European heart journal. Digital health最新文献

Aims: Acute kidney injury (AKI) is a frequent and severe complication in critically ill patients with cardiovascular instability. Current risk scores rely on delayed renal biomarkers such as serum creatinine (sCr) and blood urea nitrogen (BUN). We aimed to develop and validate machine learning (ML) models predicting AKI and major adverse kidney events (MAKE) exclusively from systemic physiological and haemodynamic data.

Methods and results: Two ML models were trained on the MIMIC-IV database: one including (sCr+/BUN+) and one excluding (sCr-/BUN-) renal parameters. External validation was performed in the eICU database and in a cohort of burn ICU patients from AP-HP. Model performance was assessed for early AKI and MAKE prediction up to 100 h before diagnosis. Systemic haemodynamic and physiological variables were the strongest predictors of AKI. In MIMIC-IV, the sCr-/BUN- model achieved auROC 0.78 at 72 h, approaching the sCr+/BUN+ model. In eICU, it outperformed the biomarker-based model at later time points (auROC 0.73). In the burn ICU cohort-representing a high-stress systemic environment-it maintained robust accuracy (auROC 0.75 at 24 h, 0.77 at 72 h). For MAKE prediction, the sCr-/BUN- model achieved auROC 0.87 (burn cohort), 0.67 (eICU), and 0.77 (MIMIC-IV). Median lead time for AKI prediction exceeded 70 h.

Conclusion: AI models based solely on non-renal parameters can accurately predict AKI and MAKE, even under extreme systemic stress such as severe burns. Haemodynamic signatures carry sufficient information to anticipate kidney dysfunction well in advance, opening the way to real-time, proactive cardio-renal risk stratification in ICU patients with acute heart failure, cardiogenic shock, and after cardiac surgery.

Aims: Accurate diagnoses contribute to the improvement of clinical workflows and the enhancement of patient care. Commercially available automated electrocardiogram (ECG) interpretation systems require manual review by physicians despite their widespread use. This study investigates the frequency and characteristics of the modifications from automated ECG reports in routine clinical practice.

Methods and results: We retrospectively analysed 159 630 ECGs from 2011 to 2023 and compared automated preliminary ECG reports generated by the GE Marquette™ 12SL ECG analysis programme with finalized reports by physicians. A modification was defined as any textual difference between the initial and final reports. Our analysis revealed that 31.3% of all ECG reports underwent some forms of modification by physicians. We analysed the frequency of 69 pre-defined ECG-related terms before and after physician review, categorizing modifications as unchanged, deleted, or newly added. Modifications were more frequent for ECGs performed during off-hours, in patients with higher ventricular rates and longer QRS durations. At the term-level, diagnoses such as 'prolonged QT interval' (newly added from 5.6% of original reports) and 'electronic ventricular pacemaker' (newly added from 3.6% of original reports) were frequently added by physicians, while diagnoses like 'inferior infarct' and 'anterior infarct' were frequently deleted from automated ECG reports (32.0% and 44.6% automated reports with these terms required removals).

Conclusion: This large-scale real-world study demonstrated the high frequency of physicians' modification in automated ECG interpretation. The identified patterns of modifications highlight the limitations of current rule-based systems in handling complex cases and nuanced ECG findings.

Aims: Unenhanced chest CT is frequently used to assess lung malignancy and parenchymal disease. Harnessing CT data to quantify cardiac and vascular structures has the potential to improve the diagnosis of heart failure and pulmonary hypertension (PH). This study aims to develop a deep learning model to segment and analyse cardiothoracic structures from unenhanced CT images to diagnose PH, pre-capillary PH and PH associated with left heart disease (LHD).

Methods and results: A twelve-structure cardiothoracic segmentation model was developed using an institutional cohort (n = 55, 35/9/11 training/validation/testing). Model performance was evaluated using Dice similarity coefficients (DSC). Volumetric measurements were compared to manual values using intra-class correlation (ICC) and visually assessed by four observers using an external cohort (n = 50, from 26 hospitals). Univariable and multivariable regression analyses were performed using a cohort of 368 patients (254/114 training/testing). Receiver-operating characteristic curves were plotted and the area under the curves (AUC) with confidence intervals (CI) were calculated. The model yielded a DSC segmentation performance of ≥0.87 for 9/12 segmented structures and ICC > 0.95 for 10/12 structures. Most of the segmented structures scored as excellent in the external cohort visual assessment. Diagnostic accuracy for predicting PH was high [AUC = 0.88 (CI: 0.80-0.96), sensitivity = 70%, specificity = 100%], including pre-capillary PH [AUC = 0.84 (CI: 0.74-0.94), sensitivity = 72%, specificity = 94%] and PH-LHD [AUC = 0.86 (CI: 0.79-0.93), sensitivity = 94%, specificity = 63%].

Conclusion: A fully automated model for multi-structure cardiothoracic segmentation on unenhanced CT is achievable. The model can predict PH and identify patients with pre-capillary PH and PH-LHD with promising performance.

Aims: Stress echocardiography (SE) is widely used for assessing coronary artery disease, but volumetric chamber analysis during SE is limited by time-consuming manual tracings and operator-dependent variability. Automated evaluation may overcome these barriers and enhance efficiency.

Methods and results: This multi-centre study included 240 participants undergoing pharmacological SE for ischaemic heart disease evaluation from five sites in four countries. SE imaging data from apical four-chamber and two-chamber views were acquired during rest and stress phases. Expert cardiologists manually traced endocardial borders for left ventricular (LV), left atrial (LA) and right ventricular (RV), right atrial (RA) areas, which were compared to machine learning (ML) derived measurements. Image quality was categorized as optimal, good, fair, or poor, and its influence on ML performance was analysed. Statistical methods included Intraclass Correlation Coefficients (ICCs), Bland-Altman testing, and within-patient coefficient of variation. The yield of the ML algorithm demonstrated consistency across rest and stress phases. It demonstrated strong agreement with cardiologists for LV and LA volumes, with ICCs ranging from 0.84 to 0.93 across rest and stress conditions. RA and RV areas measurements showed moderate correlations, with better agreement at rest than during stress phases. Image quality significantly influenced ML performance, as poor-quality images reduced diagnostic yield.

Conclusion: AI-driven volumetric analysis is a reliable method for quantifying left-sided heart chambers during pharmacological SE, with results closely matching expert measurements. Moderate reliability for right-sided chambers highlights the need for high-quality imaging and standardized protocols. AI integration may streamline SE workflows and support improved clinical decision-making.

Aims: Coded healthcare data are now commonly used in clinical research. This study aimed to assess the transparency of reporting within heart failure studies and employ machine learning to facilitate larger-scale evaluation.

Methods & results: A systematic search of EMBASE and MEDLINE (2015-2020) identified 4279 heart failure studies with accessible Extensible Markup Language published in the top 25 journals by impact factor. Manual extraction in a random sample of 170 studies by independent human reviewers characterized 40 studies (23.5%) that used coded healthcare data, with 34 of these (85%) reporting doing so and only 19 (47.5%) providing clear descriptions of dataset construction and linkage. Another 420 studies underwent manual annotation to further train a Natural Language Processing (NLP) model designed for this study to automate and upscale review. The NLP model processed 3689 studies with a high level of internal accuracy (area under the receiver operating characteristic curve 0.97 and F1 score 0.96). Overall, the NLP approach identified 782 studies (21.2%) that reported coded healthcare data usage (95% CI 19.8-20.9%). No correlation was found between the reporting of coded healthcare data use and the publication year (r = ^-0.05; P = 0.21) or citation count (r = ^-0.13; P = 0.12).

Conclusion: One-fifth of contemporary heart failure research articles are already reporting the use of coded healthcare data, with at-scale evaluation facilitated by a machine-learning model. The limited transparency on how coded healthcare data were used in studies highlights the need for quality standards such as the CODE-EHR framework for the use of healthcare data in research.

Aims: To propose a novel deep learning-based method, the eLVMass-Net, for the estimation of left ventricular mass (LVM) based on 12-lead electrocardiograms (ECGs).

Methods and results: We developed a deep learning model for LVM estimation using raw ECG signals, demographic data, and ECG parameters as input by using TW-CVAI dataset (n = 1459). Synchronized single-heartbeat waveforms were processed using a temporal convolutional network (TCN). Ground-truth LVM values were obtained from coronary computed tomography angiography. We performed external validation on an independent NTUH dataset (n = 2579). To account for sex-specific differences in left ventricular remodelling and body habitus, we further developed separate models for males and females. We compared the performance of the eLVMass-Net, with two state-of-the-art (SOTA) models.Non-sex-specific eLVMass-Net achieved a mean absolute error (MAE) of 14.3 ± 0.7 g and a mean absolute percentage error (MAPE) of 12.9 ± 1.1% between predicted and ground-truth LVM values under five-fold cross-validation. The eLVMass-Net outperformed two SOTA models in terms of both LVM estimation and left ventricular hypertrophy (LVH) classification. Sex-specific design was superior in LVH classification based on estimated LVM (c-statistic: 0.77 ± 0.05 for male model; 0.75 ± 0.05 for female model; 0.70 ± 0.02 for non-sex-specific model; P  < 0.01 between both sex-specific models vs. non-sex-specific model). The saliency maps revealed gender-specific differences in how the model weighted ST-T segment features for LVM prediction.

Conclusion: The proposed eLVMass-Net outperformed previously published approaches by ECG pre-processing with synchronized single heartbeat extraction and TCN as ECG encoder. Additionally, the development of sex-specific models proved to be a rational approach.

Aims: Automatic echocardiographic measurements using artificial intelligence have shown promising results; however, they have not been compared with manual measurements regarding heart failure (HF) progression and algorithm runtime.

Methods and results: Data came from the prospective HF study MyoVasc (NCT04064450), which involved a highly standardized 5-h examination, including comprehensive echocardiography, at a dedicated study centre between January 2013 and April 2018. Worsening of HF was a primary composite endpoint, recorded by structured follow-up, death certificates, and medical records. The automated assessment was performed using EchoDL, eight 3D convolutional neural networks (CNNs) trained to predict clinical parameters. Manual and automatic left ventricular ejection fraction (LVEF), E/E'-ratio and left ventricular mass (LVM) demonstrated a good intraclass correlation coefficient {LVEF: 0.75 [95% confidence interval (CI) 0.75-0.77], E/E'-ratio: 0.59 [CI 0.56-0.61], LVM: 0.64 [CI 0.62-0.66]}. After a median follow-up of 3.8 years (IQR 2.1-5.0), 470 patients experienced worsening of HF. In multivariable Cox analysis, comparison of manually and automatically assessed LVEF, E/E'-ratio and LVM demonstrated risk estimates slightly in favour of the CNNs. Direct comparison of C-indices showed significantly better model performance for automatically determined LVEF (0.71 vs. 0.73, P = 0.038) and E/E'-ratio (0.64 vs. 0.66, P = 0.013) and a trend for LVM (0.66 vs. 0.68, P = 0.063). Echo-DL required an average of 1053.4 ms (95% CI 1050.7-1056.0) to analyse a four-second-long echocardiogram.

Conclusion: Automated analysis of echocardiograms using 3D CNNs was comparable to manual measurements in predicting HF-specific outcomes. Echo-DL offers potential time savings and improved risk prediction in clinical settings, allowing integration into echocardiographic hardware.

Aims: Epicardial adipose tissue (EAT), located within the pericardial sac, has emerged as a biomarker for coronary artery disease (CAD) progression. This study aimed to develop and validate a deep learning-based system for automated EAT volume quantification using non-contrast computed tomography (NCCT) scans from a large, multi-centre, pan-Asian cohort.

Methods and results: A total of 1243 NCCT patient scans from three centres were used to train and internally validate a deep learning model based on 3D UNet++ architecture for pericardium segmentation, followed by intensity thresholding to derive EAT volume. Epicardial adipose tissue quantification required ∼30 s per scan. The final model was evaluated on an external testing cohort of 160 patients, including 90 non-Asian individuals. In this cohort, AI-predicted EAT volumes showed excellent agreement with expert annotations (r = 0.975; P < 0.0001). The Bland-Altman analysis demonstrated a mean bias of -5.2 cm³with 95% limits of agreement from -25.1 to 14.7 cm³. Among the non-Asian subgroup, model performance remained strong (r = 0.970; bias, -3.2 cm³; limits of agreement, -25.1-18.7 cm³). AI-derived EAT volume was independently associated with obstructive CAD (odds ratio 1.11; 95% confidence interval, 1.04-1.19; P = 0.004), after adjusting for confounders. The global χ² statistic increased from 81.7 with coronary calcium score alone to 93.3 when EAT volume was added (P = 0.001), indicating improved risk prediction.

Conclusion: We developed and validated a deep learning system for automated EAT volume quantification from NCCT scans. The model demonstrated high accuracy and generalizability across ethnically diverse populations, supporting its potential for routine EAT assessment and CAD risk stratification.

Trial registration: ClinicalTrials.gov Identifier: NCT05509010.

Aims: There is a lack of studies investigating the pathophysiologic and phenotypic distinctiveness of aortic stenosis (AS). This heterogeneity has important implications for identifying optimal intervention timing and potential medical management. This study seeks to identify phenogroups of AS using unsupervised machine learning to improve risk stratification.

Methods and results: A total of 349 patients with asymptomatic AS from the PROGRESSA study were included in this analysis. Echocardiographic, clinical and blood sample data were used in the unsupervised clustering process. Longitudinal echocardiographic data were used to evaluate AS progression. Five clusters of patients were revealed using 18 variables selected by an unsupervised machine learning algorithm. Amongst them, aortic valvular phenotype, mean gradient, peak jet velocity (V_peak), and left ventricle stroke volume were selected as discriminatory variables. Following the clustering process, characteristics differed between clusters, including age, body mass index, and sex ratio (all P < 0.001). Of note, cluster 1 showed higher AS severity at baseline with significantly higher initial V_peak (344 [314; 376] cm/s) and calcium score (1257 [806; 1837] UA) (P < 0.001). Patients from cluster 1 had a faster AS progression (progression of V_peak = 22 [9; 39] cm/s/year), and calcium score (213 [111; 307] UA/year) (P < 0.001). Cluster 1 was also associated with a higher composite risk of mortality and aortic valve replacement when adjusted for age, sex, and baseline AS severity (P < 0.001).

Conclusion: Artificial intelligence-guided phenotypic classification revealed 5 distinct groups and enhanced risk stratification of patients with AS. This approach may be useful to optimize and individualize medical and interventional management of AS.

Aims: Non-ischaemic dilated cardiomyopathy (NIDCM) is a major cause of heart failure (HF) and heart transplantation (HTx), characterized by heterogeneity in aetiology, clinical phenotype, and disease progression. Nevertheless, precision medicine-based diagnostics and treatment strategies for NIDCM remain lacking. This proof-of-concept study aimed to stratify NIDCM patients by pathological features and identify those at high-risk for malignant arrhythmia (MA) and rapid progression to end-stage HF.

Methods and results: 293 NIDCM-HTx patients were included in this study. A total of 3516 heart tissue slides from six representative sites of each patient were analyzed using deep learning-based computational pathology (DL-CPath) and unsupervised clustering to identify pathological subgroups (PGs): PGA, PGB, and PGC. PGA was characterized by interstitial fibrosis, cardiomyocyte vacuolization, microvascular intimal hyperplasia, and myocyte disarray, and had the highest rates of MA (P = 0.03) and the shortest interval from diagnosis to HTx (P = 0.03). PGB showed focal fibrosis, whereas PGC demonstrated the mildest histopathological alterations. For clinical features, PGA showed elevated levels of blood biomarkers indicative of myocardial and secondary organ injury. PGB was associated with extensive fibrosis and significant impairment of ejection fraction. PGC presented with the mildest clinical abnormalities. Although LMNA mutation was a significant non-DL-CPath high-risk factor for MA and rapid NIDCM progression, its distribution did not differ significantly across PGs (P = 0.786).

Conclusion: DL-based pathological classification effectively extracted clinically-meaningful imaging features and enabled the identification of high-risk NIDCM subgroup. Each PG exhibited unique histopathological and clinical characteristics, highlighting distinct phenotypes and risk profiles.