Intelligent Detection of Arrhythmia Episodes in Dialysis Patients

This work discusses the design of an automatic detector of arrhythmia episodes in patients submitted to dialysis. The system aims to operate on portable devices in real-time, allowing a faster response of healthcare workers to possible intercurrence episodes. The detection is based on processing short windows of samples extracted from the electrocardiogram signal around the R-wave peak in raw format. A comprehensive study evaluating several classification techniques and class-imbalance strategies is conducted based on the MIT-BIH Arrhythmia Database. Besides, a new procedure for tuning the sample window length based on an experimental feature importance cumulative distribution is proposed. Results show that a Random Forest classifier, trained with minority class oversampling, is cost-effective regarding complexity and computational cost, achieving an accuracy of 98.7% for windows sizes as small as 105 samples.