Deep feature extraction for representing and classifying time series cases: Towards an interpretable approach in haemodialysis

Case-based retrieval and K-NN classification techniques are suitable for assessing haemodialysis treatment efficiency and for identifying risk situations. In this domain, cases involve time series data, that need to undergo a feature extraction phase in order to reduce dimensionality and to speed up similarity calculation. In this paper, we propose a deep learning architecture for time series feature extraction, based on the use of a convolutional autoencoder. Deep features provide a better time series representation with respect to features produced by the Discrete Cosine Transform (DCT). Indeed, in our experiments, K-NN classification based on deep features has outperformed the DCT-based one. We are also working in the direction of improving interpretability, by using case retrieval results obtained in a different feature space (defined on the basis of domain knowledge) to explain the outputs provided by the adoption of the deep learning technique.