Wavelet neural networks combine the advantages of fast wavelet analysis and adaptive network optimization. They receive widespread attention for physiological signal interpretation. The infrastructures of current wavelet neural networks are either loosely associated or intrinsically synthesized. The former systems are advantageous in flexible structure, while the latter ones are oriented to global optimization. In this study we propose a new discriminant wavelet modeling by incorporating the famous method of Fisher's Linear Discrimination. It is then possible to construct a series of linear discriminant wavelet networks that inherit flexible infrastructure but achieve global optimization. Experiments on a well-known benchmark database effectively support this novel scheme for wavelet neural networks.