ECG-Based Deep Learning and Clinical Risk Factors to Predict Atrial Fibrillation.

BACKGROUND: Artificial intelligence (AI)-enabled analysis of 12-lead ECGs may facilitate efficient estimation of incident atrial fibrillation (AF) risk. However, it remains unclear whether AI provides meaningful and generalizable improvement in predictive accuracy beyond clinical risk factors for AF. METHODS: We trained a convolutional neural network (ECG-AI) to infer 5-year incident AF risk using 12-lead ECGs in patients receiving longitudinal primary care at Massachusetts General Hospital (MGH). We then fit 3 Cox proportional hazards models, composed of ECG-AI 5-year AF probability, CHARGE-AF clinical risk score (Cohorts for Heart and Aging in Genomic Epidemiology-Atrial Fibrillation), and terms for both ECG-AI and CHARGE-AF (CH-AI), respectively. We assessed model performance by calculating discrimination (area under the receiver operating characteristic curve) and calibration in an internal test set and 2 external test sets (Brigham and Women's Hospital [BWH] and UK Biobank). Models were recalibrated to estimate 2-year AF risk in the UK Biobank given limited available follow-up. We used saliency mapping to identify ECG features most influential on ECG-AI risk predictions and assessed correlation between ECG-AI and CHARGE-AF linear predictors. RESULTS: The training set comprised 45 770 individuals (age 55±17 years, 53% women, 2171 AF events) and the test sets comprised 83 162 individuals (age 59±13 years, 56% women, 2424 AF events). Area under the receiver operating characteristic curve was comparable using CHARGE-AF (MGH, 0.802 [95% CI, 0.767-0.836]; BWH, 0.752 [95% CI, 0.741-0.763]; UK Biobank, 0.732 [95% CI, 0.704-0.759]) and ECG-AI (MGH, 0.823 [95% CI, 0.790-0.856]; BWH, 0.747 [95% CI, 0.736-0.759]; UK Biobank, 0.705 [95% CI, 0.673-0.737]). Area under the receiver operating characteristic curve was highest using CH-AI (MGH, 0.838 [95% CI, 0.807 to 0.869]; BWH, 0.777 [95% CI, 0.766 to 0.788]; UK Biobank, 0.746 [95% CI, 0.716 to 0.776]). Calibration error was low using ECG-AI (MGH, 0.0212; BWH, 0.0129; UK Biobank, 0.0035) and CH-AI (MGH, 0.012; BWH, 0.0108; UK Biobank, 0.0001). In saliency analyses, the ECG P-wave had the greatest influence on AI model predictions. ECG-AI and CHARGE-AF linear predictors were correlated (Pearson r: MGH, 0.61; BWH, 0.66; UK Biobank, 0.41). CONCLUSIONS: AI-based analysis of 12-lead ECGs has similar predictive usefulness to a clinical risk factor model for incident AF and the approaches are complementary. ECG-AI may enable efficient quantification of future AF risk.

Patient contrastive learning: A performant, expressive, and practical approach to electrocardiogram modeling.

Supervised machine learning applications in health care are often limited due to a scarcity of labeled training data. To mitigate the effect of small sample size, we introduce a pre-training approach, Patient Contrastive Learning of Representations (PCLR), which creates latent representations of electrocardiograms (ECGs) from a large number of unlabeled examples using contrastive learning. The resulting representations are expressive, performant, and practical across a wide spectrum of clinical tasks. We develop PCLR using a large health care system with over 3.2 million 12-lead ECGs and demonstrate that training linear models on PCLR representations achieves a 51% performance increase, on average, over six training set sizes and four tasks (sex classification, age regression, and the detection of left ventricular hypertrophy and atrial fibrillation), relative to training neural network models from scratch. We also compared PCLR to three other ECG pre-training approaches (supervised pre-training, unsupervised pre-training with an autoencoder, and pre-training using a contrastive multi ECG-segment approach), and show significant performance benefits in three out of four tasks. We found an average performance benefit of 47% over the other models and an average of a 9% performance benefit compared to best model for each task. We release PCLR to enable others to extract ECG representations at https://github.com/broadinstitute/ml4h/tree/master/model_zoo/PCLR.

Deep learned representations of the resting 12-lead electrocardiogram to predict at peak exercise.

AIMS: To leverage deep learning on the resting 12-lead electrocardiogram (ECG) to estimate peak oxygen consumption (V˙O2peak) without cardiopulmonary exercise testing (CPET). METHODS AND RESULTS: V ˙ O 2 peak estimation models were developed in 1891 individuals undergoing CPET at Massachusetts General Hospital (age 45 ± 19 years, 38% female) and validated in a separate test set (MGH Test, n = 448) and external sample (BWH Test, n = 1076). Three penalized linear models were compared: (i) age, sex, and body mass index ('Basic'), (ii) Basic plus standard ECG measurements ('Basic + ECG Parameters'), and (iii) basic plus 320 deep learning-derived ECG variables instead of ECG measurements ('Deep ECG-V˙O2'). Associations between estimated V˙O2peak and incident disease were assessed using proportional hazards models within 84 718 primary care patients without CPET. Inference ECGs preceded CPET by 7 days (median, interquartile range 27-0 days). Among models, Deep ECG-V˙O2 was most accurate in MGH Test [r = 0.845, 95% confidence interval (CI) 0.817-0.870; mean absolute error (MAE) 5.84, 95% CI 5.39-6.29] and BWH Test (r = 0.552, 95% CI 0.509-0.592, MAE 6.49, 95% CI 6.21-6.67). Deep ECG-V˙O2 also outperformed the Wasserman, Jones, and FRIEND reference equations (P < 0.01 for comparisons of correlation). Performance was higher in BWH Test when individuals with heart failure (HF) were excluded (r = 0.628, 95% CI 0.567-0.682; MAE 5.97, 95% CI 5.57-6.37). Deep ECG-V˙O2 estimated V˙O2peak <14 mL/kg/min was associated with increased risks of incident atrial fibrillation [hazard ratio 1.36 (95% CI 1.21-1.54)], myocardial infarction [1.21 (1.02-1.45)], HF [1.67 (1.49-1.88)], and death [1.84 (1.68-2.03)]. CONCLUSION: Deep learning-enabled analysis of the resting 12-lead ECG can estimate exercise capacity (V˙O2peak) at scale to enable efficient cardiovascular risk stratification.

Researchers here present data describing a method of estimating exercise capacity from the resting electrocardiogram. Electrocardiogram estimation of exercise capacity was accurate and was found to predict the onset of the wide range of cardiovascular diseases including heart attacks, heart failure, arrhythmia, and death.This approach offers the ability to estimate exercise capacity without dedicated exercise testing and may enable efficient risk stratification of cardiac patients at scale.

BMI-adjusted adipose tissue volumes exhibit depot-specific and divergent associations with cardiometabolic diseases.

For any given body mass index (BMI), individuals vary substantially in fat distribution, and this variation may have important implications for cardiometabolic risk. Here, we study disease associations with BMI-independent variation in visceral (VAT), abdominal subcutaneous (ASAT), and gluteofemoral (GFAT) fat depots in 40,032 individuals of the UK Biobank with body MRI. We apply deep learning models based on two-dimensional body MRI projections to enable near-perfect estimation of fat depot volumes (R2 in heldout dataset = 0.978-0.991 for VAT, ASAT, and GFAT). Next, we derive BMI-adjusted metrics for each fat depot (e.g. VAT adjusted for BMI, VATadjBMI) to quantify local adiposity burden. VATadjBMI is associated with increased risk of type 2 diabetes and coronary artery disease, ASATadjBMI is largely neutral, and GFATadjBMI is associated with reduced risk. These results - describing three metabolically distinct fat depots at scale - clarify the cardiometabolic impact of BMI-independent differences in body fat distribution.

Identifying high risk clinical phenogroups of pulmonary hypertension through a clustering analysis.

INTRODUCTION: The classification and management of pulmonary hypertension (PH) is challenging due to clinical heterogeneity of patients. We sought to identify distinct multimorbid phenogroups of patients with PH that are at particularly high-risk for adverse events. METHODS: A hospital-based cohort of patients referred for right heart catheterization between 2005-2016 with PH were included. Key exclusion criteria were shock, cardiac arrest, cardiac transplant, or valvular surgery. K-prototypes was used to cluster patients into phenogroups based on 12 clinical covariates. RESULTS: Among 5208 patients with mean age 64±12 years, 39% women, we identified 5 distinct multimorbid PH phenogroups with similar hemodynamic measures yet differing clinical outcomes: (1) "young men with obesity", (2) "women with hypertension", (3) "men with overweight", (4) "men with cardiometabolic and cardiovascular disease", and (5) "men with structural heart disease and atrial fibrillation." Over a median follow-up of 6.3 years, we observed 2182 deaths and 2002 major cardiovascular events (MACE). In age- and sex-adjusted analyses, phenogroups 4 and 5 had higher risk of MACE (HR 1.68, 95% CI 1.41-2.00 and HR 1.52, 95% CI 1.24-1.87, respectively, compared to the lowest risk phenogroup 1). Phenogroup 4 had the highest risk of mortality (HR 1.26, 95% CI 1.04-1.52, relative to phenogroup 1). CONCLUSIONS: Cluster-based analyses identify patients with PH and specific comorbid cardiometabolic and cardiovascular disease burden that are at highest risk for adverse clinical outcomes. Interestingly, cardiopulmonary hemodynamics were similar across phenogroups, highlighting the importance of multimorbidity on clinical trajectory. Further studies are needed to better understand comorbid heterogeneity among patients with PH.

Silhouette images enable estimation of body fat distribution and associated cardiometabolic risk.

Inter-individual variation in fat distribution is increasingly recognized as clinically important but is not routinely assessed in clinical practice, in part because medical imaging has not been practical to deploy at scale for this task. Here, we report a deep learning model trained on an individual's body shape outline-or "silhouette" -that enables accurate estimation of specific fat depots of interest, including visceral (VAT), abdominal subcutaneous (ASAT), and gluteofemoral (GFAT) adipose tissue volumes, and VAT/ASAT ratio. Two-dimensional coronal and sagittal silhouettes are constructed from whole-body magnetic resonance images in 40,032 participants of the UK Biobank and used as inputs for a convolutional neural network to predict each of these quantities. Mean age of the study participants is 65 years and 51% are female. A cross-validated deep learning model trained on silhouettes enables accurate estimation of VAT, ASAT, and GFAT volumes (R2: 0.88, 0.93, and 0.93, respectively), outperforming a comparator model combining anthropometric and bioimpedance measures (ΔR2 = 0.05-0.13). Next, we study VAT/ASAT ratio, a nearly body-mass index (BMI)-and waist circumference-independent marker of metabolically unhealthy fat distribution. While the comparator model poorly predicts VAT/ASAT ratio (R2: 0.17-0.26), a silhouette-based model enables significant improvement (R2: 0.50-0.55). Increased silhouette-predicted VAT/ASAT ratio is associated with increased risk of prevalent and incident type 2 diabetes and coronary artery disease independent of BMI and waist circumference. These results demonstrate that body silhouette images can estimate important measures of fat distribution, laying the scientific foundation for scalable population-based assessment.

A Deep Learning Model for Inferring Elevated Pulmonary Capillary Wedge Pressures From the 12-Lead Electrocardiogram.

Background: Central hemodynamic parameters are typically measured via pulmonary artery catherization-an invasive procedure that involves some risk to the patient and is not routinely available in all settings. Objectives: This study sought to develop a noninvasive method to identify elevated mean pulmonary capillary wedge pressure (mPCWP). Methods: We leveraged data from 248,955 clinical records at the Massachusetts General Hospital to develop a deep learning model that can infer when the mPCWP >15 mmHg using the 12-lead electrocardiogram (ECG). Of these data, 242,216 records were used to pre-train a model that generates useful ECG representations. The remaining 6,739 records contain encounters with direct measurements of the mPCWP. Eighty percent of these data were used for model development and testing (5,390), and the remaining records comprise a holdout set (1,349) that was used to evaluate the model. We developed an associated unreliability score that identifies when model predictions are likely to be untrustworthy. Results: The model achieves an area under the receiver operating characteristic curve (AUC) of 0.80 ± 0.02 (test set) and 0.79 ± 0.01 (holdout set). Model performance varies as a function of the unreliability, where patients with high unreliability scores correspond to a subgroup where model performance is poor: for example, patients in the holdout set with unreliability scores in the highest decile have a reduced AUC of 0.70 ± 0.06. Conclusions: The mPCWP can be inferred from the ECG, and the reliability of this inference can be measured. When invasive monitoring cannot be expeditiously performed, deep learning models may provide information that can inform clinical care.

Inherited basis of visceral, abdominal subcutaneous and gluteofemoral fat depots.

For any given level of overall adiposity, individuals vary considerably in fat distribution. The inherited basis of fat distribution in the general population is not fully understood. Here, we study up to 38,965 UK Biobank participants with MRI-derived visceral (VAT), abdominal subcutaneous (ASAT), and gluteofemoral (GFAT) adipose tissue volumes. Because these fat depot volumes are highly correlated with BMI, we additionally study six local adiposity traits: VAT adjusted for BMI and height (VATadj), ASATadj, GFATadj, VAT/ASAT, VAT/GFAT, and ASAT/GFAT. We identify 250 independent common variants (39 newly-identified) associated with at least one trait, with many associations more pronounced in female participants. Rare variant association studies extend prior evidence for PDE3B as an important modulator of fat distribution. Local adiposity traits (1) highlight depot-specific genetic architecture and (2) enable construction of depot-specific polygenic scores that have divergent associations with type 2 diabetes and coronary artery disease. These results - using MRI-derived, BMI-independent measures of local adiposity - confirm fat distribution as a highly heritable trait with important implications for cardiometabolic health outcomes.

Deep learning on resting electrocardiogram to identify impaired heart rate recovery.

Background and Objective: Postexercise heart rate recovery (HRR) is an important indicator of cardiac autonomic function and abnormal HRR is associated with adverse outcomes. We hypothesized that deep learning on resting electrocardiogram (ECG) tracings may identify individuals with impaired HRR. Methods: We trained a deep learning model (convolutional neural network) to infer HRR based on resting ECG waveforms (HRRpred) among UK Biobank participants who had undergone exercise testing. We examined the association of HRRpred with incident cardiovascular disease using Cox models, and investigated the genetic architecture of HRRpred in genome-wide association analysis. Results: Among 56,793 individuals (mean age 57 years, 51% women), the HRRpred model was moderately correlated with actual HRR (r = 0.48, 95% confidence interval [CI] 0.47-0.48). Over a median follow-up of 10 years, we observed 2060 incident diabetes mellitus (DM) events, 862 heart failure events, and 2065 deaths. Higher HRRpred was associated with lower risk of DM (hazard ratio [HR] 0.79 per 1 standard deviation change, 95% CI 0.76-0.83), heart failure (HR 0.89, 95% CI 0.83-0.95), and death (HR 0.83, 95% CI 0.79-0.86). After accounting for resting heart rate, the association of HRRpred with incident DM and all-cause mortality were similar. Genetic determinants of HRRpred included known heart rate, cardiac conduction system, cardiomyopathy, and metabolic trait loci. Conclusion: Deep learning-derived estimates of HRR using resting ECG independently associated with future clinical outcomes, including new-onset DM and all-cause mortality. Inferring postexercise heart rate response from a resting ECG may have potential clinical implications and impact on preventive strategies warrants future study.

Cohort design and natural language processing to reduce bias in electronic health records research.

Electronic health record (EHR) datasets are statistically powerful but are subject to ascertainment bias and missingness. Using the Mass General Brigham multi-institutional EHR, we approximated a community-based cohort by sampling patients receiving longitudinal primary care between 2001-2018 (Community Care Cohort Project [C3PO], n = 520,868). We utilized natural language processing (NLP) to recover vital signs from unstructured notes. We assessed the validity of C3PO by deploying established risk models for myocardial infarction/stroke and atrial fibrillation. We then compared C3PO to Convenience Samples including all individuals from the same EHR with complete data, but without a longitudinal primary care requirement. NLP reduced the missingness of vital signs by 31%. NLP-recovered vital signs were highly correlated with values derived from structured fields (Pearson r range 0.95-0.99). Atrial fibrillation and myocardial infarction/stroke incidence were lower and risk models were better calibrated in C3PO as opposed to the Convenience Samples (calibration error range for myocardial infarction/stroke: 0.012-0.030 in C3PO vs. 0.028-0.046 in Convenience Samples; calibration error for atrial fibrillation 0.028 in C3PO vs. 0.036 in Convenience Samples). Sampling patients receiving regular primary care and using NLP to recover missing data may reduce bias and maximize generalizability of EHR research.

Deep Learning to Predict Cardiac Magnetic Resonance-Derived Left Ventricular Mass and Hypertrophy From 12-Lead ECGs.

BACKGROUND: Classical methods for detecting left ventricular (LV) hypertrophy (LVH) using 12-lead ECGs are insensitive. Deep learning models using ECG to infer cardiac magnetic resonance (CMR)-derived LV mass may improve LVH detection. METHODS: Within 32 239 individuals of the UK Biobank prospective cohort who underwent CMR and 12-lead ECG, we trained a convolutional neural network to predict CMR-derived LV mass using 12-lead ECGs (left ventricular mass-artificial intelligence [LVM-AI]). In independent test sets (UK Biobank [n=4903] and Mass General Brigham [MGB, n=1371]), we assessed correlation between LVM-AI predicted and CMR-derived LV mass and compared LVH discrimination using LVM-AI versus traditional ECG-based rules (ie, Sokolow-Lyon, Cornell, lead aVL rule, or any ECG rule). In the UK Biobank and an ambulatory MGB cohort (MGB outcomes, n=28 612), we assessed associations between LVM-AI predicted LVH and incident cardiovascular outcomes using age- and sex-adjusted Cox regression. RESULTS: LVM-AI predicted LV mass correlated with CMR-derived LV mass in both test sets, although correlation was greater in the UK Biobank (r=0.79) versus MGB (r=0.60, P<0.001 for both). When compared with any ECG rule, LVM-AI demonstrated similar LVH discrimination in the UK Biobank (LVM-AI c-statistic 0.653 [95% CI, 0.608 -0.698] versus any ECG rule c-statistic 0.618 [95% CI, 0.574 -0.663], P=0.11) and superior discrimination in MGB (0.621; 95% CI, 0.592 -0.649 versus 0.588; 95% CI, 0.564 -0.611, P=0.02). LVM-AI-predicted LVH was associated with incident atrial fibrillation, myocardial infarction, heart failure, and ventricular arrhythmias. CONCLUSIONS: Deep learning-inferred LV mass estimates from 12-lead ECGs correlate with CMR-derived LV mass, associate with incident cardiovascular disease, and may improve LVH discrimination compared to traditional ECG rules.

Deep learning to estimate cardiac magnetic resonance-derived left ventricular mass.

Background: Cardiac magnetic resonance (CMR) is the gold standard for left ventricular hypertrophy (LVH) diagnosis. CMR-derived LV mass can be estimated using proprietary algorithms (eg, InlineVF), but their accuracy and availability may be limited. Objective: To develop an open-source deep learning model to estimate CMR-derived LV mass. Methods: Within participants of the UK Biobank prospective cohort undergoing CMR, we trained 2 convolutional neural networks to estimate LV mass. The first (ML4Hreg) performed regression informed by manually labeled LV mass (available in 5065 individuals), while the second (ML4Hseg) performed LV segmentation informed by InlineVF (version D13A) contours. We compared ML4Hreg, ML4Hseg, and InlineVF against manually labeled LV mass within an independent holdout set using Pearson correlation and mean absolute error (MAE). We assessed associations between CMR-derived LVH and prevalent cardiovascular disease using logistic regression adjusted for age and sex. Results: We generated CMR-derived LV mass estimates within 38,574 individuals. Among 891 individuals in the holdout set, ML4Hseg reproduced manually labeled LV mass more accurately (r = 0.864, 95% confidence interval [CI] 0.847-0.880; MAE 10.41 g, 95% CI 9.82-10.99) than ML4Hreg (r = 0.843, 95% CI 0.823-0.861; MAE 10.51, 95% CI 9.86-11.15, P = .01) and InlineVF (r = 0.795, 95% CI 0.770-0.818; MAE 14.30, 95% CI 13.46-11.01, P < .01). LVH defined using ML4Hseg demonstrated the strongest associations with hypertension (odds ratio 2.76, 95% CI 2.51-3.04), atrial fibrillation (1.75, 95% CI 1.37-2.20), and heart failure (4.67, 95% CI 3.28-6.49). Conclusions: ML4Hseg is an open-source deep learning model providing automated quantification of CMR-derived LV mass. Deep learning models characterizing cardiac structure may facilitate broad cardiovascular discovery.