The application of Data Analysis methods to temporal sequences is limited because of the inherent high temporal dimension of the data. One promising solution consists in performing dimensionality reduction before analysing the sequences. We propose a new technique that reduces the temporal dimension of a multivariate sequence while preserving correlations between the sequence's descriptive features, a highly important property for many Data Analysis process. We illustrate the effectiveness of the proposed technique for a multivariate temporal sequence of an intensive care unit application.