The thesis explores methods of emotional state estimation based on data mining of physiological signals, such as electrocardiogram (ECG), skin conductance, skin temperature and respiration, methods for finding the dominant physiological features for emotional state estimation, and the general architecture of the corresponding systems for the emotional state estimation that adapt to the individual in real-time. The major architectural components of the emotional state estimator are identified, which refer to the analysis before starting the session, estimation and adaptation in real-time during the session, and the required databases. Further architectural decomposition describes in detail the steps of data selection for estimator learning, raw physiological signal processing, physiological feature calculation, physiological feature reduction, estimator supervised learning, as well as the databases of all the participants, of the current participant, of the stored parameters of the estimator etc. Time constraints for the estimation and estimator adaptation in real time are explained. As an example of the analysis needed to meet the imposed time constraints, time analysis and parallelization of tasks that constitute real-time emotional state estimation and adaptation are conducted. The problem of finding the dominant physiological features for emotional state estimation is formulated, and associated methods for finding such features are proposed. The accuracy of emotional state estimators developed in line with the proposed architecture is experimentally tested on the problem of discrete emotions classification based on the systematic combination of various methods of feature reduction and classification. The results of applying the methods for dominant physiological feature identification are also analyzed. Estimation of valence and arousal is conducted with the use of multilayer perceptron with SFFS feature reduction, a combination that exhibited the highest classification accuracy for discrete emotions.