A new class of prediction-based independent training (PBIT) networks for temporal patterns classification is propose. Our approach combines the universal NN approximator and TDNN. The input vectors of PBIT are consecutively created by the time-delayed segment of a pattern. The network is robust since all input segments contribute in equal share to the classification result. To demonstrate the feasibility of PBIT, extensive simulations on ECG classification have been conducted. They have all relatively good generalization accuracy. IIR filters can be adopted as preprocessors to extract information out of the time sequence so that smaller networks will suffice. For a comparative study, we have included another independent training classifier -hidden Markov model (HMM) - for temporal patterns. We have also included in the study, some mutually training (MT) models, e.g. the (static) decision-based neural network (DBNN). Hybrid IT and MT techniques are also proposed to further improve classification accuracy.