Detecting Digoxin Toxicity by Artificial Intelligence-Assisted Electrocardiography

Abstract

Although digoxin is important in heart rate control, the utilization of digoxin is declining due to its narrow therapeutic window. Misdiagnosis or delayed diagnosis of digoxin toxicity is common due to the lack of awareness and the time-consuming laboratory work that is involved. Electrocardiography (ECG) may be able to detect potential digoxin toxicity based on characteristic presentations. Our study attempted to develop a deep learning model to detect digoxin toxicity based on ECG manifestations. This study included 61 ECGs from patients with digoxin toxicity and 177,066 ECGs from patients in the emergency room from November 2011 to February 2019. The deep learning algorithm was trained using approximately 80% of ECGs. The other 20% of ECGs were used to validate the performance of the Artificial Intelligence (AI) system and to conduct a human-machine competition. Area under the receiver operating characteristic curve (AUC), sensitivity, and specificity were used to evaluate the performance of ECG interpretation between humans and our deep learning system. The AUCs of our deep learning system for identifying digoxin toxicity were 0.912 and 0.929 in the validation cohort and the human-machine competition, respectively, which reached 84.6% of sensitivity and 94.6% of specificity. Interestingly, the deep learning system using only lead I (AUC = 0.960) was not worse than using complete 12 leads (0.912). Stratified analysis showed that our deep learning system was more applicable to patients with heart failure (HF) and without atrial fibrillation (AF) than those without HF and with AF. Our ECG-based deep learning system provides a high-accuracy, economical, rapid, and accessible way to detect digoxin toxicity, which can be applied as a promising decision supportive system for diagnosing digoxin toxicity in clinical practice.

Keywords: artificial intelligence; deep learning algorithm; digoxin toxicity; electrocardiogram.

Figure 1

Performance comparison in digoxin toxicity recognition in the human-machine competition. The receiver operating characteristic curves (ROC curves) were made by the predictions of the deep learning model in the validation cohort (AI-all) and the human-machine competition (AI-sub), respectively. The red and blue points represent visiting staff members and residents, respectively. The triangle and square marks represent the emergency physicians and cardiologists, respectively. The areas under curve (AUCs) were 0.912 and 0.929 in the validation cohort and the human-machine competition, respectively. The sensitivity was 84.6%, the specificity was 96.6%, and the probability was 33.4% in the validation cohort. The positive predictive value in the validation cohort was 1.98%, which provided an F1 score of 3.87%. prob = probalility; sens = sensitivity; spec = specificity.

Figure 2

Selected ECGs (electrocardiographies) of consistent and inconsistent assessments given by the deep learning model and the human experts. (A) The ECG was correctly identified as high risk of digoxin toxicity by both human experts and AI machine. (B) The ECG was correctly identified as low risk of digoxin toxicity by AI machine, but not by human experts. (C) The ECG was correctly identified as high risk of digoxin toxicity by human experts, but not by AI machine. (D) The ECG has been misdiagnosed by both human experts and AI machine.

Figure 3

Lead-specific ROC (Receiver operating characteristic) curves in the validation cohort. ROC curves with specificity on the x-axis and sensitivity on the y-axis. The cases were defined as the digoxin toxicity ECGs, and the controls were defined as the normal ECGs.

Figure 4

ROC curves under significant stratified variables. (A) Better prediction of the AI model may be noted with a larger AUC in patients known to have heart failure; (B) Worse prediction of the AI model with smaller AUC for patients with atrial fibrillation.