Kidney Week

Abstract: TH-OR069

ECG12Net: A Deep Learning Algorithm Capable of Suprahuman Detection of Hypokalemia and Hyperkalemia by Electrocardiography

Session Information

• Fluid and Electrolytes: Clinical Resesearch
  November 07, 2019 | Location: 144, Walter E. Washington Convention Center
  Abstract Time: 04:42 PM - 04:54 PM

Category: Fluid and Electrolytes

• 902 Fluid and Electrolytes: Clinical

Authors

• Lin, Shih-Hua P., Tri-Service General Hospital, Neihu, Taiwan

Shih-Hua P. Lin, MD



As a clinician, I am mostly involved in clinical diagnosis and management of acid-base and electrolyte disorders and collaborate with worldwide experts and laboratories. I also help with the molecular diagnosis for the patients with inherited renal tubular disorders such as renal tubular acidosis, Bartter’s syndrome, Gitelman’s syndrome, and nephrogenic diabetes insipidus, et al. I also do several animal studies, focusing on disease (Gitelman’s syndrome, Gordon’s syndrome and isolated proximal renal tubular acidosis)-causing knockin mice to explore the pathophysiology and rescue therapy, as well as global and kidney-specific knock-out mice in the upstream and downstream regulator of thiazide-sensitive sodium chloride co-transporter (NCC). I participated in KDIGO Expert Meeting for Gitelman’s syndrome and serves as an editorial board in JASN from 2018. Recently, I am interested in the application of deep learning model (DLM) to make EKG-based detection of severe dyskalemia.

• Sung, Chih-Chien, Tri-Service General Hospital, Neihu, Taiwan

Chih-Chien Sung,

• Lin, Chien-Ming, Tri-Service General Hospital, National Defense Medical Center, Taipei, Taiwan

Chien-Ming Lin,

• Hsu, Yu-Juei, Division of Nephrology, Tri-Service General Hospital, Taipei, Taiwan

Yu-Juei Hsu,

Background

Detection of dyskalemias (hypo- and hyperkalemia), important causes of sudden cardiac death, currently depends on laboratory tests. Since cardiac tissue is very sensitive to dyskalemias, electrocardiography (ECG) may be able to uncover clinically important dyskalemias before laboratory results. Our study aimed to develop a deep learning model, ECG12Net, to detect dyskalemias based on ECG presentation and to evaluate the logic and performance of this model.

Methods

Between May 2011 and December 2016, 66,321 ECG records with corresponding serum potassium (K⁺) concentrations were obtained from 40,180 patients admitted to the emergency department. ECG12Net is an 82-layer convolutional neural network, which estimates serum K⁺ concentration. Six clinicians (three emergency physicians and three cardiologists) participated a human-machine competition. We used sensitivity and specificity as evaluation measures to compare the performance of ECG12Net with these physicians.

Results

In a human-machine competition including 300 ECGs of different serum K⁺ concentrations, the area under curve in detecting hypo- and hyperkalemia by ECG12Net was 0.926 and 0.958, respectively, which was significantly better than that of our best clinicians. Moreover, the sensitivities and specificities of detecting hypokalemia and hyperkalemia were 96.7% and 83.3%, and 93.3% and 97.8%, respectively. In the test set including 13,222 ECGs, ECG12Net had the same performance with sensitivities for severe hypokalemia/hyperkalemia achieving 95.6% and 84.5%, respectively with the mean absolute error of 0.531. The specificities of detecting hypokalemia and hyperkalemia were 81.6% and 96.0%, respectively.

Conclusion

A deep learning model based on 12-lead ECG may help physicians to promptly recognize severe dyskalemias and thereby reduce cardiac events.