Robust Algorithmic Detection of Cardiac Pathologies from Short Periods of RR Data

Abstract

Numerous research efforts and clinical testing have confirmed validity of heart rate variability (HRV) analysis as one of the cardiac diagnostics modalities. Recently we have illustrated that building meta-indicators on the base of existing indicators from nonlinear dynamics (NLD) using boosting-like ensemble learning techniques could help to overcome one of the main restrictions of all NLD and linear indicators – requirement of long time series for stable calculation. We demonstrate universality of such meta-indicators and discuss operational details of their practical usage. We show that classifiers trained on short RR segments (down to several minutes) could achieve reasonable accuracy (classification rate ≈80-85% and higher). These indicators calculated from longer RR segments could be applicable for accurate diagnostics of the developed pathologies with classification rate approaching 100%. In addition, it is feasible to discover single “normal-abnormal” meta-classifier capable of detecting multiple abnormalities. Rare abnormalities and complex physiological states can be effectively classified by a new approach - ensemble decomposition learning (EDL).