Automated and Interpretable Patient ECG Profiles for Disease Detection, Tracking, and Discovery

Abstract

Background: The ECG remains the most widely used diagnostic test for characterization of cardiac structure and electrical activity. We hypothesized that parallel advances in computing power, machine learning algorithms, and availability of large-scale data could substantially expand the clinical inferences derived from the ECG while at the same time preserving interpretability for medical decision-making.

Methods and results: We identified 36 186 ECGs from the University of California, San Francisco database that would enable training of models for estimation of cardiac structure or function or detection of disease. We segmented the ECG into standard component waveforms and intervals using a novel combination of convolutional neural networks and hidden Markov models and evaluated this segmentation by comparing resulting electrical intervals against 141 864 measurements produced during the clinical workflow. We then built a patient-level ECG profile, a 725-element feature vector and used this profile to train and interpret machine learning models for examples of cardiac structure (left ventricular mass, left atrial volume, and mitral annulus e-prime) and disease (pulmonary arterial hypertension, hypertrophic cardiomyopathy, cardiac amyloid, and mitral valve prolapse). ECG measurements derived from the convolutional neural network-hidden Markov model segmentation agreed with clinical estimates, with median absolute deviations as a fraction of observed value of 0.6% for heart rate and 4% for QT interval. Models trained using patient-level ECG profiles enabled surprising quantitative estimates of left ventricular mass and mitral annulus e' velocity (median absolute deviation of 16% and 19%, respectively) with good discrimination for left ventricular hypertrophy and diastolic dysfunction as binary traits. Model performance using our approach for disease detection demonstrated areas under the receiver operating characteristic curve of 0.94 for pulmonary arterial hypertension, 0.91 for hypertrophic cardiomyopathy, 0.86 for cardiac amyloid, and 0.77 for mitral valve prolapse.

Conclusions: Modern machine learning methods can extend the 12-lead ECG to quantitative applications well beyond its current uses while preserving the transparency that is so fundamental to clinical care.

Keywords: heart rate; hypertension; machine learning; mitral valve prolapse; work flow.

Figure 1.

Workflow for ecgAI project. The workflow consisted of training an ECG segmentation model and using a selected group of ECGs to train interpretable models to estimate cardiac structure and function and detect and track disease. Concordance with measurements from the GE MUSE system was used for validation of segmentation as well as a filter for segmentation quality. The number of ECGs used for the various tasks is indicated in parentheses. For disease detection, a slash separates the number of cases and control ECGs used. Curved rectangles represent training data; ellipses represent algorithms; and standard rectangles represent other data types. CNN indicates convolutional neural network; HCM, hypertrophic cardiomyopathy; HMM, hidden Markov model; MVP, mitral valve prolapse; NSR, normal sinus rhythm; PAH, pulmonary arterial hypertension; LA, left atrium; and LV, left ventricle.

Figure 2.

ecgAI method of ECG segmentation. A, Architecture of convolutional neural network used for ECG segmentation. Gray rectangles represent layers with dimensions listed below. The notation for each layer indicates the size of the input (eg, 2000 ms, initially) by the number of leads (eg, 12) by the number of filters. The size of the filter is specified in the body of the rectangle. B, Architecture of hidden Markov model (HMM) used after convolutional neural network (CNN) based segmentation. Gray boxes represent states that are traversed in order in the ECG. C, Example of CNN-HMM output for an ECG. CNN-HMM based classes are shown below the image. The ST and T wave segments have been combined.

Figure 3.

Comparison of ecgAI (hidden Markov model [HMM]+convolutional neural network [CNN]) derived measurements and MUSE/University of California, San Francisco (UCSF) values for 4 commonly reported ECG measurements. The scatterplot depicts 35 466 comparisons. The line y=x is drawn to help identify any bias. The unit for heart rate is beats per minute while that of the other 3 metrics is milliseconds.

Figure 4.

Estimating cardiac structure and function using patient-level ECG profiles. Bland-Altman plots comparing estimation of left ventricular mass index (LVMi; A) and mitral annulus medial e′ (B) values using ECG alone compared with echo-derived values. Number of studies depicted in comparison is shown. Orange, red, and blue dashed lines delineate the central 50%, 75%, and 95% of patients, as judged by difference between automated and manual measurements. The solid gray line indicates the median. Receiver operating characteristic (ROC) curves for classification models for left ventricular hypertrophy (C) and diastolic dysfunction (D). The area under the ROC curve (AUROC) is indicated. Variable importance for LVMi (E) and mitral annulus e′ (F) estimation models. The predictors most important for each model are highlighted with the relative importance indicated by the shading (white to blue). Informative intervals are depicted below the plot while lead-specific segments of the ECG are highlighted on the voltage trace.

Figure 5.

Detecting disease using patient-level ECG profiles. Receiver operating characteristic (ROC) curves (with area under the ROC curve [AUROC] indicated) for disease detection models for pulmonary arterial hypertension (PAH; A), hypertrophic cardiomyopathy (HCM; C), and cardiac amyloid (CA; E). Corresponding variable importance plots (B, D, F, and H) with coloring as in Figure 4. Violin plots indicating distribution of the top predictive feature in cases and controls for PAH (G), HCM (H), and CA (I). Precision-recall curves are depicted in Figure II in the Data Supplement.

Figure 6.

ECG-profile based models can be used to track changes in patient ECGs in pulmonary arterial hypertension (PAH). A, Scores for PAH detection for individual patients with measurements for 2 or more years. A median of all scores for each year is computed, and lines are drawn connecting scores for different years for each patient. The blue dashed line indicates a score threshold with 90% specificity and 80% sensitivity for a diagnosis of PAH. Purple, red, and yellow lines highlight score trajectories for 3 patients with dramatic variation in scores, crossing this threshold. B, Variation in ECG patterns for leads I and V1 from 2010 to 2017 for patient highlighted in purple in A. Over time, as the PAH score increases, the QRS axis swings progressively rightward (lead I), the P wave height grows, and the R′ wave in lead V1 increases in size.