European heart journal. Digital health最新文献

Aims: Large language models (LLMs) have shown potential in clinical decision support, but the influence of prompt design on their performance, particularly in complex cardiology decision-making, is not well understood.

Methods and results: We retrospectively reviewed 231 patients evaluated by our Heart Team for severe aortic stenosis, with treatment options including surgical aortic valve replacement, transcatheter aortic valve implantation, or medical therapy. We tested multiple prompt-design strategies using zero-shot (0-shot), Chain-of-Thought (CoT), and Tree-of-Thought (ToT) prompting, combined with few-shot prompting, free/guided-thinking, and self-consistency. Patient data were condensed into standardized vignettes and queried using GPT4-o (version 2024-05-13, OpenAI) 40 times per patient under each prompt (147 840 total queries). Primary endpoint was mean accuracy; secondary endpoints included sensitivity, specificity, area under the curve (AUC), and treatment invasiveness. Guided-thinking-ToT achieved the highest accuracy (94.04%, 95% CI 90.87-97.21), significantly outperforming few-shot-ToT (87.16%, 95% CI 82.68-91.63) and few-shot-CoT (85.32%, 95% CI 80.59-90.06; P < 0.0001). Zero-shot prompting showed the lowest accuracy (73.39%, 95% CI 67.48-79.31). Guided-thinking-ToT yielded the highest AUC values (up to 0.97) and was the only prompt whose invasiveness did not differ significantly from Heart Team decisions (P = 0.078). An inverted quadratic relationship emerged between few-shot examples and accuracy, with nine examples optimal (P < 0.0001). Self-consistency improved overall accuracy, particularly for ToT-derived prompts (P < 0.001).

Conclusion: Prompt design significantly impacts LLM performance in clinical decision-making for severe aortic stenosis. Tree-of-Thought prompting markedly improved accuracy and aligned recommendations with expert decisions, though LLMs tended toward conservative treatment approaches.

Aims: The recurrence rate of ventricular tachycardia (VT) after ablation remains high due to the difficulty in locating VT critical sites. This study proposes a machine learning approach for improved identification of ablation targets based on intracardiac electrograms (EGMs) features derived from standard substrate mapping in a chronic myocardial infarction (MI) porcine model.

Methods and results: Thirteen pigs with chronic MI underwent invasive electrophysiological studies using multipolar catheters (Advisor™ HD grid, EnSite Precision™). Fifty-six substrate maps and 35 068 EGMs were collected during sinus rhythm and pacing from multiple sites, including left, right, and biventricular pacing. Ventricular tachycardia was induced in all pigs, and a total of 36 VTs were localized and mapped with early, mid-, and late diastolic components of the circuit. Mapping sites within 6 mm from these critical sites were considered as potential ablation targets. Forty-six signal features representing functional, spatial, spectral, and time-frequency properties were computed from each bipolar and unipolar EGM. Several machine learning models were developed to automatically localize ablation targets, and logistic regressions were used to investigate the association between signal features and VT critical sites. Random forest provided the best accuracy based on unipolar signals from sinus rhythm map, provided an area under the curve of 0.821 with sensitivity and specificity of 81.4% and 71.4%, respectively.

Conclusion: This study demonstrates for the first time that machine learning approaches based on EGM features may support clinicians in localizing targets for VT ablation using substrate mapping. This could lead to the development of similar approaches in VT patients.

Aims: Heart failure (HF) is associated with high mortality and reduced quality of life (QoL). Interventions encouraging a healthy lifestyle and self-care can reduce morbidity and HF-related hospitalizations. We conducted a randomized controlled trial (RCT) to assess the impact of a digital health programme on QoL and clinical outcomes of patients. The programme included remote patient monitoring (RPM), self-care, HF education, and empowered positive lifestyle changes.

Methods and results: Patients (n = 175) received standard-of-care (SoC) at a HF outpatient clinic (control, n = 89) or SoC plus a digital health programme (intervention, n = 86) for 6 months, followed by a 6-month maintenance period. Compliance with RPM was 93% at 6 months. No significant between-group difference was found in the primary endpoint (health-related QoL), except in an exploratory subgroup of New York Heart Association class III patients, where the intervention group had a significantly smaller QoL decline (P = 0.023). For secondary endpoints, the intervention group had significantly greater improvements in self-care at 6 months (P < 0.001) and 12 months (P = 0.003), and in disease-specific knowledge at 12 months (P = 0.001). Several exploratory endpoints favoured the intervention, with significant improvements in triglycerides (P = 0.012), HbA1c (P = 0.014), and fasting glucose (P = 0.010). The TG/HDL cholesterol ratio and TG/glucose index improved significantly at both 6 and 12 months in between-group comparisons.

Conclusion: Although the digital programme did not improve health-related QoL, it led to benefits in other important outcomes such as self-care, disease-specific knowledge, and several key metabolic parameters.

Aims: Artificial intelligence (AI) applied to a single-lead electrocardiogram (AI-ECG) can detect impaired left ventricular systolic dysfunction [LVSD: left ventricular ejection fraction (LVEF) ≤ 40%]. This study aimed to determine if AI-ECG can also predict the two-year risk of major adverse cardiovascular events (MACE) and all-cause mortality independent of LVSD.

Methods and results: Clinical outcomes after two-year follow-up were collected on patients who attended for routine echocardiography and received simultaneous single-lead-ECG recording for AI-ECG analysis. MACE and all-cause mortality were compared by Cox regression, measured against the classification of LVEF > or ≤40%. A subgroup analysis was performed on patients with echocardiographic LVEF ≥ 50%. With previously established thresholds, 'positive' AI-ECG was defined as an LVEF-predicted ≤40%, and negative AI-ECG signified an LVEF-predicted >40%; 1007 patients were included for analysis (mean age, 62.3 years; 52.4% male). 339 (33.7%) had an AI-ECG-predicted LVEF ≤ 40% and had a higher MACE rate (LVEF ≤ 40% vs. >40%: 34.2% vs.11.9%; adjusted hazard ratio (aHR) 1.93; 95% CI, 1.39-2.69; P < 0.001), primarily driven by increased mortality (23% vs. 9.6%; P < 0.001; aHR 1.56; 95% CI, 1.06-2.29; P = 0.0239). In patients with echocardiographic LVEF ≥ 50%, there was a higher incidence of MACE in those with an AI-ECG 'false positive' prediction of LVEF ≤ 40% (27.2% vs.11.9%; P < 0.001; aHR 1.71 and 95% CI, 1.11-2.47) and all-cause mortality (20.4% vs. 9.6%; P < 0.001; aHR 1.59, 95% CI, 1.09-2.42).

Conclusion: An AI-ECG algorithm designed to detect LVEF ≤ 40% can also identify patients at risk of MACE and all-cause mortality from single-lead ECG recording-independent of actual LVEF on echo. This requires further evaluation as a point-of-care risk stratification tool.

Aims: To evaluate feasibility, efficacy, and safety of standardized medical titration at home using telemonitoring. Treatment for heart failure with reduced ejection fraction (HFrEF) has advanced rapidly, emphasizing swift initiation and titration of guideline-directed medical therapy (GDMT) to improve outcomes. Implementing this in practice remains a significant challenge for healthcare. This study proposes a standardized home-based titration process incorporating home-based monitoring (HBM) to enhance GDMT titration, reduce delays, and limit the need for in-clinic assessment visits.

Methods and results: 60 patients were enrolled in this open cohort study. Standardized pre-specified titration schedules in combination with HBM were evaluated. Outcome measures included the time to optimal medical therapy (OMT), doses of GDMT at 8 weeks and 6 months, and safety evaluation through adverse events. The median time to OMT was 48 days (IQR 42-60). All participants achieved OMT within 6 months. At 8 weeks, 73%, 85%, and 88% had reached target doses for beta-blockers, ACE inhibitors, and mineral receptor antagonists, respectively. All participants reached SGLT2i target dosage. By 6 months, 62%, 73%, 80%, and 97% were on target doses for these medications, and 43% had achieved target doses for all four GDMT drugs. No serious adverse events occurred during titration.

Conclusion: We present a novel and promising approach for achieving OMT and high GDMT doses in patients with HFrEF. The utilization of standardized protocols has the potential to optimize the titration process of GDMT, and with HBM support, it can be accomplished with few in-clinic visits.

Aims: The APOLLO-QR (APPlying smartphOne for piLLs intake cOnfirmation by QR code reading) study assessed the congruence between a quick response (QR) code-based digital self-reporting and pill count in measuring medication adherence.

Methods and results: The APOLLO-QR pilot, observational study prospectively included patients owning a smartphone accepting to undergo a home-telemonitoring of ticagrelor adherence by sending feedback of each pill intake through an email generated by framing a QR code placed on the medication packaging. Ticagrelor adherence was measured at 1 and 3 months by pill count allowing to calculate accuracy of the digital self-reporting in estimating drug adherence by assessing the correspondence between the number of received feedback emails and the number of pills taken from those prescribed. Among 109 patients, 30-day adherence to ticagrelor was 98.6 ± 2.6% as measured by pill count vs. 88.9 ± 10.4% as assessed by the number of feedback emails sent by the digital self-reporting, which provided an accuracy in estimating drug adherence of 90.1 ± 10.1%. Similar results were achieved at three months among the 95 patients (87.2%) continuing the study. Only nine patients (8.3%) missed sending four consecutive feedback emails of whom three (2.8%) had voluntarily discontinued ticagrelor within 1 month. A high patient satisfaction emerged from responses to a questionnaire showing that tested telemonitoring was consistently perceived as easy, convenient, and useful, although the need for more interactivity was suggested.

Conclusion: The QR code-based self-reporting of pill intake showed a high accuracy in estimating medication adherence and yielded a good patient satisfaction, suggesting a potential for its clinical applicability.

Aims: ECG monitoring is often required during critical phases of illness. To evaluate the role of modern technology and advanced analytical algorithms artificial intelligence compared with standard-of care, we undertook a prospective, head-to-head comparison of a novel, cable-free, patch-based, and AI-enhanced electrocardiography system (CardioSenseSystem) with standard of care (SOC) ECG monitoring. Patients who had undergone cardiac surgery at a large university hospital (Skåne University Hospital, Sweden) were simultaneously monitored by both systems, and alarms and monitoring interruptions were recorded.

Methods and results: Forty-nine patients were recruited. The CardioSenseSystem system demonstrated significantly higher sensitivity, correctly detecting 364 critical red alarms vs. 12 for SOC (P < 0.0001), and lower rates of high priority false alarms (0.3% vs. 40%; P < 0.0001). Monitoring interruptions were markedly reduced (114 s/day vs. 584 s/day; P < 0.0001). Handling time per patient day was significantly shorter (256 s vs. 880 s). The CardioSenseSystem system also reduced alarm fatigue, with fewer disturbances per patient per hour (0.03 vs. 0.11; P < 0.0001).

Conclusion: The CardioSenseSystem system delivered significant advantages over conventional ECG monitoring in post-cardiac surgery patients. Its high sensitivity, reduced false alarms, fewer monitoring interruptions, and decreased handling time suggest that it may enhance patient outcomes and clinical efficiency, warranting broader application in acute-care settings.

Aims: Holter monitoring is a high prevalent technique to detect various heart pathologies. Its use has progressively increased over time with the consequent expenditure of time to interpret its results. We aim to evaluate the validity of the Cardiologs software as well as the clinical utility and potential benefits derived from the inclusion of an artificial intelligence (AI)-based software in the clinical routine of the cardiology service.

Methods and results: Concordance analyses were performed to determine the degree of correlation between the results reported by the Cardiologs software and cardiologists regarding a list of variables for 498 Holter records included in the study. Sensitivity, specificity, positive and negative prediction values, positive and negative likelihood ratios, and odds ratio were calculated. The preliminary analysis reported good correlation between the reported observations by the cardiologists involved in this study (Kappa = 0.67; P < 0001). Furthermore, an excellent concordance was found between software and cardiologists in the detection of atrial fibrillation, ventricular extrasystoles and sinus pauses of >3 s, moderate for supraventricular extrasystoles (Kappa > 0.80 in all cases), but weak or poor correlations in the rest of the variables studied. The global correlation was moderate (Kappa = 0.43; P < 0.001). Notably, the software showed sensitivity of 99.4%, negative predictive value of 99.5%, and negative likelihood ratio of 0.010, highlighting its clinical usefulness in correctly identify normal tests.

Conclusion: The inclusion of an AI-based software for reading Holter tests may have great impact in distinguishing normal Holter tests, leading to time savings and improved clinical efficiency.