Abstract:
Sudden, drastic changes in blood ion concentrations are common and clinically silent in patients with renal disease, from which around 42 million people in the US suffer. While otherwise healthy patients can tolerate mild hyperkalemia (i.e. high serum potassium), this condition can be life-threatening in patients suffering from cardiac disease. A non-invasive, robust, unobtrusive method for estimation of potassium and calcium concentrations in blood serum would be an important clinical advancement, providing basis for continuous monitoring an early warning for high-risk patients. During hemodialysis, patients underwent a high-resolution 12-lead electrocardiogram (ECG) recording and had 1-7 extracorporeal blood tests taken. Such data from 127 sessions of overall 42 patients was used as basis for this work. Signal processing methods were applied for signal-to-noise Ratio (SNR) improvement, followed by an averaging procedure condensing the continuous signal down to single beat ECG templates. Each template was then reduced to 14 features, whose correlations with the targets were positively assessed in previous works. Two novel features relating to the curvature of the T wave were also included. For the purpose of feature extraction a novel algorithm for finding onset and offset of the T wave was developed, which outperformed existing algorithms in a test scenario with artificially generated ECG signals. Initially, a global method for estimating potassium and calcium based on the features was developed. 112 different parametrizations of shallow neural network were tested, which all led to unsatisfactory estimation resolution.Features that exhibited the best linear dependencies with the targets out of the 14 initially chosen features were selected. Using these features and applying a 2-point calibration for each session, a simple linear model was set up, which delivered an estimation error of 0.09 mmol/l ± 0.54 mmol/l (mean ± standard deviation) for potassium and −0.01 mmol/l ± 0.12 mmol/l for calcium, after a 10-fold cross-validation. When the calibration procedure was done only with the information of the first session for each patient the estimation results were −0.2 mmol/l ± 0.68 mmol/l for potassium and −0.00 mmol/l ± 0.14 mmol/l for calcium. The methods for estimation of potassium and calcium blood ion concentration using clinical ECG data could not achieve the results of studies using simulated data. Nevertheless, the novel method presented in this work, featuring a patient-specific calibration, performs up to par with state of the art methods, exhibiting acceptable accuracy, making it clinically relevant for first-level remote monitoring of high-risk patients.