Two-Stage Classifier for Resource Constrained On-Board Cardiac Arrhythmia Detection

Abstract

Early automatic detection of cardiac arrhythmia in an ambulatory health monitoring setup can reduce the mortality rates. In the existing approaches on multiclass arrhythmia detection (MCAD) techniques, adopting either time-domain morphological features (TDMFs), fails to extract hidden pattern from the single-lead electrocardiogram (ECG) resulting in poor accuracy (Ac), or adopting deep-learning (DL) techniques becomes too complex or heavy in architecture, thus hindering their easy implementation in low resource architecture for wearable applications. This work presents a judicious hybridization of both approaches in an end-to-end MCAD using two-stage classifier to achieve higher Ac and low latency while detecting 15 classes of cardiac rhythms [14 types of arrhythmia beats or normal sinus rhythms (NSRs)]. The memory resource saving is achieved through the beat skeletonization process and two-stage feature minimization. The stage#1 classifier uses 14 TDMF features fed to a random forest (RF) regressor and provides probability values for NSR, atrial, or ventricular conduction problem. The stage#2 classifier, another RF, accepts stacked autoencoder (SAE) features, TDMF features, and output of stage#1 to achieve a blind test Ac of 99.21% using only 13 features while evaluated with full Massachusetts Institute of Technology-Beth Israel (MIT-BIH) arrhythmia database (mitdb) records. A stand-alone on-board application of the two-stage MCAD using ARMv6 single-core microcontroller yields a detection latency of 782 ms and consumes 390 kB of memory per cardiac cycle, thus proving a feasible solution for wearable cardiac monitors. It also provides the competitive performance with the State of the art (SOTA) methods and outperforms the limited available research on hardware implementation in terms of number of classes handled and Ac.