Supervised Nonnegative Tucker Decomposition for Computational Phenotyping

With the availability of Electronic Health Records (EHR) data, lots of predictive tasks in medical practice seem solvable by building predictive models. However, EHR data always contains various medical concepts (e.g., diagnosis, medicines, lab tests) with high dimensions and mass correlations among them. To avoid the curse of dimensionality, the representation learning or dimensional reduction process is needed before prediction. Traditional methods either could not mine the correlations of different dimensions in datasets or lack of interpretability. Therefore we proposed a supervised tucker tensor decomposition (SNTD) which is guided by the prediction tasks for representation learning. Detailly speaking, SNTD constrains factor matrices by adding the label information of task and generating regularization term towards the non-negative tucker decomposition (NTD). Compared to the CP decomposition based methods, SNTD has better performance for that more flexible core tensor can contribute to the better learning of correlations among dimensions largely. We also demonstrate the accuracy and interpretability of our approach on an real-world EHR data for hospitalization task. Our result shows SNTD can not only obtain better performance of prediction than baseline methods, but also achieve the interpretability of representations.