Abstract
Modern patient data tends to be large-scale and multi-dimensional, containing both spatial and temporal features. Learning good spatio-temporal features from large patient data is a challenging task, especially when there are missing observations. In this paper, we propose a spatio-temporal autoencoder (STAE), an unsupervised deep learning scheme, to learn features from large-scale and high-dimensional patient data with missing observations. Through both spatial and temporal encoding, STAE is able to automatically identify patterns and dependencies in the patient data, even with missing values, and learn a compact representation of each patient for better classification. Publicly available electroencephalogram (EEG) data are extracted from the UCI Machine Learning Repository to test and support our findings. Through simulations, we compare STAE with several baseline feature selection methods and demonstrate its effectiveness in the presence of missing data.