Survival Topic Models for Predicting Outcomes for Trauma Patients

Data mining techniques have been proposed to predict mortality for ICU patients using their demographic data, measurements and notes from doctors and nurses. Most of these techniques suffer from two main drawbacks. First, they model the mortality prediction problem as a binary classification problem, while ignoring the time of death as continuous values. Second, they use topic models to analyze the notes, while ignoring the relationship between measurements, notes and mortality/discharge outcomes. In this paper we propose a novel model called the survival topic model (SVTM), which models patients' measurements, notes and mortality/discharge jointly, and predicts the probability of mortality/discharge as functions of time. The idea is that each patient has a latent distribution of disease conditions, which we call topics. These conditions generate the measurements and notes and determine the patients' mortality. We derive a mean-field variational inference algorithm for this model. We fitted the SVTM with two outcomes on Medical Information Mart for Intensive Care III (MIMIC III) trauma patients data and obtained some important topics. Also, we demonstrated the relationships between these topics.