The works presented in this dissertation originate from our years of research on cardiovascular health monitoring at home. Different prototyping systems have been developed in our group for cardiovascular health monitoring. However, it is found in our practice that computerized interpretation of the recorded cardiovascular physiological signals is far more challenging than expected. The biggest challenge comes from the intrinsic physiological signal variability due to pathophysiological artifacts, instrumental inaccuracy and operational inconsistency. All of them alter the morphologies and rhythms of cardiovascular physiological signal substantially. At the same time, there is yet no unanimous solution to discriminate and eliminate abovementioned variability in cardiovascular physiological signals. This dissertation is focused on integrating hard computing (HC) and soft computing (SC) for computerized physiological signal interpretation. Neural networks are chosen as the delegate of SC for adaptive clustering and supervised classification. Among various HC paradigms, wavelet transform is a well-established technology for unified time-frequency analysis. To attack physiological signal variability, we are interested in the essential components by means of wavelet analysis. Those components are supposed resistant to noises and artifacts. Nevertheless, there is yet no unanimous criterion for the choice of essential wavelet components. Two criteria were examined and evaluated in this dissertation for adaptive wavelet modeling. The first one is oriented to energy maximization. It attempts to select out those strong wavelet components. Actually, many popular schemes of adaptive wavelet modeling, such as wavelet scale maxima, relative wavelet energies and regional wavelet entropies, exactly follow this energy-oriented criterion. The second criterion directing adaptive wavelet modeling is oriented to morphological similarity. In essence, it is desired ii to optimize wavelet modeling to approximate the original physiological signals accurately with a compact representation. The refined wavelet components are supposed optimal to the essential signal components, meanwhile are resistant to noises and artifacts. The paradigms of wavelet shrinkage, matching pursuits and wavelet regression networks obey such morphology-oriented criterion for adaptive wavelet modeling. All of them were further examined in comparison with the fully integrated wavelet networks for adaptive clustering and supervised classification of cardiovascular physiological signals. In general, a wavelet network replaces the neuronal transfer functions in conventional neural networks by wavelet basis functions. Hence the optimization of wavelet modeling and neural networks can be unified as a whole. In practice, we found that wavelet networks suffer from many challenging issues in system architecture, network initialization, optimal evolution, and so forth. On the contrary, the modular integration of wavelet analysis and neural networks owns structural simplicity and implemental flexibility. Nevertheless, the morphology- or energy-oriented paradigm of adaptive wavelet modeling keeps blind to the subsequent clustering or classification. In other words, their performance can not be guaranteed as an overall system for physiological signal interpretation. Therefore we proposed two novel wavelet modeling strategies in this dissertation. The first scheme, termed as principal wavelet modeling (PWM), incorporates the theories and methods of principal component analysis into adaptive wavelet modeling. The choice of wavelet components are based on neither energy maximization nor morphological similarity, but their clustering distributions in the unified time-frequency space. Theoretically speaking, the PWM is optimal for adaptive clustering of cardiovascular physiological signals. On the other hand, we took advantages of the theories and methods of linear discriminant analysis for the second paradigm of adaptive wavelet modeling, termed as discriminant wavelet modeling (DWM). The adaptive wavelet modeling is hereby oriented to classification optimization directly. The performance was validated carefully in both small- and large-scale benchmark databases. In terms of ECG beat clustering, the wavelet neural network using novel PWM and FCM achieved the accuracy 64.7%, which was better than those morphology-oriented HBF (57.5%) and MP (53.4%). Coming to ECG classification, the wavelet neural network based on the novel DWM and kNN (error rate 10.34±0.35%) was fairly good against the energy-oriented WSM (13.31±0.30%) and the morphology-oriented MP (18.75±0.67%), but worse than the morphology-oriented HBF (8.45 ± 0.26%). As a conclusion, neither morphology- nor energy-oriented adaptive wavelet modeling is able to guarantee the global performance for optimal clustering and classification. In contrast, the novel PWM and DWM are robust against physiological signal variability, thus effective for computerized physiological signal interpretation