Development of Pediatric Hypokalemia Prediction Deep-Learning System Based on Wearable Single-Lead Electrocardiography in Real Time

초록

Background:Bartter syndrome (BS) is one of the most well-known hereditary tubular disorders, posing significant challenges for patients in maintaining electrolyte and fluid balance. In severe cases, muscle weakness, paralysis, arrhythmias, and even sudden death can occur, emphasizing the importance of continuous potassium level monitoring for appropriate management. While recent studies have explored deep learning models using 12-lead electrocardiograms to predict abnormal potassium levels, research on monitoring serum potassium levels in patients' home settings is lacking. Therefore, this study aims to develop a deep learning model for detecting hypokalemia using wearable devices' electrocardiogram data.Methods:This study employed oversampling of positive data and utilizing a generative adversarial neural network-based data augmentation process. A classification model is employed to verify the quality of data, ensuring the utilization of high-quality datasets. Subsequently, the study selects ResNet-50, MobileNet-v3, and EfficientNet-B3 models to compare their performance following dataset augmentation.Results:The performance of hypokalemia prediction models was evaluated by comparing ROC (Receiver Operating Characteristic) curve values. The best AUC (Area Under the Curve) of the hypokalemia detection model was 0.98.Conclusion:This study developed a deep learning model for hypokalemia detection using data measured by wearable devices, focusing on pediatric patients with BS. Continuous monitoring of potassium levels through wearable devices to predict abnormal potassium concentrations in the bloodstream is expected to aid in the healthcare management of patients with BS and support medical decision-making by healthcare professionals.