Julian S. Haimovich

Right ventricular (RV) size and function inform prognosis in many cardiovascular diseases, but assessment by two-dimensional transthoracic echocardiography is challenging. Here we show that a deep learning model, Dimensional Reconstruction of Imaging Data-Right Ventricle (DROID-RV), can estimate RV end-diastolic diameter and identify RV systolic dysfunction from transthoracic echocardiography videos. We trained DROID-RV in 17,141 ambulatory cardiology patients (51,188 studies and 383,881 video clips) and validated model predictions in 9,112 primary care patients and 831 patients with RV ejection fraction assessed by three-dimensional transthoracic echocardiography. Across multiple test samples, DROID-RV accurately estimates RV end-diastolic diameter and identifies RV dysfunction, although model prediction error increases at extremes of RV size and function. We additionally show that RV size and function predictions are associated with incident mortality, atrial fibrillation and heart failure. Our work demonstrates how deep learning can automate RV assessment and risk stratification at scale using two-dimensional echocardiography.

Background: Stressor‐associated atrial fibrillation (AF) refers to new‐onset AF that occurs with a reversible, acute stressor. Identifying individuals at highest risk for AF recurrence is essential to guide management. Although clinical factors have shown limited value, the utility of contemporary artificial intelligence (AI)‐enabled models using the 12‐lead ECG to estimate recurrence risk remains unknown. Methods: We retrospectively analyzed consecutive primary care and cardiology patients with stressor‐associated AF occurring during hospitalization. We quantified the cumulative incidence of recurrence accounting for death as a competing risk. We investigated the relationship between time‐varying recurrence and a composite end point of AF‐related adverse events (stroke, heart failure, all‐cause death) using Cox models. We then developed and validated a penalized regression model to predict recurrence using clinical factors, stressor type, and AF risk estimates from a previously validated ECG‐based AI model. Results: We analyzed 3371 patients with stressor‐associated AF (mean age, 69±12 years; 40% women). Over a median of 3.7 years (interquartile range, 1.8–7.2), the 10‐year cumulative incidence of AF recurrence was 41% (95% CI, 39–44). AF recurrence was strongly associated with AF‐related adverse events (hazard ratio, 2.24 [95% CI, 1.81–2.76]). A model incorporating clinical factors, stressor type, and ECG‐based AI model AF risk estimates (clinical–AI) discriminated AF recurrence (area under the receiver operating characteristic curve, 0.768 [95% CI, 0.707–0.830]) favorably compared with clinical features (area under the receiver operating characteristic curve, 0.707 [95% CI, 0.642–0.772]; P <0.05). Conclusions: AF recurrence rates following stressor‐associated AF are considerable and are associated with substantially higher risk of adverse cardiovascular events. Models incorporating ECG‐based AI risk estimates may prioritize individuals for intensive monitoring and preventive interventions.

Background: Stressor-associated atrial fibrillation (AF), defined as newly diagnosed AF in the setting of a reversible physiological stressor, is common. However, risk factors, outcomes, and current management practices remain poorly understood. Methods: We analyzed data from VITAL-AF, a pragmatic, cluster-randomized AF screening trial conducted 2018-2019 comprising adults aged ≥65 across 16 primary care practices affiliated with Massachusetts General Hospital. All participants had longitudinal follow-up for adjudicated incident AF (including whether stressor-associated vs non-stressor-associated [‘primary’]) and clinical outcomes. We compared associations between clinical AF risk factors and incident AF group (stressor-associated vs primary) using Fine-Gray models handling death and each AF group as competing risks. We also quantified anticoagulation (OAC) initiation rates following AF diagnosis. We then fit Cox proportional hazards models to quantify associations between incident AF group (as a time-varying covariate) and a composite endpoint of major bleeding, stroke, and all-cause mortality, with adjustment for CHA _₂ DS _₂ -VASc (stroke) and ATRIA (bleeding) scores and time-varying OAC exposure. Results: We analyzed data from VITAL-AF, a pragmatic, cluster-randomized AF screening trial conducted 2018-2019 comprising adults aged ≥65 across 16 primary care practices affiliated with Massachusetts General Hospital. All participants had longitudinal follow-up for adjudicated incident AF (including whether stressor-associated vs non-stressor-associated [‘primary’]) and clinical outcomes. We compared associations between clinical AF risk factors and incident AF group (stressor-associated vs primary) using Fine-Gray models handling death and each AF group as competing risks. We also quantified anticoagulation (OAC) initiation rates following AF diagnosis. We then fit Cox proportional hazards models to quantify associations between incident AF group (as a time-varying covariate) and a composite endpoint of major bleeding, stroke, and all-cause mortality, with adjustment for CHA _₂ DS _₂ -VASc (stroke) and ATRIA (bleeding) scores and time-varying OAC exposure. Conclusions: Among over 30,000 primary care patients with adjudicated AF events, nearly one-third of incident AF cases were stressor-associated. Despite similar risk factor profiles and comparably high rates of adverse outcomes, OAC initiation is lower when AF is stressor-associated. Future work is needed to define optimal approaches to prevention, surveillance, and management of stressor-associated AF.

Background: Despite standardised approaches, subjective assessment and inconsistent diagnostic testing for chest pain in the emergency department (ED) drive costs, disparities and adverse outcomes. Artificial intelligence offers potential to automate and improve risk stratification. Methods and results: Using a retrospective cohort of 15 048 patients presenting to the ED of a tertiary care hospital, we trained a neural network classifier (‘Chest Pain-AI’ or ‘CP-AI’) to predict a 7-day composite endpoint of major cardiovascular diagnoses including myocardial infarction, pulmonary embolism, aortic dissection and all-cause mortality. Inputs to CP-AI included age, sex, cardiac biomarkers (D-dimer or troponin I or T positivity) and numerical representations of presenting 12-lead ECGs. ECG representations were derived using a publicly available deep learning model known as patient contrastive learning of representations. In an external validation set of 14 476 patients, we evaluated CP-AI against comparator models, including a ‘Biomarker Model’ incorporating clinical data (age, sex, biomarker positivity), based on both the area under the receiver operating characteristic curve (AUROC) and area under the precision-recall curve (AUPRC). CP-AI outperformed the Biomarker Model in prediction of the 7-day composite endpoint with an AUROC of 0.82 (95% CI 0.81 to 0.83) vs 0.79 (95% CI 0.78 to 0.81) and an AUPRC of 0.46 (95% CI 0.44 to 0.49) vs 0.35 (95% CI 0.33 to 0.37) (p<0.05 for both comparisons). Conclusions: CP-AI, a fully automated neural network classifier, demonstrated superior performance in the prediction of 7-day major cardiovascular diagnoses for patients presenting with acute chest pain compared with conventional models trained on demographics and cardiac biomarkers. CP-AI may standardise and expedite risk stratification of patients presenting to the ED with chest pain.

Background: Cardiac magnetic resonance imaging (CMR) is a powerful diagnostic modality that provides detailed quantitative assessment of cardiac anatomy and function. Automated extraction of CMR measurements from clinical reports that are typically stored as unstructured text in electronic health record systems would facilitate their use in research. Existing machine learning approaches either rely on large quantities of expert annotation or require the development of engineered rules that are time-consuming and are specific to the setting in which they were developed. Objective: We hypothesize that the use of pretrained transformer-based language models may enable label-efficient numerical extraction from clinical text without the need for heuristics or large quantities of expert annotations. Here, we fine-tuned pretrained transformer-based language models on a small quantity of CMR annotations to extract 21 CMR measurements. We assessed the effect of clinical pretraining to reduce labeling needs and explored alternative representations of numerical inputs to improve performance. Methods: Our study sample comprised 99,252 patients that received longitudinal cardiology care in a multi-institutional health care system. There were 12,720 available CMR reports from 9280 patients. We adapted PRAnCER (Platform Enabling Rapid Annotation for Clinical Entity Recognition), an annotation tool for clinical text, to collect annotations from a study clinician on 370 reports. We experimented with 5 different representations of numerical quantities and several model weight initializations. We evaluated extraction performance using macroaveraged F1-scores across the measurements of interest. We applied the best-performing model to extract measurements from the remaining CMR reports in the study sample and evaluated established associations between selected extracted measures with clinical outcomes to demonstrate validity. Results: All combinations of weight initializations and numerical representations obtained excellent performance on the gold-standard test set, suggesting that transformer models fine-tuned on a small set of annotations can effectively extract numerical quantities. Our results further indicate that custom numerical representations did not appear to have a significant impact on extraction performance. The best-performing model achieved a macroaveraged F1-score of 0.957 across the evaluated CMR measurements (range 0.92 for the lowest-performing measure of left atrial anterior-posterior dimension to 1.0 for the highest-performing measures of left ventricular end systolic volume index and left ventricular end systolic diameter). Application of the best-performing model to the study cohort yielded 136,407 measurements from all available reports in the study sample. We observed expected associations between extracted left ventricular mass index, left ventricular ejection fraction, and right ventricular ejection fraction with clinical outcomes like atrial fibrillation, heart failure, and mortality. Conclusions: This study demonstrated that a domain-agnostic pretrained transformer model is able to effectively extract quantitative clinical measurements from diagnostic reports with a relatively small number of gold-standard annotations. The proposed workflow may serve as a roadmap for other quantitative entity extraction.