Intelligent Detection of Hypokalemia Based on 12-Lead ECG Using Two-stream Deep Learning Model

Hypokalemia is one of the most common electrolyte disorders in clinic. The detection of hypokalemia mainly depends on the detection of serum potassium concentration. Previous studies have shown that with the decrease of serum potassium ion concentration, ECG will show corresponding characteristics. In this paper, 12-lead ECG is used for intelligent detection of hypokalemia. After six artificial features based on ECG are extracted, a two-stream deep learning model is trained by using these features and 12-lead ECG to detect hypokalemia. The AVC of the two-stream model on the verification set is 0.84, and the AVC on the test set is 0.82. After taking the best working point, on the verification set, the sensitivity is 81.45%, the specificity is 74.21 %, and the recognition accuracy is 77.82%, while on the test set, the sensitivity is 77.54%, the specificity is 74.28%, and the recognition accuracy is 75.91 %. The results show that these time-domain features can significantly improve the recognition accuracy of hypokalemia.