Exploring unsupervised multivariate time series representation learning for chronic disease diagnosis

The application of various sensors in hospitals has enabled the widespread utilization of multivariate time series signals for chronic disease diagnosis in the data-driven world. The key challenge is how to model the complex temporal (linear and nonlinear) correlations among multiple longitudinal variables. Due to scarcity of labels in practice, unsupervised learning methods have already become indispensable. However, state-of-the-art approaches mainly focus on the extraction of linear correlation-induced feature connectivity network, e.g., Pearson correlation, partial correlation, etc. To this end, for chronic disease (e.g., Parkinson disease) diagnosis, an unsupervised representation learning method is first designed to obtain informative correlation-aware signals from multivariate time series data. At the core is a contrastive learning framework with a graph neural network (GNN) encoder to capture the inter-correlation and intra-correlation of multiple longitudinal variables. Then, the previously learned representations are sent to a simple fully connected neural network, which can be trained using fewer labels compared with end-to-end complex supervised learning models. Further, to assist the decision-making process in the high-stake chronic disease detection task, model uncertainty quantification is enabled according to evidential theory. The experimental results on two public Parkinson’s disease data sets show the expressiveness of the learned embeddings, and the final lightweight classifier achieves the best performance.