Predicting hyperkalemia by a two-staged artificial neural network

Hyperkalemia-induced arrhythmias are commonly observed in emergency medicine. Clinically, hyperkalemia is verified through laboratory tests, which takes 20 minutes or more to generate results. To increase the efficiency of treatment, the identification and classification of hyperkalemia directly from electrocardiogram are crucial. In this study, we developed a two-stage artificial neural network (ANN) to classify the disorder of electrolyte from 20 normal and 20 moderate hyperkalemic (5.4 - 7.4 mmole/L) individuals. The first stage of ANN, consisting of two multi-layer back-propagation ANNs, was fed with selected features extracted directly from 12-lead electrocardiograms. The second stage was composed of a two-layer back-propagation ANN receiving input from the outcomes of previous stage. A total of 17 features were selected, among which 12 features including T wave amplitude and duration from V1to V6 leads served as the inputs of the first ANN in the first stage and the other 5 features including P wave amplitude and duration, QRS duration, PR interval, averaged RR interval from lead II served as the inputs of the second ANN in the first stage. By training the two-stage ANN with 30 normal and 30 hyperkalemic cases, the results indicated that the accuracy is 62.5% with sensitivity 12/20. In conclusion, the algorithm developed in this study can predict hyperkalemia more efficiently compared to the prediction by experienced clinicians whose accuracy is 50 % with sensitivity 7/20. It can be expected that the sensitivity and the accuracy will be promoted once the training data increase.