related artery (IRA) before coronary angiography in non–ST-segment elevation myocardial infarction (NSTEMI) is difficult because current approaches have limited sensitivity. We evaluated whether a noninvasive 90-second optically pumped magnetometer magnetocardiography (OPM-MCG) recording can estimate pre-angiography IRA probabilities in these patients and provide information beyond standard-care variables. Methods and Results This prospective two-campus study within one hospital system included a Campus 1 development cohort (n=493) and a geographically separate Campus 2 test cohort (n=118). All patients underwent OPM-MCG before angiography. An XGBoost model using 65 prespecified OPM-MCG features was developed with nested cross-validation, locked, and evaluated once in Campus 2. IRA adjudication followed a core-laboratory hierarchy. Discrimination, calibration, and clinical utility were compared with ECG and TTE. Among 611 patients, 43% had LAD, 23% LCx, and 34% RCA IRA. In Campus 2, overall accuracy was 63.6%, macro-F1 was 60.0%; LAD/LCx/RCA vessel-specific areas under the receiver operating characteristic curve were 0.87/0.79/0.79 and top-1 sensitivity was 76.6%/34.5%/69.0%. Accuracy was 51.7% for matched ECG-only XGBoost and 55.9% for integrated standard care. In an exploratory stacked analysis, adding locked OPM-MCG probabilities to standard care increased accuracy to 67.8% (paired difference 11.9%, 95% CI 4.2–19.5; P=0.008). Conclusions In haemodynamically stable patients with NSTEMI already selected for invasive management, a locked OPM-MCG model provided informative vessel-level probabilities in Campus 2 and complementary signal with standard care. Performance was strongest for LAD and RCA. Prospective multicentre, cross-platform testing is needed before clinical use. Trial registration number: ChiCTR2300070211.
Abstract Background Optical coherence tomography (OCT) provides excellent resolution at the cost of shallow penetration, limiting its ability for plaque burden (PB) assessment. A deep learning (DL) algorithm was therefore developed to provide fully automated assessment of PB and vessel size from OCT. Objectives We aimed to perform a head-to-head comparison of AI-powered OCT versus intravascular ultrasound (IVUS) for PB quantification, using histopathology as the reference standard. Methods Ten post-mortem human hearts were harvested. All the 29 major epicardial coronary arteries were perfused and underwent OCT and IVUS pullbacks. After imaging, the arteries were serially sectioned at 1-mm intervals for histopathology. Lumen and vessel contours were derived fully automatically from OCT using the DL algorithm and semi-automatically from IVUS. Histopathological measurements were obtained by experienced analysts. Statistical analyses accounted for within-patient clustering. Results A total of 852 paired OCT-IVUS-histopathology cross-sections were obtained. AI-powered OCT showed strong correlations with histopathology for the assessment of lumen area (r=0.89 [95% CI: 0.88–0.91]), vessel area (r=0.85 [95% CI: 0.83–0.87]), and PB (r=0.76 [95% CI: 0.73–0.79]). The corresponding correlations between IVUS and histopathology were 0.90 (95% CI: 0.89–0.92), 0.86 (95% CI: 0.84–0.88), and 0.79 (95% CI: 0.76–0.81), respectively. Comparisons of the OCT–histopathology and IVUS–histopathology correlation coefficients yielded p=0.336 for lumen area, p=0.444 for vessel area, and p=0.161 for PB. Both OCT and IVUS showed good diagnostic performance in identifying PB>65%, with the area under the receiver operating characteristic curve being 0.92 (95% CI: 0.90-0.94) and 0.92 (95% CI: 0.88-0.95), respectively. Conclusions In this histopathology-validation study of 10 post-mortem human hearts, AI-powered OCT enabled fully automated plaque burden quantification and correlated strongly with histopathology, with no statistically significant difference between the OCT–histopathology and IVUS–histopathology correlations.
Abstract Aims Artificial intelligence (AI) models trained on electrocardiograms (ECGs) can estimate age, with deviations from chronological age (Δage) serving as markers of accelerated cardiovascular ageing. We evaluated whether transformer-based AI models trained on globally diverse, heterogeneous ECG datasets outperform convolutional neural network (CNN)-based models in generalizability and physiological interpretability. Methods and results Using 11 public datasets (435 203 recordings) from four continents and an independent external cohort (14 462 recordings), we trained and validated transformer and CNN architectures. The transformer model trained on global data achieved strong correlation with chronological age and maintained robust performance in the external cohort (r = 0.90), outperforming both a CNN trained on global data and a model trained exclusively on a homogenous population-based cohort (r = 0.88). Individuals with Δage ≥+8 years exhibited a two-fold increased mortality risk (hazard ratio = 2.08, 95% confidence interval 1.93–2.24). Explainability analysis identified the P-wave, particularly in lead V1, as the dominant contributor to age prediction, consistent with known electrophysiological changes associated with atrial remodelling and ageing. The model’s attribution to early atrial conduction features was consistent across age strata, peaking in midlife (30–50 years), and correlated positively with disease burden, including hypertension and smoking. Conclusion Our findings establish that transformer models trained on globally heterogeneous data generalize well and can provide insights to plausible ageing signatures. Furthermore, Δage may offer insight into underlying physiological ageing processes and associated risk, though further validation is needed to define its clinical utility in risk stratification.
Abstract Aim To identify sleep stage-specific heart rate variability (HRV) phenotypes and evaluate their associations with long-term cardiovascular outcomes. Methods Data were obtained from the Sleep Heart Health Study (SHHS). After exclusion of participants with missing clinical covariates or outcome data, 4,029 individuals were included. Linear and non-linear HRV features were extracted from stage-specific electrocardiographic segments during N2, N3, and REM sleep. Results Three HRV-informed phenotypes were identified within the selected clustering framework across N2, N3, and REM sleep stages: Stable (Cluster 0), Burden (Cluster 1), and High-risk (Cluster 2). Compared with Stable phenotype, High-risk phenotype was consistently associated with increased cardiovascular risk, with strongest association observed for incident CVD in N2 stage (HR 1.468, 95% CI 1.280–1.684, P < 0.001). Similar associations were observed in N3 and REM stages. Sequential adjustment for AHI, oxygen desaturation index, arousal index, and sleep efficiency resulted in only minimal attenuation of hazard estimates, indicating that the prognostic associations were largely independent of conventional respiratory measures. In contrast, Burden phenotype was consistently associated with lower all-cause mortality across all three sleep stages (HR range 0.83–0.85, all P < 0.001). Subgroup analyses stratified by OSA severity demonstrated broadly consistent prognostic associations across AHI strata. Conclusions Incorporating sleep stage-specific HRV features into conventional clinical phenotyping provides complementary prognostic information beyond traditional respiratory severity metrics. HRV-informed phenotyping may improve long-term cardiovascular risk stratification in patients with obstructive sleep apnea.
