Source-linked AI summary
Deep neural networks can predict mortality from 12-lead electrocardiogram voltage data
Sushravya Raghunath, Alvaro E. Ulloa Cerna, Linyuan Jing, David P. vanMaanen, Joshua Stough, Dustin N. Hartzel, Joseph B. Leader, H. Lester Kirchner, Christopher W. Good, Aalpen A. Patel, Brian P. Delisle, Amro Alsaid, Dominik Beer, Christopher M. Haggerty, Brandon K. Fornwalt
TL;DR
Can deep neural networks predict one-year mortality directly from 12-lead ECG voltage-time traces, including ECGs interpreted as normal? The paper evaluates this approach on a large longitudinal ECG dataset and reports strong mortality prediction, with an average AUC of 0.85 and high performance among normal ECGs.
Problem
The study asks whether deep neural networks can extract prognostic information about one-year mortality directly from 12-lead ECG voltage-time traces.
Method
The study trains and five-fold cross-validates a convolutional neural network on 1.8 million resting 12-lead ECGs to predict one-year mortality.
Results
AUC was 0.847 for one-year mortality prediction, remained high for physician-interpreted normal ECGs, and predicted groups differed in longer-term survival.
Takeaways & Limitations
Deep learning can provide prognostic information from resting 12-lead ECGs, including cases interpreted as normal by cardiologists.
Abstract
from arXiv · showhide
The electrocardiogram (ECG) is a widely-used medical test, typically consisting of 12 voltage versus time traces collected from surface recordings over the heart. Here we hypothesize that a deep neural network can predict an important future clinical event (one-year all-cause mortality) from ECG voltage-time traces. We show good performance for predicting one-year mortality with an average AUC of 0.85 from a model cross-validated on 1,775,926 12-lead resting ECGs, that were collected over a 34-year period in a large regional health system. Even within the large subset of ECGs interpreted as 'normal' by a physician (n=297,548), the model performance to predict one-year mortality remained high (AUC=0.84), and Cox Proportional Hazard model revealed a hazard ratio of 6.6 (p<0.005) for the two predicted groups (dead vs alive one year after ECG) over a 30-year follow-up period. A blinded survey of three cardiologists suggested that the patterns captured by the model were generally not visually apparent to cardiologists even after being shown 240 paired examples of labeled true positives (dead) and true negatives (alive). In summary, deep learning can add significant prognostic information to the interpretation of 12-lead resting ECGs, even in cases that are interpreted as 'normal' by physicians.
Discussion:
Using nearly 1.8 million ECGs collected over 34 years, the study showed that deep neural networks can predict one-year mortality directly from 12-lead voltage-time traces. The model also provided longer-term prognostic information and identified patterns beyond extensive clinical measurements.
- Nearly 1.8 million ECGs from ~400 thousand patients collected over 34 years were used to predict one-year mortality directly from 12-lead ECG voltage-time data.The study described this endpoint as highly clinically relevant.
- The voltage-time DNN outperformed a machine-learning model using 39 clinical numerical measurements and diagnostic patterns, suggesting novel prognostic patterns.
- Accuracy for three cardiologists identifying the true-positive ECG was assessed using 401 paired normal ECGs and then an independent labeled set of 240 ECGs.All ECG pairs were matched for age and sex.
- AUC 0.847 was achieved for one-year mortality prediction, while median survival was over 3 times longer for patients predicted alive than dead.
Methods: ECG and patient data:
The study analyzed artifact-free resting 12-lead ECG voltage-time traces from a large institutional database collected between 1984 and 2018, linked to follow-up or death data. Clinical reports, age, sex, and survival information were extracted for characterization and outcome ascertainment, under institutional review board approval with waived consent.
- ECG and patient data: 2.6 million standard 12-lead ECG traces were extracted from the institutional MUSE database, acquired between 1984 and 2018.The retained recordings had 2.5-second traces for 12 leads and 10-second traces for V1, II, and V5, without significant artifacts.
- ECG and patient data: Resting ECGs were retained only when follow-up lasted at least one year or death occurred within one year.Patients were included if their recordings were artifact-free and met the specified trace-duration requirements.
- ECG and patient data: The institutional review board approved the study with a waiver of consent.The study used clinical ECG and electronic health record data.
- ECG and patient data: Clinical ECG reports were parsed into 30 diagnostic pattern classes and 9 continuous ECG measurements.The tabular ECG features were available for 75% of all available ECGs.
- ECG and patient data: Only patients older than 18 years at ECG acquisition were included, with age, sex, and survival time referenced to the acquisition date.Death data were checked regularly against death-index databases, but survival was not assumed without a known living encounter.
Model development and evaluation:
The study developed a five-branch convolutional neural network that processed concurrent ECG lead groups and evaluated 1-year mortality prediction using 5-fold cross-validation. Global Average Pooling was retained because it matched LSTM performance with shorter runtimes.
- Architecture: The network used 5 branches, each receiving 3 ECG leads acquired concurrently over 2.5 seconds.The branches represented [I, II, III], [aVR, aVL, aVF], [V1, V2, V3], [V4, V5, V6], and [V1-long, II-long, V5-long].
- Architecture: The convolutional branches applied one-dimensional convolution, batch normalization, ReLU activations, and Global Average Pooling before dense-layer concatenation.The first four branches used 4 convolutional blocks, while the last used 6, followed by six dense layers ending in a sigmoid unit.
- Architecture: LSTM replacement produced similar performance but significantly longer runtimes, so the final model used Global Average Pooling layers.
- Evaluation: The study cross-validated an XGB classifier on clinically reported ECG measures using the same folds to compare prognostic efficacy with voltage-time-trace models.For paired comparison, the DNNs were retrained on the approximately 75% of studies with available ECG measures.
Survival analysis:
Survival analyses stratified ECGs by DNN predictions using a 0.5 likelihood threshold and modeled mortality with Cox proportional hazards in normal and abnormal ECG subsets.
- Survival analysis:: Kaplan–Meier survival analysis stratified patients by DNN model prediction at a likelihood threshold of 0.5.Follow-up data were censored based on the most recent encounter.
- Survival analysis:: Cox proportional hazards models regressed mortality on DNN-predicted alive versus dead classifications separately for normal and abnormal ECG subsets.