Discretization of continuous ECG based risk metrics using asymmetric and warped entropy measures

Abstract

We investigate several entropy based approaches to finding cut points for discretizing continuous ECG-based risk metrics. We describe two existing approaches, Shannon entropy and asymmetric entropy, and one new approach, warped entropy. The approaches are used to find cut points for the end point of cardiovascular death for three risk metrics: heart rate variability (HRV LF-HF), morphological variability (MV) and deceleration capacity (DC). When trained on multiple instances of training set containing 2813 patients, warped entropy yielded the most robust cut-offs. The performance of the cutoffs obtained using warped entropy from the training sets was compared with those in the literature using a Naive Bayes classifier on corresponding test sets. Each test set contained 1406 patients. The resulting classifier resulted in a significantly (p<;0.05) improved recall rate at the expense of a lower precision.