Deep neural networks can predict mortality from 12-lead electrocardiogram voltage data

Discussion:

• AUC was 0.847 for one-year mortality, and the voltage-time DNN outperformed a machine-learning model using 39 clinically derived ECG features.The analysis used nearly 1.8 million ECGs collected from about 400,000 patients over 34 years.Source: Discussion · p. 6
• Median survival was over 3 times longer among patients whose ECGs were predicted to be alive than among those predicted to be dead.Source: Discussion · p. 6
• ECGs read as normal retained high mortality-prediction accuracy, and model features were not generally apparent to cardiologists even after 240 labeled true-positive and true-negative pairs.Source: Discussion · p. 6, Figure 3 · p. 6

Methods: ECG and patient data:

• 1.8 million resting ECGs acquired from 1984 to 2018 were retained after artifact and follow-up criteria, then resampled to 500 Hz.The initial institutional database contained 2.6 million standard 12-lead traces; retained recordings had 2.5-second traces for 12 leads and 10-second traces for V1, II, and V5.Source: Methods: ECG and patient data · p. 7
• Five-fold cross-validation kept each patient out of either the training or test sets, with 5% of training data reserved for validation.Patient status was assigned by cross-referencing clinical encounters and the institutional death index; outcomes were approximately balanced in the validation set.Source: Methods: ECG and patient data · p. 7
• Clinical reports were parsed into 30 diagnostic pattern classes and 9 continuous ECG measurements, and the tabular features were available for 75% of ECGs.An ECG was classified as abnormal when at least one diagnostic abnormality was flagged.Source: Methods: ECG and patient data · p. 7
• Only patients older than 18 years at ECG acquisition were included; age and survival time were calculated from the ECG date, and sex was extracted from EHR data.Survival was not assumed without a known living encounter; death records were regularly checked against death-index databases.Source: Methods: ECG and patient data · p. 7

• Five CNN branches grouped leads recorded concurrently over 2.5 seconds, with three long leads spanning 10 seconds, to account for changing trace morphology, particularly during arrhythmias.Source: Model development and evaluation · p. 7
• The branches used 1D convolution, batch normalization, ReLU and global average pooling, then fed concatenated outputs through six dense layers to a sigmoid output.The first four branches had four convolutional blocks each and the long-lead branch had six; Adam used a 1e-5 learning rate and batch size 2048.Source: Model development and evaluation · p. 7
• Replacing global average pooling with LSTM gave similar performance but substantially longer run times, so the final model used global average pooling.Source: Model development and evaluation · p. 8
• Five-fold cross-validation evaluated one-year mortality prediction; training stopped after 10 epochs without validation-loss improvement, with a 500-epoch maximum.Each fold was independently fit on five GPUs, and each epoch took about 10 minutes.Source: Model development and evaluation · pp. 8–9
• To compare voltage-time traces with clinically reported ECG measures, XGB used the same folds and DNNs were retrained on the roughly 75% of studies with measures; XGB outperformed random forest.Source: Model development and evaluation · p. 9

Survival analysis:

• Kaplan–Meier survival analyses stratified follow-up by DNN prediction, using a 0.5 likelihood threshold and censoring at the most recent encounter.The analysis used available follow-up data.Source: Survival analysis · p. 9
• Cox models regressed mortality on predicted alive/dead class separately in normal and abnormal ECG subsets, reporting mean estimates and 95% confidence bounds across five folds.Survival analyses used lifelines version 0.18.6 in Python.Source: Survival analysis · p. 9