June 13, 2019 | (4.00 PM)
USA (United States of America)
In the presented study, we sought to automatically quantify the mortality risk associated with all risk factors mentioned in an EHR text. Given the multitude of relevant concepts in such a task, the issue of taking into account the influence of the context for each of them was particularly challenging. We formulated it as a multi-task learning problem and solved through linear regression with appropriately designed regularization. The evaluation showed the proposed solution effectively assessing risk even for previously unseen factors and the overall performance of risk quantification reaching the level of human annotators' agreement. The obtained results point to the great potential of linear regularized methods in highly multi-dimensional and multi-task challenges, such as information extraction from electronic health records.