- Keywords: Feature learning, Spatio-temporal correlations, Multi-channel time-series data, Healthcare applications
- TL;DR: We explore the use of autoencoders to do spatio-temporal feature learning and demonstrate that the order in which one stacks the autoencoders significantly impacts performance.
- Abstract: In modern medicine, patient vital sign information is often collected as high-dimensional multi-channel time series data, which contains both spatial and temporal information. In this paper, we propose a hybrid feature learning model containing both spatial and temporal autoencoders to learn deep feature representations of time series data. We use a publicly available electroencephalograph (EEG) dataset to evaluate our model's classification performance and compare the results to: (i) using raw data as features, and (ii) features learned from various combinations of spatial and temporal autoencoders. Our findings highlight that the way in which we exploit spatial and temporal correlations makes a significant difference and demonstrate the effectiveness of our model in processing multivariate time series patient data.