Keywords: time series, classification, deep learning, representation learning, cerebral palsy, information fusion
Heterogeneous datasets, consisting of both structured and unstructured information, are prevalent in the biomedical domain. Time series, specifically, are frequently encountered in the form of vital signs, gait kinematics and longitudinal health indicators. Current approaches dealing with heterogeneous data handle the different types of data separately, or simply append representations of the time series data to the existing structured fields. We introduce ShortFuse, a method which introduces hybrid layers into CNN-based or LSTM-based architectures built for temporal features. These layers use structured information as distinct inputs, which are used to parametrize, guide, and enrich the representations. ShortFuse preserves the sequential structure of time series in a way that explicitly models interactions and dependencies with structured covariates. We show that the proposed method of introducing structured covariates in the learning process results in improved performance over the state-of-the-art for three biomedical classification tasks, including the prediction of surgery outcome for children with Cerebral Palsy.