In this paper a time-varying autoregressive (AR) modeling technique was used to model Ventricular Arrhythmias (VA) recorded from isolated rabbit hearts. In parametric modeling of such rhythms, the apparent repetition rate becomes the dominant component in the frequency estimate. Thus, smaller changes in the frequency content are masked by the depolarization rate of the cardiac cell. Instead of investigating the frequency location of the dominant poles, a bandwidth estimate was employed. This estimate better describes the energy distribution associated with the dominant poles of the AR-model. Bipolar ElectroCardioGram (ECG) recorded from hearts under different conditions were investigated. Typically, changes in the relative bandwidth of 100% could be observed for recordings having the frequency location estimate of the dominant poles displaced by only 5%. In addition, recordings of hearts which spontaneously went from one type of to another were analyzed. In all cases the bandwidth measure was superior to the frequency measure, when used as a feature to detect changes of various nature.