In-hospital Mortality Prediction for Trauma Patients Using Cost-sensitive MedLDA

In intensive care units (ICUs), mortality prediction using vital sign or demographics of patients yields helpful information to support the decision-making of intensivists. Clinical texts recorded by medical staff tend to be valuable for prediction. However, text data are not applicable to outcome prediction of the regression framework in a direct way. In addition, learning of prediction models of such outcomes is a class of imbalanced data problem because the number of survivors is greater than the number of dead patients in most ICUs. To address these difficulties, we present Cost-Sensitive MedLDA: a supervised topic model employing cost-sensitive learning. The model realizes a prediction model from heterogeneous data such as vital signs, demographic information, and clinical text in an imbalanced class problem. Through experimentation and discussion, we demonstrate that the model has two benefits for use in medical fields: 1) our model has high prediction performance for minority instances while maintaining good performance for majority instances even if the training set is imbalanced data; 2) our model can reveal some characteristics that are associated with bad outcomes from the use of clinical texts.