Valvular heart diseases (VHDs) are an abnormal activity of the heart caused by a damage of one of the four heart valves. The impedance cardiography (ICG) is a non-invasive method employed to identify and classify the heart abnormalities. Despite its importance, there are not many works in scientific literature that use the ICG method in order to diagnose VHDs. Therefore, this paper deals with the ICG signal processing for the classification of normal (N) and various VHDs classes of heartbeats. In this work, six types of heartbeat classes of VHD are used, namely: aortic insufficiency (AOI), aortic stenosis (AOS), aortic disease (AOD), mitral disease (MD), mitral-aortic heart disease (MAOHD) and tricuspid insufficiency + mitral disease (TI + MD). The classification of these heartbeat classes is performed using a combination among statistical, morphological and spectral features. For each ICG heartbeat, the statistical features (median, mean, standard deviation, kurtosis, skewness, central moment and Shannon entropy) are computed from the first four intrinsic mode functions (IMFs) calculated using the empirical mode decomposition (EMD) technique. These statistical features are subjected to principal component analysis (PCA) to reduce the dimensionality. Then, the morphological features are extracted by calculating maximums and minimums of the peaks existing in the ICG heartbeat signal and determining the intervals which separate these peaks. Besides, the spectral features are calculated from the first three harmonics of each heartbeat ICG spectrum. From these three types of features, we selected the most significant of them by applying the analysis of variance (ANOVA) test. In order to obtain the best classification performance, three combination of the selected features are investigated and tested using the decision tree (DT), the random forest (RF) and the support vector machine (SVM) classifiers: i) spectral + morphological, ii) statistical + morphological, iii) statistical + morphological + spectral. The achieved results showed that the use of the third combination of features coupled with the RF classifier gave the highest overall accuracy of 96.34 %, average class-accuracy of 98.95 %, average specificity of 99.38 %, average positive predictivity of 96.73 % and average negative predictivity of 99.44 % using 10-fold cross validation. Thus, our developed method seems robust and very efficient for automatic detection of normal and VHDs ICG classes.