Design and Development of a Machine Learning Model for Detection of Serum Potassium using Electrocardiograms

Abstract

Chronic kidney disease (CKD), one of the fastest-growing causes of death worldwide, newlinesignificantly increases the risk of hyperkalemia (elevated potassium levels). If left undetected, newlinehyperkalemia can lead to life-threatening heart complications. The current method of detecting this newlinepotassium imbalance involves frequent blood tests, which are invasive, time-consuming, and newlineimpractical for CKD patients requiring continuous monitoring, such as before and after dialysis newlinesessions. This research aims to develop a non-invasive method for detecting serum potassium levels newlineusing electrocardiogram (ECG) signals. Electrolytes such as sodium, potassium, magnesium, newlinecalcium, and chloride are essential for maintaining fluid balance, nerve transmission, and muscle newlinefunction. Imbalances in these electrolytes can result in serious health conditions, including newlinearrhythmias, gastrointestinal issues, kidney dysfunction, and sudden cardiac arrest. This study newlineadopts machine learning and deep learning techniques to analyse ECG signals for detecting newlinepotassium imbalances non-invasively. Various feature extraction methods including statistical, newlinemorphological, empirical mode decomposition (EMD), and automatic feature extraction was newlineexplored to identify the most effective approach. By comparing multiple models and feature newlineextraction techniques, this research aims to develop an accurate predictive model that offers early newlinewarning signs of hyperkalemia, potentially preventing life-threatening cardiac events. newlineIn this research data was extracted from the MIMIC-IV ECG matched subset and the newlineMIMIC-IV Clinical database(laboratory test reports data), focusing on cases where the time newlinedifference between ECG and potassium tests is less than four hours to ensure temporal alignment newlinewith the clinical condition of hyperkalemia. After pre-processing, 1503 patients data were newlineselected, with each ECG recording comprising 12 leads of 10-second duration, sampled at 500 Hz. newlineThe research compared five approaches to detecting hyperkalemia