Grant: Canada Research Chairs

We aim to develop next-generation tools and methods to analyze text streams and sensor data collected for disease and wellness monitoring. There are three objectives. (Objective 1) To develop tools to extract targeted information from clinical documents (e.g., pathology reports, radiology reports). These tools would reduce false positives in cancer reporting, facilitating cancer treatment selection, and stratifying risks for cancer relapses. (Objective 2) To analyze user-generated content to predict changes in an individual’s mood, and to predict self-harm and various psychiatry conditions. (Objective 3) To generate effective privacy-preserving synthetic data that promote data sharing of sensitive medical data and content. For Objective 1, our general approach is to apply methods in representation learning, which finds a new feature space for down-stream analysis. The prime example of representation learning is the Bidirectional Encoder Representations from Transformers (BERT). To extract targeted information (e.g., histology, tumour size) from a span of text in an unstructured clinical document, we will leverage the similarity analysis capability offered by BERT. We will also explore a “question-answering” approach by incorporating the “Text-to-Text Transfer Transformer” (T5) neural network model. A key issue is the design of appropriate questions to ask so that BERT and T5 can appropriately interpret the questions and extract the targeted information. BC Provincial Health Services Authority will provide 100,000+ documents for large scale validation. For Objective 2, we will use four large corpora to build models to capture various cognitive and affect states. We will also explore how to incorporate common sense knowledge and reasoning abilities in the models. Furthermore, for various psychiatry conditions, detecting semantic coherence of written text would be important, and we will conduct coherence and discourse analysis with appropriate use of transfer learning schemes. For synthetic data generation, we will focus on generating clinical text and time series data. For text, we will use large pre-trained language models, and fine-tune them in a memory-efficient way. For time series data, we will focuses on generation of multivariate time series originating from a single source. Finally, we will conduct explicit privacy assessments, so that data custodians and owners can tradeoff between privacy protection and utility based on their needs.