Detecting QT prolongation from a single-lead ECG with deep learning.

For a number of antiarrhythmics, drug loading requires a 3-day hospitalization with continuous monitoring for QT-prolongation. Automated QT monitoring with wearable ECG monitors would enable out-of-hospital care. We therefore develop a deep learning model that infers QT intervals from ECG Lead-I-the lead that is often available in ambulatory ECG monitors-and use this model to detect clinically meaningful QT-prolongation episodes during Dofetilide drug loading. QTNet-a deep neural network that infers QT intervals from Lead-I ECG-was trained using over 3 million ECGs from 653 thousand patients at the Massachusetts General Hospital and tested on an internal-test set consisting of 633 thousand ECGs from 135 thousand patients. QTNet is further evaluated on an external-validation set containing 3.1 million ECGs from 667 thousand patients at another healthcare institution. On both evaluations, the model achieves mean absolute errors of 12.63ms (internal-test) and 12.30ms (external-validation) for estimating absolute QT intervals. The associated Pearson correlation coefficients are 0.91 (internal-test) and 0.92 (external-validation). Finally, QTNet was used to detect Dofetilide-induced QT prolongation in a publicly available database (ECGRDVQ-dataset) containing ECGs from subjects enrolled in a clinical trial evaluating the effects of antiarrhythmic drugs. QTNet detects Dofetilide-induced QTc prolongation with 87% sensitivity and 77% specificity. The negative predictive value of the model is greater than 95% when the pre-test probability of drug-induced QTc prolongation is below 25%. These results show that drug-induced QT prolongation risk can be tracked from ECG Lead-I using deep learning.

ECG-guided non-invasive estimation of pulmonary congestion in patients with heart failure.

Quantifying hemodynamic severity in patients with heart failure (HF) is an integral part of clinical care. A key indicator of hemodynamic severity is the mean Pulmonary Capillary Wedge Pressure (mPCWP), which is ideally measured invasively. Accurate non-invasive estimates of the mPCWP in patients with heart failure would help identify individuals at the greatest risk of a HF exacerbation. We developed a deep learning model, HFNet, that uses the 12-lead electrocardiogram (ECG) together with age and sex to identify when the mPCWP > 18 mmHg in patients who have a prior diagnosis of HF. The model was developed using retrospective data from the Massachusetts General Hospital and evaluated on both an internal test set and an independent external validation set, from another institution. We developed an uncertainty score that identifies when model performance is likely to be poor, thereby helping clinicians gauge when to trust a given model prediction. HFNet AUROC for the task of estimating mPCWP > 18 mmHg was 0.8 [Formula: see text] 0.01 and 0.[Formula: see text] 0.01 on the internal and external datasets, respectively. The AUROC on predictions with the highest uncertainty are 0.50 [Formula: see text] 0.02 (internal) and 0.[Formula: see text] 0.04 (external), while the AUROC on predictions with the lowest uncertainty were 0.86 ± 0.01 (internal) and 0.82 ± 0.01 (external). Using estimates of the prevalence of mPCWP > 18 mmHg in patients with reduced ventricular function, and a decision threshold corresponding to an 80% sensitivity, the calculated positive predictive value (PPV) is 0.[Formula: see text] 0.01when the corresponding chest x-ray (CXR) is consistent with interstitial edema HF. When the CXR is not consistent with interstitial edema, the estimated PPV is 0.[Formula: see text] 0.02, again at an 80% sensitivity threshold. HFNet can accurately predict elevated mPCWP in patients with HF using the 12-lead ECG and age/sex. The method also identifies cohorts in which the model is more/less likely to produce accurate outputs.

QTNet: Deep Learning for Estimating QT Intervals Using a Single Lead ECG.

QT prolongation often leads to fatal arrhythmia and sudden cardiac death. Antiarrhythmic drugs can increase the risk of QT prolongation and therefore require strict post-administration monitoring and dosage control. Measurement of the QT interval from the 12-lead electrocardiogram (ECG) by a trained expert, in a clinical setting, is the accepted method for tracking QT prolongation. Recent advances in wearable ECG technology, however, raise the possibility of automated out-of-hospital QT tracking. Applications of Deep Learning (DL) - a subfield within Machine Learning - in ECG analysis holds the promise of automation for a variety of classification and regression tasks. In this work, we propose a residual neural network, QTNet, for the regression of QT intervals from a single lead (Lead-I) ECG. QTNet is trained in a supervised manner on a large ECG dataset from a U.S. hospital. We demonstrate the robustness and generalizability of QTNet on four test-sets; one from the same hospital, one from another U.S. hospital, and two public datasets. Over all four datasets, the mean absolute error (MAE) in the estimated QT interval ranges between 9ms and 15.8ms. Pearson correlation coefficients vary between 0.899 and 0.914. By contrast, QT interval estimation on these datasets with a standard method for automated ECG analysis (NeuroKit2) yields MAEs between 22.29ms and 90.79ms, and Pearson correlation coefficients 0.345 and 0.620. These results demonstrate the utility of QTNet across distinct datasets and patient populations, thereby highlighting the potential utility of DL models for ubiquitous QT tracking.Clinical Relevance- QTNet can be applied to inpatient or ambulatory Lead-I ECG signals to track QT intervals. The method facilitates ambulatory monitoring of patients at risk of QT prolongation.

Predicting outcomes in patients with aortic stenosis using machine learning: the Aortic Stenosis Risk (ASteRisk) score.

OBJECTIVE: To use echocardiographic and clinical features to develop an explainable clinical risk prediction model in patients with aortic stenosis (AS), including those with low-gradient AS (LGAS), using machine learning (ML). METHODS: In 1130 patients with moderate or severe AS, we used bootstrap lasso regression (BLR), an ML method, to identify echocardiographic and clinical features important for predicting the combined outcome of all-cause mortality or aortic valve replacement (AVR) within 5 years after the initial echocardiogram. A separate hold out set, from a different centre (n=540), was used to test the generality of the model. We also evaluated model performance with respect to each outcome separately and in different subgroups, including patients with LGAS. RESULTS: Out of 69 available variables, 26 features were identified as predictive by BLR and expert knowledge was used to further reduce this set to 9 easily available and input features without loss of efficacy. A ridge logistic regression model constructed using these features had an area under the receiver operating characteristic curve (AUC) of 0.74 for the combined outcome of mortality/AVR. The model reliably identified patients at high risk of death in years 2-5 (HRs ≥2.0, upper vs other quartiles, for years 2-5, p<0.05, p=not significant in year 1) and was also predictive in the cohort with LGAS (n=383, HRs≥3.3, p<0.05). The model performed similarly well in the independent hold out set (AUC 0.78, HR ≥2.5 in years 1-5, p<0.05). CONCLUSION: In two separate longitudinal databases, ML identified prognostic features and produced an algorithm that predicts outcome for up to 5 years of follow-up in patients with AS, including patients with LGAS. Our algorithm, the Aortic Stenosis Risk (ASteRisk) score, is available online for public use.

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.

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.

DeepStrain: A Deep Learning Workflow for the Automated Characterization of Cardiac Mechanics.

Myocardial strain analysis from cinematic magnetic resonance imaging (cine-MRI) data provides a more thorough characterization of cardiac mechanics than volumetric parameters such as left-ventricular ejection fraction, but sources of variation including segmentation and motion estimation have limited its wider clinical use. We designed and validated a fast, fully-automatic deep learning (DL) workflow to generate both volumetric parameters and strain measures from cine-MRI data consisting of segmentation and motion estimation convolutional neural networks. The final motion network design, loss function, and associated hyperparameters are the result of a thorough ad hoc implementation that we carefully planned specific for strain quantification, tested, and compared to other potential alternatives. The optimal configuration was trained using healthy and cardiovascular disease (CVD) subjects (n = 150). DL-based volumetric parameters were correlated (>0.98) and without significant bias relative to parameters derived from manual segmentations in 50 healthy and CVD test subjects. Compared to landmarks manually-tracked on tagging-MRI images from 15 healthy subjects, landmark deformation using DL-based motion estimates from paired cine-MRI data resulted in an end-point-error of 2.9 ± 1.5 mm. Measures of end-systolic global strain from these cine-MRI data showed no significant biases relative to a tagging-MRI reference method. On 10 healthy subjects, intraclass correlation coefficient for intra-scanner repeatability was good to excellent (>0.75) for all global measures and most polar map segments. In conclusion, we developed and evaluated the first end-to-end learning-based workflow for automated strain analysis from cine-MRI data to quantitatively characterize cardiac mechanics of healthy and CVD subjects.

Identifying unreliable predictions in clinical risk models.

The ability to identify patients who are likely to have an adverse outcome is an essential component of good clinical care. Therefore, predictive risk stratification models play an important role in clinical decision making. Determining whether a given predictive model is suitable for clinical use usually involves evaluating the model's performance on large patient datasets using standard statistical measures of success (e.g., accuracy, discriminatory ability). However, as these metrics correspond to averages over patients who have a range of different characteristics, it is difficult to discern whether an individual prediction on a given patient should be trusted using these measures alone. In this paper, we introduce a new method for identifying patient subgroups where a predictive model is expected to be poor, thereby highlighting when a given prediction is misleading and should not be trusted. The resulting "unreliability score" can be computed for any clinical risk model and is suitable in the setting of large class imbalance, a situation often encountered in healthcare settings. Using data from more than 40,000 patients in the Global Registry of Acute Coronary Events (GRACE), we demonstrate that patients with high unreliability scores form a subgroup in which the predictive model has both decreased accuracy and decreased discriminatory ability.

Choosing Clinical Variables for Risk Stratification Post-Acute Coronary Syndrome.

Most risk stratification methods use expert opinion to identify a fixed number of clinical variables that have prognostic significance. In this study our goal was to develop improved metrics that utilize a variable number of input parameters. We first used Bootstrap Lasso Regression (BLR) - a Machine Learning method for selecting important variables - to identify a prognostic set of features that identify patients at high risk of death 6-months after presenting with an Acute Coronary Syndrome. Using data derived from the Global Registry of Acute Coronary Events (GRACE) we trained a logistic regression model using these features and evaluated its performance on a development set (N = 43,063) containing patients who have values for all features, and a separate dataset (N = 6,363) that contains patients who have missing feature values. The final model, Ridge Logistic Regression with Variable Inputs (RLRVI), uses imputation to estimate values for missing features. BLR identified 19 features, 8 of which appear in the GRACE score. RLRVI had modest, yet statistically significant, improvement over the standard GRACE score on both datasets. Moreover, for patients who are relatively low-risk (GRACE≤87), RLRVI had an AUC and Hazard Ratio of 0.754 and 6.27, respectively, vs. 0.688 and 2.46 for GRACE, (p < 0.007). RLRVI has improved discriminatory performance on patients who have values for the 8 GRACE features plus any subset of the 11 non-GRACE features. Our results demonstrate that BLR and data imputation can be used to obtain improved risk stratification metrics, particularly for patients who are classified as low risk using traditional methods.

Solution structure and biochemical characterization of a spare part protein that restores activity to an oxygen-damaged glycyl radical enzyme.

Glycyl radical enzymes (GREs) utilize a glycyl radical cofactor to carry out a diverse array of chemically challenging enzymatic reactions in anaerobic bacteria. Although the glycyl radical is a powerful catalyst, it is also oxygen sensitive such that oxygen exposure causes cleavage of the GRE at the site of the radical. This oxygen sensitivity presents a challenge to facultative anaerobes dwelling in areas prone to oxygen exposure. Once GREs are irreversibly oxygen damaged, cells either need to make new GREs or somehow repair the damaged one. One particular GRE, pyruvate formate lyase (PFL), can be repaired through the binding of a 14.3 kDa protein, termed YfiD, which is constitutively expressed in E. coli. Herein, we have solved a solution structure of this 'spare part' protein using nuclear magnetic resonance spectroscopy. These data, coupled with data from circular dichroism, indicate that YfiD has an inherently flexible N-terminal region (residues 1-60) that is followed by a C-terminal region (residues 72-127) that has high similarity to the glycyl radical domain of PFL. Reconstitution of PFL activity requires that YfiD binds within the core of the PFL barrel fold; however, modeling suggests that oxygen-damaged, i.e. cleaved, PFL cannot fully accommodate YfiD. We further report that a PFL variant that mimics the oxygen-damaged enzyme is highly susceptible to proteolysis, yielding additionally truncated forms of PFL. One such PFL variant of ~ 77 kDa makes an ideal scaffold for the accommodation of YfiD. A molecular model for the rescue of PFL activity by YfiD is presented.

Machine Learning Improves Risk Stratification After Acute Coronary Syndrome.

The accurate assessment of a patient's risk of adverse events remains a mainstay of clinical care. Commonly used risk metrics have been based on logistic regression models that incorporate aspects of the medical history, presenting signs and symptoms, and lab values. More sophisticated methods, such as Artificial Neural Networks (ANN), form an attractive platform to build risk metrics because they can easily incorporate disparate pieces of data, yielding classifiers with improved performance. Using two cohorts consisting of patients admitted with a non-ST-segment elevation acute coronary syndrome, we constructed an ANN that identifies patients at high risk of cardiovascular death (CVD). The ANN was trained and tested using patient subsets derived from a cohort containing 4395 patients (Area Under the Curve (AUC) 0.743) and validated on an independent holdout set containing 861 patients (AUC 0.767). The ANN 1-year Hazard Ratio for CVD was 3.72 (95% confidence interval 1.04-14.3) after adjusting for the TIMI Risk Score, left ventricular ejection fraction, and B-type natriuretic peptide. A unique feature of our approach is that it captures small changes in the ST segment over time that cannot be detected by visual inspection. These findings highlight the important role that ANNs can play in risk stratification.

Hidden States within Disordered Regions of the CcdA Antitoxin Protein.

The bacterial toxin-antitoxin system CcdB-CcdA provides a mechanism for the control of cell death and quiescence. The antitoxin protein CcdA is a homodimer composed of two monomers that each contain a folded N-terminal region and an intrinsically disordered C-terminal arm. Binding of the intrinsically disordered C-terminal arm of CcdA to the toxin CcdB prevents CcdB from inhibiting DNA gyrase and thereby averts cell death. Accurate models of the unfolded state of the partially disordered CcdA antitoxin can therefore provide insight into general mechanisms whereby protein disorder regulates events that are crucial to cell survival. Previous structural studies were able to model only two of three distinct structural states, a closed state and an open state, that are adopted by the C-terminal arm of CcdA. Using a combination of free energy simulations, single-pair Förster resonance energy transfer experiments, and existing NMR data, we developed structural models for all three states of the protein. Contrary to prior studies, we find that CcdA samples a previously unknown state where only one of the disordered C-terminal arms makes extensive contacts with the folded N-terminal domain. Moreover, our data suggest that previously unobserved conformational states play a role in regulating antitoxin concentrations and the activity of CcdA's cognate toxin. These data demonstrate that intrinsic disorder in CcdA provides a mechanism for regulating cell fate.

Beatquency domain and machine learning improve prediction of cardiovascular death after acute coronary syndrome.

Frequency domain measures of heart rate variability (HRV) are associated with adverse events after a myocardial infarction. However, patterns in the traditional frequency domain (measured in Hz, or cycles per second) may capture different cardiac phenomena at different heart rates. An alternative is to consider frequency with respect to heartbeats, or beatquency. We compared the use of frequency and beatquency domains to predict patient risk after an acute coronary syndrome. We then determined whether machine learning could further improve the predictive performance. We first evaluated the use of pre-defined frequency and beatquency bands in a clinical trial dataset (N = 2302) for the HRV risk measure LF/HF (the ratio of low frequency to high frequency power). Relative to frequency, beatquency improved the ability of LF/HF to predict cardiovascular death within one year (Area Under the Curve, or AUC, of 0.730 vs. 0.704, p < 0.001). Next, we used machine learning to learn frequency and beatquency bands with optimal predictive power, which further improved the AUC for beatquency to 0.753 (p < 0.001), but not for frequency. Results in additional validation datasets (N = 2255 and N = 765) were similar. Our results suggest that beatquency and machine learning provide valuable tools in physiological studies of HRV.

A Structure-free Method for Quantifying Conformational Flexibility in proteins.

All proteins sample a range of conformations at physiologic temperatures and this inherent flexibility enables them to carry out their prescribed functions. A comprehensive understanding of protein function therefore entails a characterization of protein flexibility. Here we describe a novel approach for quantifying a protein's flexibility in solution using small-angle X-ray scattering (SAXS) data. The method calculates an effective entropy that quantifies the diversity of radii of gyration that a protein can adopt in solution and does not require the explicit generation of structural ensembles to garner insights into protein flexibility. Application of this structure-free approach to over 200 experimental datasets demonstrates that the methodology can quantify a protein's disorder as well as the effects of ligand binding on protein flexibility. Such quantitative descriptions of protein flexibility form the basis of a rigorous taxonomy for the description and classification of protein structure.

Mollack: a web server for the automated creation of conformational ensembles for intrinsically disordered proteins.

UNLABELLED: Intrinsically disordered proteins (IDPs) play central roles in many biological processes. Consequently, an accurate description of the disordered state is an important step towards a comprehensive understanding of a number of important biological functions. In this work we describe a new web server, Mollack, for the automated construction of unfolded ensembles that uses both experimental and molecular simulation data to construct models for the unfolded state. An important aspect of the method is that it calculates a quantitative estimate of the uncertainty in the constructed ensemble, thereby providing an objective measure of the quality of the final model. Overall, Mollack facilitates structure-function studies of disordered proteins. AVAILABILITY AND IMPLEMENTATION: http://cmstultz-mollack.mit.edu CONTACT: cmstultz@mit.edu SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Expanding the Range of Protein Function at the Far End of the Order-Structure Continuum.

The traditional view of the structure-function paradigm is that a protein's function is inextricably linked to a well defined, three-dimensional structure, which is determined by the protein's primary amino acid sequence. However, it is now accepted that a number of proteins do not adopt a unique tertiary structure in solution and that some degree of disorder is required for many proteins to perform their prescribed functions. In this review, we highlight how a number of protein functions are facilitated by intrinsic disorder and introduce a new protein structure taxonomy that is based on quantifiable metrics of a protein's disorder.