Limited added value of ECG-based mortality prediction in COVID-19 patients using machine learning

ECG-based machine learning models were able to identify predictors of mortality in patients with COVID-19 in the first 72 hours after admission. The added value of prediction models based on ECG features was present but limited [1]. 

Dr Hidde Bleijendaal (Amsterdam University Medical Center, the Netherlands) presented a study aimed to evaluate whether ECG-based machine learning models can predict all-cause, in-hospital mortality in COVID-19 patients and to identify ECG features associated with mortality [1]. Included were 882 patients admitted with COVID-19 in 7 different Dutch hospitals. Raw-format 12-lead ECGs recorded after admission (<72 hours) were collected, manually assessed, and annotated using pre-defined ECG features. Using data from 5 of the 7 centres (n=634), 2 mortality prediction models were developed. The first prediction model was a multivariate logistic regression (LASSO) model adding manually extracted ECG features and was used to identify ECG features associated with mortality. The second was a deep learning model (DNN) developed using raw-format ECGs, age, and sex. To implement transfer learning, a pre-trained model (large dataset, different task) fine-tuned for current prediction task was used. A baseline model was created using only age and sex to evaluate the added value of ECG. Data from 2 other centres (n=248) were used for external validation.

Performance of both prediction models was similar, with a mean area under the ROC of 0.76 (95% CI 0.68–0.82) for the LASSO model (sensitivity 0.86, specificity 0.54) and 0.77 (95% CI 0.70-0.83) for the DNN in the external validation cohort (sensitivity 0.86, specificity 0.57). Respective results for the baseline model were slightly lower or similar: AUC was 0.76 (95% CI 0.68–0.82), sensitivity was 0.87, and specificity was 0.49. After adjustment for age and sex, increased ventricular rate, right bundle branch block, ST-depression, and low QRS voltages remained as significant predictors for mortality in COVID-19 patients.

Dr Bleijendaal concluded, “Predication of mortality in this dataset in COVID-19 patients is mostly based on age and sex. However, by adding ECG data we did improve the AUC slightly. The added value of the ECG seems to be present but is limited.”

1. 
   
   1. Bleijendaal H. Electrocardiogram-based mortality prediction in patients with COVID-19 using machine learning. EHRA 2021 Congress, 23-25 April.

 

Copyright ©2021 Medicom Medical Publishers



Posted on 16 June, 202124 June, 2021