Mahalanobis-Taguchi system (MTS) was proposed by Genichi Taguchi for the study of multidimensional systems and has been effectively applied in many fields. The traditional MTS method focuses on the dimensionality reduction of multidimensional systems and the measurement of abnormal degree of observations. In order to further improve the effectiveness of the MTS method, many studies related to the MTS method have been conducted, including the construction of Mahalanobis space, threshold determination, selection of useful variables, and so on. However, only four studies have been conducted to analyze the potential causes of abnormal observations diagnosed by the MTS method. Meanwhile, these studies ignore the counter-correlations between variables, which will make it impossible to find the potential causes of some abnormal observations and therefore unable to effectively improve these abnormal observations. This paper aims to improve the traditional MTS method and then propose an effective method to analyze the potential causes of abnormal observations diagnosed by the improved MTS method. First, this paper proposes to divide the traditional MTS method into two phases. In phase one, the scaled Mahalanobis distance function is still used to measure the abnormal degree of observation, while in phase two the weighted Mahalanobis distance function should be used to make the diagnosis more reasonable and accurate. Second, a new decomposition method of weighted Mahalanobis distance is proposed to identify potential causes of abnormal observations based on Mason-Young-Tracy (MYT) decomposition method. The method focuses not only on whether individual variables are outside their reference ranges, but also on whether the linear relationships between variables are consistent with the reference group, which makes the analysis of abnormal causes more comprehensive and reliable. Finally, the blood viscosity diagnosis system in a hospital was improved using the improved MTS method. The results showed that the variables in the diagnosis system decreased by 37.5%, and the differentiation between the abnormal conditions and the health group increased by 102%. At the same time, the potential causes of an abnormal observation were identified based on the MYT decomposition method of weighted Mahalanobis distance. The above results verified the feasibility and effectiveness of the method in this paper.