Author + information
- Chang H. Kim and
- Sadeer Al-Kindi
While various electrocardiogram (ECG) characteristics have been associated with increased cardiovascular mortality (CVM), utility of aggregated, machine-derived ECG measurements for prediction of CVM remains unknown.
Using data from 8,432 participants in the third National Health and Nutrition Examination Survey (NHANES III, 1988–1992, linked with mortality data), we applied machine learning (ML) techniques to 62 ECG features and demographic data to predict 10-year CVM. Data were split 80:20 for training and test, with multiple imputation for missing features and synthetic data to address class imbalance in the training set. ML models (logistic regression (LR), neural network, gradient boosting machine (GBM), support vector machine, and random forest) were fitted using 10-fold cross validation.
A total of 720 cardiovascular deaths were observed (event rate: 8.5%). Models trained on the original data set showed high accuracy but poor sensitivity (<0.10). Synthetic data allowed for improved sensitivity across all models. LR showed balanced accuracy (0.74) and sensitivity (0.76), while GBM performed best in terms of overall accuracy (0.88) and specificity (0.94). In comparison to LR, other ML models showed worse sensitivity but better specificity (Figure).
ML algorithms can predict 10-year CVM based on demographic and automated ECG measurements. Varying model characteristics may be useful for risk stratification in different clinical scenarios.
Posters Hall_Hall A
Monday, March 30, 2020, 9:45 a.m.-10:30 a.m.
Session Title: Prevention: Clinical 7
Abstract Category: 32. Prevention: Clinical
Presentation Number: 1420-084
- 2020 American College of Cardiology Foundation