Similarity-based LSTMs for Time Series Representation Learning in the Presence of Structured Covariates

Time series are prevalent in biomedical data, from high-frequency vital signals and longitudinal health indicators to gait kinematics and activity monitoring. One typical approach for the analysis of this data, frequently encountered in the biomedical literature, is to extract a set of features by applying tools such as PCA. Such general purpose techniques are insufficient to yield a complete view of time series, creating a need for more salient representations, in a setting where, often, only a limited amount of labels and annotations are available for this purpose. To address this issue, we introduce a deep learning framework which takes structured covariates into account when learning time series representations. The lower layers of the deep architecture include an embedding which captures parameters specific to the time series, such as frequency and amplitude, early in the transformation. This allows the rest of the network to operate on a higher-level abstraction. Our experiments show that this modification enables the prediction of osteoarthritis-related pain with an accuracy of 74%, compared to the 67% obtained by the top-performing method used by domain experts. In addition, we also modeled a component that directly encodes sample similarity, which ensures that features relevant to that sample are given more weight in issuing the prediction. This similarity-based LSTM achieved an accuracy of 94% in classifying running injuries based on the gait kinematics, compared to the 86% yielded by the state-of-the-art for this application. Our model is an alternative to painstaking feature engineering and achieves the goal of learning predictive features with a small amount of training samples.