Prediction of hyperkalemia in dogs from electrocardiographic parameters using an artificial neural network

Objective: To predict severe hyperkalemia from single electrocardiogram (EC G) tracings. Methods: Ten conditioned dogs each underwent this protocol thr ee times: Under isoflurane anesthesia, 2 mEq/kg/hr of potassium chloride wa s given intravenously until P-waves were absent from the ECG and ventricula r rates decreased greater than or equal to 20% in less than or equal to5 mi nutes. Serum potassium levels (K+) were measured at regular intervals with concurrent digital storage of lead II of the surface EGG. A three-layer art ificial neural network with four hidden nodes was trained to predict K+ fro m 15 separate elements of corresponding ECG data. Data were divided into a training set and a test set. Sensitivity, specificity, and diagnostic accur acy for recognizing hyperkalemia were calculated for the test set based on a prospectively defined K+ = 7.5. Results: The model produced data for 189 events; 139 were placed in the training set and 50 in the test set. The tes t set had 37 potassium levels at or above 7.5 mmol/L. The neural network ha d a sensitivity of 89% (95% CI = 75% to 97%) and a specificity of 77% (95% CI = 46% to 95%) in recognizing these. The positive likelihood ratio was 3. 87. Overall accuracy of this model was 86% (95% CI = 73% to 94%). Mean (+/- SD) difference between predicted and actual Kf values was 0.4 +/- 2.0 (95% CI = -0.2 to 1.0). Conclusions: An artificial neural network can accuratel y diagnose experimental hyperkalemia using ECG parameters. Further work cou ld potentially demonstrate its usefulness in bedside diagnosis of human sub jects.