Leveraging electronic health records (EHR) data for healthcare predictive analytics has been receiving growing attention in recent years. High-throughput phenotyping is one of the analytics task where machine learning algorithms are used to derive phenotypes (sets of clinical conditions) from the EHR data to characterize patients of different diseases. In this project, we propose a deep tensor factorization framework for inferring highly interpretable phenotypes and dynamic patient representations from multi-modal EHR data. The proposed framework contains a temporal tensor model as its core for capturing (a) the interaction of the structured information (like diagnosis, medication, and lab tests), (b) the underlying phenotypes (as tensor factors), and (c) the temporal evolution of the phenotype portion (dynamic representation), as part of the model learning. As the temporal evolution of the health condition of a patient is complex in nature, deep models like recurrent neural network and neural Hawkes process can be integrated for regularizing the dynamic representations. In addition, the proposed framework can be integrated with a deep network architecture to learn to extract features from physiological time series like vital signs and ECG waveforms so that the associated predictive analytics tasks can be carried out in a patient-specific manner.
This project is supported by the General Research Fund (GRF), Research Grants Council (RGC), Hong Kong SAR, China (Project 12201219).
For further information on this research topic, please contact Dr. William K. CHEUNG.