European heart journal. Digital health最新文献

Aims: Atrial fibrillation (AF) is a common arrythmia that increases the risk of thromboembolism. Despite the effectiveness of anticoagulation in AF, underuse remains a substantial problem. Clinical decision support (CDS) systems may increase adherence to guideline recommended anticoagulation in AF. However, evidence regarding the cost-effectiveness of these interventions is lacking. The aim of this study was therefore to evaluate the cost-effectiveness of a CDS for AF.

Methods and results: We developed a disease progression model with a Markov structure and simulated a cohort of hypothetical individuals with AF through a standard of care and a CDS strategy. The adherence to anticoagulation in the model was based on the treatment effect reported in the CDS-AF trial, which evaluated the effect of a CDS in patients with AF in the primary care in Östergötland, Sweden. The cost-effectiveness of the CDS-AF intervention compared with standard of care was determined by estimating costs and quality-adjusted life years (QALYs) gained over a lifetime time horizon and was reported as an incremental cost-effectiveness ratio (ICER) assessed against a decision-threshold of €50 000. Uncertainty was evaluated using both one-way and probabilistic sensitivity analysis (PSA). The CDS-intervention resulted in fewer ischaemic strokes but more bleedings. The mean per patient gain in QALYs was 0.012 and the ICER was €963 per QALY gained. The result of the PSA indicated a high probability that the ICER was below €50 000.

Conclusion: The CDS intervention used in the CDS-AF trial appears to yield health gains at a lower cost than typically considered cost-effective.

Trial registration: NCT02635685.

Hypertrophic cardiomyopathy (HCM) is a heterogeneous disease where, despite recent advances, accurate diagnosis, risk stratification, and personalized treatment remain challenging. Artificial intelligence (AI) offers a transformative approach to HCM by enabling rapid, precise analysis of complex data. This article reviews the current and potential applications of AI in HCM. AI enhances diagnostic accuracy by analysing electrocardiograms, echocardiography, and cardiac magnetic resonance images, differentiating HCM from other forms of left ventricular hypertrophy, identifying subtle phenotypic variations, and standardizing myocardial fibrosis assessment. Multimodal AI-driven approaches improve risk stratification, therapeutic decision-making, and monitoring of both established and novel therapies, such as cardiac myosin inhibitors. Emerging AI-driven in silico trials and digital twin platforms highlight the potential of combining data-driven and knowledge-based AI with biophysical models to simulate patient-specific disease trajectories, supporting preclinical evaluation and personalized care. As a multidisciplinary case study, the SMASH-HCM consortium is presented to illustrate how digital twin technologies and hybrid modelling can bring AI into clinical practice. Integration of genetic data further enhances AI's ability to identify at-risk individuals and predict disease progression. However, widespread AI applications raise concerns regarding data privacy, ethical considerations, and the risk of biases. Guidelines for researchers and developers-e.g. on trustworthy AI, regulatory frameworks, and transparent policies-are essential to address these possible pitfalls. As AI rapidly evolves, it has the potential to revolutionize drug discovery, disease management, and the patient journey in HCM, making interventions more precise, timely, and patient-centred.

Aims: Among patients with acute pulmonary embolism (PE), rapid identification of those with highest clinical risk can help guide life-saving treatment. However, current risk stratification algorithms involve a multistep process requiring physical exam, imaging, and laboratory results. We investigated the utility of electrocardiogram (ECG) alone to rapidly identify patients at elevated clinical risk by developing and validating a feature-based artificial intelligence (AI) model to predict clinical risk.

Methods and results: Patients who were diagnosed with PE over a 9-year period, had an ECG within 1 day of presentation, and were evaluated by our PE response team (PERT) were included. A feature-based random forest model was trained to predict the PERT team's risk stratification from the ECG alone. The ability of the model to predict the clinical risk categorization and the accuracy of both risk stratification approaches in predicting mortality were examined on a holdout test set. Of the overall cohort of 1376 patients, 55% had a submassive (intermediate risk) or massive (high risk) PE, which were grouped together as 'severe PE'. The AI-ECG model was able to predict the clinical classification (low-risk vs. severe PE) with an AUC of 0.83 and F1 score of 0.78 in a holdout test set. A 30-day mortality and in-hospital mortality were significantly different between patients classified by the model as low vs. elevated risk.

Conclusion: AI-based analysis of 12-lead ECGs may provide a useful tool in the risk stratification of PE, allowing for rapid identification and treatment of those at highest risk of adverse outcomes.

Introduction: Sensor placement, activity type influencing wrist movements, and individual characteristics impact accuracy of wrist-worn photoplethysmography (PPG)-based heart rate (HR) monitors. This study investigated technical interventions to optimize PPG accuracy in patients with cardiac disease.

Methods and results: The Fitbit Inspire 2 PPG monitor was evaluated across three cohorts, using a Polar H10 chest strap as reference: (ⅰ) 10 healthy volunteers performed wrist movements with the monitor placed one or three fingers above the wrist to identify optimal placement; (ⅱ) 10 volunteers engaged in sport activities (walking, running, cycling, rowing); (ⅲ) 30 cardiac rehabilitation patients were monitored during exercise to assess baseline accuracy. Patients with low accuracy [mean absolute percentage error (MAPE) < 10% for <70% of training time] underwent technical interventions (sensor cleaning, forearm shaving, position fixation, and/or relocation to the volar wrist). Placement three vs. one fingers above the wrist was significantly more accurate (mean difference in MAPE: -11.4%, P < 0.001). Walking showed the highest accuracy (MAPE = 3.8%), followed by cycling (MAPE = 6.9%) and running (MAPE = 8.5%), while rowing had the lowest accuracy (MAPE = 13.4%, P < 0.001). Among CR patients, 66.7% achieved high baseline accuracy. Technical interventions improved accuracy in 50.0% of those with low baseline accuracy, but no significant predictors of optimization success were identified.

Conclusion: Accurate PPG-based monitoring requires a sensor placed higher on the wrist. Nevertheless, only two-thirds of patients are suitable for such monitoring, with improvement by technical adaptations possible (but impractical) in the others. Therefore, assessing baseline accuracy is a prerequisite before relying on these devices for activity guidance.

Aims: The role of temporary mechanical circulatory support (tMCS) after out-of-hospital cardiac arrest (OHCA) remains controversial. This study evaluates machine learning (ML) models for predicting mortality and neurological outcomes, highlighting their potential as a tool to guide early tMCS decision-making.

Methods and results: This retrospective study analysed five years of data from 564 adult non-traumatic OHCA patients treated at Marburg University Hospital. Four ML models (ANN, SVM, RF, XGBoost) were trained to predict in-hospital mortality and neurological outcome based on demographic, clinical, and treatment-related variables. Feature selection and SHAP analysis were used to optimize performance and identify patients potentially benefiting from tMCS. Overall, 144 patients (31.2%) out of 461 patients who fulfilled the inclusion criteria received tMCS: 39 left-ventricular microaxial flow pump, 76 venoarterial extracorporeal membrane oxygenation (VA-ECMO), and 29 biventricular support (ECMELLA). In 69 patients (14.9%) VA-ECMO implantation was performed as part of extracorporeal cardiopulmonary resuscitation. The survival rate of the tMCS group was 34.7% (50/144) compared to 52.7% (167/317) in the non-tMCS group. The highest predictive power for survival probability (with/without tMCS) could be achieved by XGBoost and RF when applied to the non-tMCS group. Machine learning identified 2.5% of non-tMCS patients likely to survive if treated with tMCS. In 23 (RF model) and 31 (XGBoost model) patients, the probability of survival increased by at least 5% with tMCS compared to their predicted outcome without tMCS. RF slightly outperformed XGBoost [area under the receiver operating characteristic curve (AUC) 0.85 vs. AUC 0.82].

Conclusion: XGBoost and RF models accurately predict mortality and tMCS benefit in OHCA patients, supporting ML-based personalized therapy.

Aims: Heart failure with preserved ejection fraction (HFpEF) is difficult to diagnose due to the lack of a definitive diagnostic marker; multiple tests are required, including advanced evaluations. This study aimed to develop an artificial intelligence (AI)-enabled electrocardiogram (ECG) model for predicting HFpEF.

Methods and results: This retrospective cohort study included patients from a single tertiary centre who underwent echocardiography, N-terminal prohormone of B-type natriuretic peptide measurement, and ECG within a defined timeframe. Patients were classified as HFpEF (HFA-PEFF score ≥5) or control (HFA-PEFF score <5). Patients were divided into training, validation, and test subsets at a 7:1:2 ratio for model development and validation. Using the collected ECGs, a convolutional neural network was trained to predict HFpEF; its performance was assessed using the area under the receiver operating characteristic curve (AUROC). Among the 13 081 patients included, 5795 (44.3%) were classified as HFpEF and 7286 (55.7%) were classified as control. The AI-enabled ECG model demonstrated good discriminative performance [AUROC 0.81; 95% confidence interval (CI) 0.79-0.82]. Subgroup analyses stratified by HFpEF risk factors confirmed consistent model performance. Prognostic evaluation revealed that patients with a positive AI-ECG classification experienced significantly worse outcomes relative to those with a negative classification, including higher risks of cardiac death (1.1% vs. 0.1%; hazard ratio 9.56; 95% CI 1.24-73.53; P = 0.030) and heart failure hospitalization (2.8% vs. 0.6%; hazard ratio 5.91; 95% CI 2.08-16.81; P = 0.001) at 5 year.

Conclusion: The AI-ECG model is a reliable tool for predicting HFpEF, as defined by the HFA-PEFF score, and effectively stratifies patients according to prognosis. Integration of this model into clinical practice may simplify and enhance the diagnostic process for HFpEF.

Heart failure (HF) is a global pandemic and accounts for substantial morbidity and healthcare expenditure, largely due to frequent hospitalizations. While traditionally HF patients are followed with intermittent clinical assessments, wearable technologies offer continuous, real-time monitoring, potentially enabling earlier detection and tailored interventions to prevent hospitalization. This systematic review evaluates the impact of non-invasive wearable devices on hospitalizations in HF. Following PRISMA guidelines, literature searches were conducted in PubMed and Scopus using keywords related to HF, hospitalization, and wearable technology on 1 March 2024, and re-run on 3 December 2024. Studies assessing the link between wearable devices and HF-related hospitalization rates were included. Data extraction covered population characteristics, study design, type of device, and hospitalization rates. Risk of bias was assessed using ROBINS-I and ROB-2 tools. Meta-analysis was attempted but not performed due to significant heterogeneity (I²>90%). From 2247 records, eight studies involving 1823 patients were finally analysed. Devices included ReDS, VitalPatch, ZOLL LifeVest, and ZOLL-HFMS, with follow-up ranging from 30 to 646 days. Wearable devices allowed prediction of HF hospitalization within 6.5-32 days in advance. Wearable-guided therapy compared to traditional assessment showed an 89% relative reduction at 30 days in a single-blind randomized-controlled trial, and 78% and 87% reductions in 30-day and 90-day hospitalization rates in observational studies. Although these data highlight the potential of wearable devices in HF management, future research should test predefined wearable-guided treatment algorithms on strong endpoints and address cost-effectiveness and data security in large randomized-controlled trials with longer follow-up. Registration This review was registered with PROSPERO (CRD42024519282).

Aims: Mitral and tricuspid regurgitation (MR and TR) are common in older adults and associated with substantial morbidity and mortality. While transthoracic echocardiography (TTE) is the diagnostic gold standard, access remains limited in many care settings. Artificial intelligence (AI)-based echocardiographic analysis may help address this diagnostic gap.

Methods and results: We externally validated a deep learning algorithm developed by Aisap.ai using TTE studies from the Mayo Clinic Health System (2013-23). The model analyses echocardiographic images to classify atrioventricular regurgitation severity and was evaluated against cardiologist interpretations. Performance was assessed using binary (normal-mild vs. moderate-severe) and ordinal (normal, mild, moderate, severe) classification schemes. Among 1541 eligible TTEs, the model returned predictions for 578 studies (38%). Performance analysis was limited to these cases. The MR cohort included 280 studies and the TR cohort 298. For MR, the model achieved an area under the receiver operating characteristic curve (AUC) of 0.98 [95% confidence interval (CI): 0.97-0.99], with 91% accuracy, 95% sensitivity, and 89% specificity. For TR, the AUC was 0.96 (95% CI: 0.94-0.98), with 84% accuracy, 91% sensitivity, and 80% specificity.

Conclusion: In cases where a prediction was generated, the model demonstrated high diagnostic performance in identifying clinically significant atrioventricular regurgitation. These findings support the feasibility of AI-assisted echocardiography in diverse populations, while underscoring the need for technical alignment between model requirements and local acquisition practices to ensure real-world applicability.

Aims: Data-driven clustering techniques may improve heart failure (HF) categorisation and provide prognostic insights. The present study aimed to elucidate the underlying pathophysiology of acute HF phenotypes based on pulmonary and systemic congestion at both the tissue (PTC, pulmonary tissue congestion; STC, systemic tissue congestion) and intravascular (PIVC, pulmonary intravascular congestion; SIVC, systemic intravascular congestion) level and to assess the association of identified phenotypes with a composite outcome of HF hospitalisation and death.

Methods and results: Nineteen clinical, laboratory, and echocardiographic congestion markers were analyzed using clustering techniques to identify phenotypes in patients with worsening HF in the Nancy-HF cohort (n = 741), followed by validation of the clustering model in the BIOSTAT-CHF cohort (n = 4254). Network analysis was conducted using 363 proteins to identify underlying biological pathways. Five congestion phenotypes were identified: (1) PTC-dilated left ventricle (LV), (2) PTC-HFpEF, (3) PTC, STC-atrial fibrillation (AF), (4) PIVC-dilated left atrium (LA) and LV and (5) Global congestion. Compared with the 'PTC-dilated LV' phenotype, the risk of composite outcome was higher in 'PTC, STC-AF' and 'Global' congestion phenotypes [adjusted HR: 1.74 (1.13-2.67) and 2.41 (1.60-3.63), respectively]. In BIOSTAT-CHF, 'Global' congestion phenotype was associated with significantly higher risk [HR: 1.64 (1.04-2.58)]. In network analysis, the immune response pathway was linked to all phenotypes. 'PTC-HFpEF' was related to lipid, protein and angiotensin metabolism, 'PTC, STC-AF' was related to kinase-mediated signalling, extracellular matrix organisation and TNF-regulated cell death, while 'PIVC-dilated LA & LV' was related to kinase-mediated signalling and hemostasis.

Conclusion: In worsening HF, clustering techniques identified clinical congestion profiles associated with both long-term clinical risk and differences in biomarkers, suggesting potential different underlying pathophysiologies. These clusters can be applied using the available online model to identify phenotypes as well as associated risks (https://cic-p-nancy.fr/ai-cong-hf/).

Aims: The effectiveness of telehealth care programmes in reducing mortality among patients with chronic conditions has been well established. Valuable insights into patients' conditions can be gleaned through daily telecommunication between patients and nurse case managers. We hypothesized that using natural language processing can predict acute deterioration in patients with chronic conditions in telehealth care programme based on the nursing records and speech dialogues occurring during daily telecommunication.

Methods and results: We conducted a retrospective study utilizing audio recording transcripts from telecommunication sessions between patients and nurse case managers at our telehealth care centre, along with nursing notes as input data. Pre-trained transformer-based neural network models were constructed to predict emergency room (ER) visits within a 2-week timeframe. The case group included 94 patients with 585 speech recordings and nursing records, while the control group included 36 patients with 396 speech recordings and nursing records. Our results showed that employing transcripts and a bidirectional encoder representations from transformers (BERT)-base model with a sliding window for predicting ER visits yielded moderate accuracy 0.75 (interquartile range: 0.742, 0.773). The inclusion of long short-term memory in the model did not significantly enhance accuracy. Notably, combining nursing records and transcripts as inputs exhibited superior performance, achieving an overall accuracy of 0.892 (interquartile range: 0.891, 0.893) by the six models.

Conclusion: Our study demonstrates the feasibility of predicting ER visits using telehealth dialogue transcripts and nursing notes with pre-trained transformer models. The incorporation of nursing notes significantly enhances the model's performance, providing a valuable method for improving predictive accuracy in telehealth care.