Deep Learning Decodes Latent ECG Signatures of Stress Cardiomyopathy
Ryu, J.; Harris, C.; Zhang, C.; Gong, K.; Stevens, R. D.
Show abstract
BackgroundStress cardiomyopathy (SCM) shares features with acute myocardial infarction (AMI) which may lead to misdiagnosis and misaligned management decisions. We hypothesize that features derived from the 12-lead ECG can identify cases of SCM and differentiate them from AMI. MethodsUsing data from a large registry of critically ill patients, we trained a deep learning algorithm to perform two classification tasks: (1) SCM vs. Non-SCM and (2) SCM vs. AMI. Model training was accomplished with 3 different sets of input features: raw ECG waveforms, clinical features from the electronic health record (EHR), and a fusion model combining ECG with clinical data. SHapley Additive exPlanations (SHAP) analysis was performed to identify the most influential predictive features. ResultsAmong 71,479 patients admitted to ICU, 349 (0.48%) were diagnosed with SCM while 4,507 (6.31%) had AMI. The clinical-ECG fusion model achieved best performance, with area under the precision-recall curve (AUPRC) values of 0.191 (0.146-0.243) for SCM vs. Non-SCM and 0.430 (0.371-0.488) for SCM vs. AMI, outperforming baseline AUPRCs of 0.0048 and 0.0625, respectively. The fusion model outperformed both the waveform model (p < 0.001) and the EHR model (p < 0.001) for both SCM vs. Non-SCM and SCM vs. AMI. The waveform model achieved AUPRC values of 0.089 (0.062-0.125) for SCM vs. Non-SCM and 0.309 (0.257-0.374) for SCM vs. AMI. There was no statistically significant difference in performance between the waveform model and the EHR model for both tasks. SHAP analysis highlighted female gender as well as congestive heart failure (CHF) and hypertension as the features most influential in predicting SCM. ConclusionFindings indicate that ECG waveforms contain latent information which supports the detection of SCM in patients admitted to the ICU. ECG-based deep learning screening could enable early identification and treatment of SCM and might be particularly valuable in resource-constrained environments.
Matching journals
The top 5 journals account for 50% of the predicted probability mass.
Similar papers in this journal
- QRS detection in single-lead, telehealth electrocardiogram signals: benchmarking open-source algorithms 96%
- Detecting QT prolongation From a Single-lead ECG With Deep Learning 95%
- A recurrent neural network and parallel hidden Markov model algorithm to segment and detect heart murmurs in phonocardiograms 94%
Similar papers in this journal
- Explaining Deep Neural Networks for Knowledge Discovery in Electrocardiogram Analysis 97%
- Advancing Cardiovascular Disease Diagnosis: A Robust ML Ecosystem Integrating Early Detection, Responsible AI Framework, and Causal Inference 94%
- Deep Learning-based Prediction of Early Cerebrovascular Events after Transcatheter Aortic Valve Replacement 94%
Similar papers in this journal
- Age Prediction From 12-lead Electrocardiograms Using Deep Learning: A Comparison of Four Models on a Contemporary, Freely Available Dataset 96%
- Classification of 12-lead ECGs: the PhysioNet/Computing in Cardiology Challenge 2020 94%
- Prospective validation of clinical deterioration predictive models prior to intensive care unit transfer among patients admitted to acute care cardiology wards 92%
Similar papers in this journal
- Algorithm for Predicting Valvular Heart Disease from Heart Sounds in an Unselected Cohort 96%
- Development and Validation of a Nomogram for Predicting Survival in Patients With Cardiogenic Shock 94%
- Deep learning-based end-to-end automated stenosis classification and localization on catheter coronary angiography 93%
"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.