The Visibility Graph algorithm has been proved to be a simple and efficient method to transform time series into complex network and has been widely used in time series analysis because it can inherit the dynamic characteristics of original time series in topological structure. Now, visibility graph analysis of univariate time series has driven to mature, however, most complex systems in real world are multi-dimensional, it is difficult for univariate analysis to describe the global characteristics when applied to multi-dimensional series. In this paper, a novel method for multivariate time series analysis is proposed. For patients with myocardial infarction and healthy subjects, the 12-lead electrocardiogram signals of each individual are considered as a multivariate time series, which is transformed into a multiplex visibility graph through visibility graph algorithm and then mapped to fully connected complex network. Each node of the network corresponds to a lead and the inter-layer mutual information between visibility graph of two leads represents the weight of edges. Due to the fully connected network of different groups show identical topological structure, the dynamic characteristics of different individuals cannot be uniquely represented. Therefore, we reconstruct the fully connected network according to inter-layer mutual information, when the value of inter-layer mutual information is less than the threshold we set, edge corresponding to the inter-layer mutual information is deleted. We extract average weighted degree and average weighted clustering coefficient of reconstructed networks so as to recognize the 12-lead ECG signals of healthy subjects and myocardial infarction patients. Moreover, multiscale weighted distribution entropy is also introduced to analyze relationship between the length of original time series and final recognition results. Due to higher average weighted degree and average weighted clustering coefficient of healthy subjects, their reconstructed networks show a more regular structure, higher complexity and connectivity, and could be distinguished from patients with myocardial infarction, whose reconstructed networks are sparser. Experiment results show that the identification accuracy of both parameters, average weighted degree and average weighted clustering coefficient, reach 93.3%, which can distinguish the 12-lead ECG signals of healthy people and patients with myocardial infarction, and realize the automatic detection of myocardial infarction.